CN117639602A

CN117639602A - Self-adaptive motor running state adjusting method and system

Info

Publication number: CN117639602A
Application number: CN202410107573.8A
Authority: CN
Inventors: 徐艳荣; 邓智斌; 陈程; 陈晓玲; 杨健; 邸亮
Original assignee: Shenzhen Kaisheng United Technology Co ltd
Current assignee: Shenzhen Kaisheng United Technology Co ltd
Priority date: 2024-01-26
Filing date: 2024-01-26
Publication date: 2024-03-01

Abstract

The invention relates to the technical field of PID control, in particular to a motor running state self-adaptive adjusting method and system, comprising the following steps: based on the existing motor operation data, modeling the probability relation among parameters by adopting a Bayesian network construction algorithm, refining the probability dependency relation among key parameters by a statistical learning method, and carrying out probability analysis to generate a motor operation probability network. According to the invention, through the application of a Bayesian network and a structural equation model, the understanding of complex relationships among parameters in the running state of the motor is enhanced, the identification capability of the motor performance deviation root causes is improved through the use of multivariate regression analysis, the application of a differential geometric method and a graph neural network algorithm in the aspects of dynamic load response and system interaction is improved, the adaptability and the overall efficiency of a motor system are improved, and the application of a depth deterministic strategy gradient algorithm in load distribution optimization is improved, so that a control system of the motor is flexibly adapted to various load conditions, and the energy waste is reduced.

Description

Self-adaptive motor running state adjusting method and system

Technical Field

The invention relates to the technical field of PID control, in particular to a motor running state self-adaptive adjusting method and system.

Background

PID control technology is a widely used control algorithm, particularly in the control system of an electric motor. PID stands for Proportional (pro), integral (Integral) and Derivative (Derivative), which form the basis of the PID controller. In the motor operating state adaptive adjustment method, PID control is used to automatically adjust the operating parameters of the motor to accommodate different operating conditions and load requirements. The method relies on real-time feedback, and can accurately adjust key parameters such as the speed, the position, the torque and the like of the motor.

The main purpose of the motor running state self-adaptive adjusting method is to improve the running efficiency, stability and response speed of the motor. By using PID control, this approach enables the motor to maintain optimal performance under a variety of different operating conditions. For example, in the event of sudden load changes or power fluctuations, the motor may quickly adjust its operating state to maintain a stable output. The reliability and the service life of the motor are improved, the energy consumption can be reduced, and more economical and efficient operation is realized.

The traditional motor control method has obvious defects in processing complex parameter relationships, causal inference and performance deviation identification. Due to the lack of advanced data analysis and machine learning techniques, conventional methods have difficulty in accurately understanding and predicting the interaction effects between parameters, resulting in lower efficiency in performance optimization and fault diagnosis. In addition, conventional methods have insufficient flexibility in dynamic load response and load distribution, and often cannot efficiently adapt to rapidly changing operating conditions, which results in energy inefficiency and system stability problems.

Disclosure of Invention

The invention aims to solve the defects in the prior art, and provides a motor running state self-adaptive adjusting method and system.

In order to achieve the above purpose, the present invention adopts the following technical scheme: the self-adaptive motor running state adjusting method comprises the following steps:

s1: modeling the probability relation among parameters by adopting a Bayesian network construction algorithm based on the existing motor operation data, refining the probability dependency relation among key parameters by a statistical learning method, and carrying out probability analysis to generate a motor operation probability network;

s2: based on the motor operation probability network, carrying out causal relationship inference analysis by adopting a structural equation model, analyzing causal relationship among various parameters by a data driving method, verifying the reliability of the relationship, and generating a motor causal relationship model;

s3: analyzing the root cause of the performance deviation by adopting a multivariate regression analysis method based on the motor causal relationship model, analyzing key parameters, identifying key factors influencing the motor performance, and generating a performance deviation root cause analysis result;

s4: based on the analysis result of the performance deviation root cause, adopting a differential geometric method in a control theory to design a dynamic load response mechanism of the motor, and matching differential load conditions through optimizing a control strategy and adjusting response parameters to generate a differential geometric load response control strategy;

S5: based on the differential geometric load response control strategy, adopting a graph neural network algorithm to analyze and optimize the interaction relation of a plurality of groups of motors, and strengthening the interaction efficiency among the motors through network learning and pattern recognition to generate a motor interaction network optimization model;

s6: based on the motor interaction network optimization model, a depth deterministic strategy gradient algorithm is adopted to learn and optimize self-adaptive load distribution, and a load distribution strategy is adjusted by combining motor performance parameters and real-time feedback data to generate a self-adaptive load optimization strategy;

s7: based on the self-adaptive load optimization strategy, a real-time feedback adjustment algorithm is adopted to dynamically adjust a control system of the motor, operation data are monitored in real time, control parameters are optimized according to feedback information, and a dynamic adjustment control instruction is generated.

As a further scheme of the invention, the motor operation probability network comprises probability distribution data of load rate, temperature change and current fluctuation of the motor, the motor causality relation model comprises direct causality relation between factors of temperature change and current fluctuation and motor performance, the performance deviation root cause analysis result comprises a root cause list of key parameter deviation, operation condition abnormality and environmental factors, the differential geometric load response control strategy comprises load change response adjustment, performance optimization path and control parameter setting, the motor interaction network optimization model comprises synchronous efficiency, load sharing strategy and fault response coordination among the motors, the self-adaptive load optimization strategy comprises a load distribution algorithm, response time optimization and energy consumption management strategy, and the dynamic adjustment control instruction comprises motor start-stop strategy, operation mode selection and fault prevention adjustment.

As a further scheme of the invention, based on the existing motor operation data, modeling of probability relations among parameters is performed by adopting a Bayesian network construction algorithm, probability dependency relations among key parameters are refined by a statistical learning method, probability analysis is performed, and the step of generating a motor operation probability network specifically comprises the following steps:

s101: based on the existing motor operation data, a Bayesian network structure learning algorithm is adopted to construct a network structure, the connection mode of a plurality of nodes in a network is determined by evaluating the relation among a plurality of parameters, and meanwhile, key parameters are selected to generate a Bayesian network structure;

s102: based on the Bayesian network structure, analyzing the mutual information quantity among parameters by using a mutual information calculation method, revealing the mutual dependency relationship by calculating the mutual information value of each pair of parameters, and constructing a parameter dependency relationship graph;

s103: based on the parameter dependency graph, a Bayesian network parameter learning algorithm is applied to quantitatively analyze the parameter relationship in the network, the conditional probabilities of a plurality of nodes are estimated, a probability dependency model among a plurality of parameters is constructed, and a Bayesian network probability model is generated;

s104: based on the Bayesian network probability model, a network optimization and calibration method is adopted, connection is added or removed to a network structure, the performance of the model on various data sets is tested by using a cross verification method, the model precision is optimized, and then a motor operation probability network is generated.

As a further scheme of the invention, based on the motor operation probability network, a structural equation model is adopted to conduct causal relationship inference analysis, causal relationship among various parameters is analyzed through a data driving method, reliability of the relationship is verified, and the step of generating the motor causal relationship model specifically comprises the following steps:

s201: based on the motor operation probability network, a Bayesian network analysis method is adopted, nodes in the network are selected to represent the differentiated operation parameters of the motor, meanwhile, probability relations among the nodes are defined through probability distribution estimation, potential connection among the nodes is detected and established through dependency relation learning, and then a network model structure is optimized, adjusted and confirmed through a network structure, so that a motor operation preliminary probability model is generated;

s202: based on the motor operation preliminary probability model, adopting path analysis, setting an assumed path and distributing path coefficients, then using a structural equation model to analyze the path, simultaneously verifying the strength of causal relation of the assumed path through model fitting, evaluating direct and indirect influences among a plurality of variables in the model, and generating a potential causal relation graph among motor parameters;

s203: based on potential causal relation graphs among motor parameters, cross verification is applied, data are divided into a plurality of subsets, training and verification of a model are respectively carried out by the subsets, stability and reliability of the model on the plurality of data sets are verified, meanwhile, statistical hypothesis verification is used, and the verified causal relation graphs are generated by determining statistics of horizontal evaluation causal relation;

S204: based on the verified causal relationship graph, model comprehensive optimization is adopted, the model is reconfigured, model efficiency is improved, parameter fine adjustment is carried out, and the causal relationship in motor operation is reflected through adjustment of weights and deviations, so that a motor causal relationship model is generated.

As a further scheme of the invention, based on the motor causal relationship model, the root cause of the performance deviation is analyzed by adopting a multivariate regression analysis method, key parameters are analyzed, key factors influencing the motor performance are identified, and the step of generating the analysis result of the performance deviation root cause is specifically as follows:

s301: based on the motor causality model, calculating standard deviation and mean value of each pair of variables by adopting a Pearson correlation coefficient algorithm, simultaneously carrying out product and subtraction operation on each data point of each pair of variables, accumulating results, dividing the accumulated results by the number of samples, subtracting one to obtain a correlation coefficient, identifying parameters related to motor performance deviation, and further generating a key parameter correlation analysis result;

s302: based on the key parameter correlation analysis result, carrying out centering treatment on parameter data by using a principal component analysis technology, eliminating a mean value, constructing a covariance matrix of parameters, extracting characteristic values and characteristic vectors of the covariance matrix, selecting the characteristic vectors according to a descending order of the characteristic values as principal components, and projecting original data onto the principal components to generate a parameter principal component analysis result;

S303: based on the analysis result of the parameter principal components, a multiple linear regression analysis technology is applied, a plurality of principal components are selected as independent variables, the motor performance deviation is used as a dependent variable, meanwhile, the coefficient in a regression equation is calculated through a least square method, the product sum of the independent variables and the dependent variable is summed, and a regression model is built by using the coefficient, so that a multiple linear regression model is generated;

s304: based on the multiple linear regression model, diagnosing and verifying the model, checking whether the distribution of residual errors accords with normal distribution, simultaneously evaluating the analysis capability of the model to the data variability by using a decision coefficient, and performing F test to generate a performance deviation root cause analysis result;

the pearson correlation coefficient algorithm adopts the formula:

；

wherein,and->For paired data points>And->For the mean value of the corresponding variables>Is the correlation coefficient that is calculated.

As a further scheme of the invention, based on the analysis result of the performance deviation root cause, a differential geometric method in a control theory is adopted to design a dynamic load response mechanism of the motor, and the steps of generating the differential geometric load response control strategy by optimizing the control strategy and adjusting response parameters and matching differential load conditions are specifically as follows:

S401: based on the analysis result of the performance deviation root cause, a state space modeling method is adopted, a dynamic equation is established by defining the relation between the motor performance parameter and the state variable, and the performance change of the motor under various operation conditions is expressed in a mathematical form, so that a state space performance deviation model is generated;

s402: based on the state space performance deviation model, a differential geometry method is adopted, the internal structure of motor performance change is revealed through analyzing the geometric characteristics of a state space, and meanwhile, the behavior of the model at a differential operation point is researched by using a differential geometry tool, so that a differential geometric characteristic analysis result is generated;

s403: based on the analysis result of the differential geometric characteristics, adopting a feedback control optimization method to adjust a control strategy to match various load conditions, and simultaneously designing a feedback loop and adjusting parameters of a controller to optimize the response of a motor to load change to generate an optimized load response control strategy;

s404: and based on the optimized load response control strategy, performing system parameter adjustment and comprehensive test, and generating a differential geometric load response control strategy by adjusting control parameters and testing performance under various load conditions.

As a further scheme of the invention, based on the differential geometric load response control strategy, the interactive relation of a plurality of groups of motors is analyzed and optimized by adopting a graph neural network algorithm, the interactive efficiency among the motors is enhanced by network learning and pattern recognition, and the step of generating the motor interactive network optimization model comprises the following steps:

S501: based on the differential geometric load response control strategy, a graph neural network algorithm is adopted, a plurality of components in a motor system are defined as nodes, interaction among the components is used as edges, initial characteristics including temperature and current are given to each node, and a graph neural network initial structure is generated;

s502: based on the initial structure of the graph neural network, the features of each node and the features of adjacent nodes are aggregated by adopting a graph convolution algorithm, the features of the current node are updated and the network is trained by adding up the features of weighted adjacent nodes, and an interaction mode in the system is assisted to be captured, so that a network model after graph convolution is generated;

s503: based on the network model after graph convolution, mapping analysis is carried out on graph structure data by utilizing a graph embedding technology, and topological relation and characteristic information among nodes in an original graph are reserved through optimizing node representation, so that low-dimensional graph embedding representation is generated;

s504: based on the low-dimensional graph embedded representation, the graph neural network is subjected to learning rate adjustment, network layer number and node number optimization, and the generalization and prediction precision of the model are evaluated by applying a cross-validation technology, so that a motor interaction network optimization model is generated.

As a further scheme of the invention, based on the motor interaction network optimization model, a depth deterministic strategy gradient algorithm is adopted to learn and optimize self-adaptive load distribution, and a load distribution strategy is adjusted by combining motor performance parameters and real-time feedback data, so that the self-adaptive load optimization strategy is generated by the steps of:

s601: based on the motor interaction network optimization model, adopting a graph neural network algorithm, identifying key network connection and characteristics affecting system performance by learning characteristics of a plurality of nodes in a network and relationships among the nodes, and simultaneously analyzing an interaction structure among motors to generate network characteristic comprehensive analysis;

s602: based on the network characteristic comprehensive analysis, a depth deterministic strategy gradient algorithm is applied, a network structure is analyzed by combining deep learning, meanwhile, loads of a motor system are distributed by utilizing a strategy gradient method, influence of a differentiated load distribution scheme on system performance is evaluated, automatic learning is performed, an initial load distribution strategy is built, and a preliminary self-adaptive load distribution model is generated;

s603: based on the preliminary self-adaptive load distribution model, performing iterative optimization of reinforcement learning, performing multiple tests in a simulation environment, adjusting a load distribution strategy according to test results, and simultaneously referring to real-time data feedback and multiple dynamic changes, perfecting the load distribution strategy through cyclic tests and adjustment, and further generating an optimally-adjusted load distribution model;

S604: based on the load distribution model of optimization adjustment, a performance test and evaluation method is adopted to test the model under differential working conditions and load change, evaluate performance and stability, and simultaneously perform iterative adjustment of strategy parameters, including adjustment of key parameters of learning rate and return discount rate in algorithm, optimize the load distribution capacity of the model, and generate a self-adaptive load optimization strategy.

As a further scheme of the invention, based on the self-adaptive load optimization strategy, a real-time feedback adjustment algorithm is adopted to dynamically adjust a control system of the motor, operation data is monitored in real time, control parameters are optimized according to feedback information, and the step of generating a dynamic adjustment control instruction specifically comprises the following steps:

s701: based on the self-adaptive load optimization strategy, a time sequence analysis method is adopted to analyze the time sequence of the motor operation data, the trend and the periodic variation are identified, and meanwhile, abnormal points in the data are analyzed by using an isolated forest algorithm, so that the operation condition of the motor is monitored in real time, and a real-time performance index is generated;

s702: based on the real-time performance index, a state estimation method is adopted, a Kalman filter is utilized to process noise and uncertainty factors in motor state data, the dynamic state of the motor is continuously estimated, a Bayesian network is simultaneously utilized to carry out probabilistic reasoning, the motor state change is analyzed, and a state estimation analysis result is generated;

S703: based on the state estimation analysis result, adopting an adaptive control method, adjusting a control strategy according to the difference between the actual performance and the expected performance of the motor by referring to the adaptive control, and simultaneously adjusting a neural network parameter according to the real-time feedback of the state of the motor by using the adaptive neural network control to generate an optimized control parameter;

s704: based on the optimized control parameters, a predictive control method is adopted to analyze the current state of the motor, simultaneously predict the performance in the future short period, analyze the batch historical data and the instant feedback information, identify the potential running trend and the performance change, and then adjust the control parameters by combining a proportional-integral-differential adjustment method, wherein the difference between the current output and the expected target is compared in a circulating manner in the adjustment process, dynamically adjust the control parameters, and generate a dynamic adjustment control instruction.

The system comprises a probability network construction module, a causal reasoning module, a performance deviation analysis module, a load response strategy module, an interactive network optimization module and a self-adaptive load adjustment module;

The probability network construction module is used for constructing a motor operation probability network based on motor operation data by using a Bayesian network structure learning algorithm, evaluating the relation among parameters, selecting a node connection mode, and then learning and deducing probability dependence among nodes from the data to generate the motor operation probability network;

the causal reasoning module is used for analyzing causal relations among motor operation parameters by using a structural equation model based on a motor operation probability network, and generating a motor causal relation model by establishing a statistical model and carrying out hypothesis verification and analyzing causal relations among parameters;

the performance deviation analysis module is used for analyzing the root cause of the motor performance deviation by adopting a multivariate regression analysis method based on a motor causal relationship model, identifying and quantifying key factors influencing the motor performance, and generating a performance deviation root cause analysis result;

the load response strategy module designs a load response strategy by using a differential geometry control method based on a performance deviation root cause analysis result, and simultaneously analyzes the geometry characteristics of a state space and adjusts the control strategy to match differential load conditions so as to generate a differential geometry load response control strategy;

the interaction network optimization module analyzes interaction of internal components of the motor system by adopting a graph neural network algorithm, and optimizes interaction efficiency among motors by learning characteristics and relations of nodes in a network to generate a motor interaction network optimization model;

The self-adaptive load adjustment module is used for learning and optimizing motor load distribution by using a depth deterministic strategy gradient algorithm based on a motor interaction network optimization model, and is used for adjusting a load distribution strategy by combining real-time data to generate a self-adaptive load optimization strategy.

Compared with the prior art, the invention has the advantages and positive effects that:

in the invention, through the application of the Bayesian network and the structural equation model, the understanding of the complex relationship between parameters in the motor running state is enhanced, and the analysis is more accurate and deeper on the inference of the causal relationship. The use of multivariate regression analysis improves the ability to identify the root cause of motor performance deviations, thereby achieving more effective problem resolution and performance optimization. The application of the differential geometry method and the graph neural network algorithm in the aspects of dynamic load response and system interaction greatly improves the adaptability and the overall efficiency of the motor system. The application of the depth deterministic strategy gradient algorithm on load distribution optimization enables a control system of a motor to be more flexibly adapted to various load conditions, reduces energy waste and improves the running stability of the system.

Drawings

FIG. 1 is a schematic workflow diagram of the present invention;

FIG. 2 is a S1 refinement flowchart of the present invention;

FIG. 3 is a S2 refinement flowchart of the present invention;

FIG. 4 is a S3 refinement flowchart of the present invention;

FIG. 5 is a S4 refinement flowchart of the present invention;

FIG. 6 is a S5 refinement flowchart of the present invention;

FIG. 7 is a S6 refinement flowchart of the present invention;

FIG. 8 is a S7 refinement flowchart of the present invention;

fig. 9 is a system flow diagram of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

In the description of the present invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," and the like indicate orientations or positional relationships based on the orientation or positional relationships shown in the drawings, merely to facilitate describing the present invention and simplify the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and therefore should not be construed as limiting the present invention. Furthermore, in the description of the present invention, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.

Embodiment one: referring to fig. 1, the present invention provides a technical solution: the self-adaptive motor running state adjusting method comprises the following steps:

s2: based on a motor operation probability network, carrying out causal relationship inference analysis by adopting a structural equation model, analyzing causal relationship among various parameters by a data driving method, verifying the reliability of the relationship, and generating a motor causal relationship model;

s3: based on a motor causal relationship model, analyzing the root cause of the performance deviation by adopting a multivariate regression analysis method, analyzing key parameters, identifying key factors influencing the motor performance, and generating a performance deviation root cause analysis result;

s4: based on the analysis result of the performance deviation root cause, adopting a differential geometric method in a control theory to design a dynamic load response mechanism of the motor, and matching differential load conditions by optimizing a control strategy and adjusting response parameters to generate a differential geometric load response control strategy;

S5: based on a differential geometric load response control strategy, adopting a graph neural network algorithm to analyze and optimize interaction relations of a plurality of groups of motors, and strengthening interaction efficiency among the motors through network learning and pattern recognition to generate a motor interaction network optimization model;

s6: based on a motor interaction network optimization model, a depth deterministic strategy gradient algorithm is adopted to learn and optimize self-adaptive load distribution, and a load distribution strategy is adjusted by combining motor performance parameters and real-time feedback data to generate a self-adaptive load optimization strategy;

The motor operation probability network comprises probability distribution data of load rate, temperature change and current fluctuation of the motor, the motor causality model comprises direct causality of factors of temperature change and current fluctuation and motor performance, performance deviation root cause analysis results comprise root cause lists of key parameter deviation, operation condition abnormality and environmental factors, the differential geometric load response control strategy comprises load change response adjustment, performance optimization paths and control parameter setting, the motor interaction network optimization model comprises synchronous efficiency, load sharing strategy and fault response coordination among the motors, the self-adaptive load optimization strategy comprises a load distribution algorithm, response time optimization and energy consumption management strategy, and the dynamic adjustment control instruction comprises motor start-stop strategy, operation mode selection and fault prevention adjustment.

In the step S1, probability relation modeling is carried out on the existing motor operation data through a Bayesian network construction algorithm. First, the data preprocessing stage normalizes motor operation data, including noise cancellation and normalization formats, to ensure data consistency and analyzability. Then, a bayesian network algorithm is used, which represents the dependency between the variables by means of a probability map model. In the network structure learning stage, the algorithm evaluates the condition dependence among different parameters in the data, and the network structure is determined by using methods such as maximum likelihood estimation or Bayesian information criterion. Subsequently, during the parameter learning phase, the algorithm estimates the probability distribution of each edge in the network by means of statistical learning methods, such as a desired maximization algorithm. The finally generated motor operation probability network reveals probability dependency relationship among key parameters, and lays a foundation for further analysis.

In step S2, based on the motor operation probability network, a structural equation model is adopted to conduct inference analysis on the causal relationship, and the structural equation model and a statistical model containing a plurality of regression equations are used at the stage to reveal the causal relationship among variables. By combining theoretical knowledge and data analysis, the model plays a key role in identifying potential variables and establishing linear relationships between observed variables. During analysis, parameter estimation methods such as maximum likelihood estimation are used to estimate coefficients in the model, while goodness-of-fit indicators such as chi-square statistics and root mean square errors are used to assess the accuracy of the model. Through the step, the generated motor causal relation model not only reveals causal relation among parameters, but also verifies the reliability of the relation, and provides important basis for subsequent performance analysis.

In the step S3, the root cause of the performance deviation is analyzed by using a multivariate regression analysis method based on the motor causal relationship model. The stage is focused on identifying key factors influencing motor performance, and the motor performance data is deeply analyzed by using a multivariate regression method. The method analyzes data by constructing a regression equation taking into account the influence of a plurality of independent variables on one dependent variable (performance bias). In the regression analysis process, the steps of coefficient estimation, residual analysis, statistical significance test and the like are performed to ensure the accuracy and reliability of the analysis result. Through the analysis, the obtained performance deviation root cause analysis results not only help identify key parameters affecting the motor performance, but also provide directions for adjusting the parameters.

In the step S4, a dynamic load response mechanism of the motor is designed by using a differential geometry method based on the analysis result of the performance deviation root cause, and the key of the stage is to apply a differential geometry technology in a control theory, so that the complexity of a nonlinear control system can be processed by the method. By analyzing the dynamic model of the motor, a targeted control strategy and response parameters are designed to optimize the adaptability of the motor to the differential load conditions. The differential geometry method can improve the stability and response speed of the system by adjusting parameters in the control law, such as gain and time constant. The generated differential geometric load response control strategy not only enhances the load adaptation capacity of the motor, but also improves the running efficiency of the motor.

In the S5 step, a motor interaction network is optimized by adopting a graph neural network algorithm based on a differential geometric load response control strategy. The graph neural network algorithm can effectively process complex interaction relations among multiple groups of motors by operating on a graph structure. The algorithm uses nodes and edges to represent the motor and relationships to each other, which are analyzed and optimized through web learning and pattern recognition. In the algorithm, node feature extraction and edge weight updating are key steps, and the steps enhance feature expression by propagating and aggregating node information. The generated motor interaction network optimization model not only improves interaction efficiency among motors, but also enhances overall performance of the system.

In the S6 step, based on the motor interaction network optimization model, learning and optimization of self-adaptive load distribution are performed by using a depth deterministic strategy gradient algorithm. The algorithm combines deep learning and reinforcement learning, and can process a high-dimensional state space and a continuous action space. By constantly learning and adjusting the strategy, the algorithm can optimize load distribution based on consideration of motor performance parameters and real-time feedback data. The key to the algorithm is the design and training of the strategy network and the value network through which the algorithm can predict the optimal actions and estimate their value. The generated adaptive load optimization strategy can flexibly adjust load distribution to adapt to different operating conditions.

In the step S7, based on the self-adaptive load optimization strategy, a real-time feedback adjustment algorithm is adopted to dynamically adjust a control system of the motor, and the core of the stage is to monitor the operation data of the motor in real time and optimize control parameters according to feedback information. The real-time feedback adjustment algorithm can quickly respond to the change of the system and timely adjust the control instruction to keep the optimal running state of the system. By means of a dynamic adjustment algorithm, the control system is able to fine-tune control parameters, such as proportional, integral and derivative parameters of a PID controller, based on real-time data. The dynamic adjustment not only improves the adaptability and flexibility of the motor, but also ensures the stability and high efficiency of the motor in a continuously-changing running environment. The finally generated dynamic adjustment control instruction provides accurate and efficient control for the motor, and continuous and stable operation of the system is ensured.

Referring to fig. 2, based on the existing motor operation data, modeling of probability relationships between parameters is performed by using a bayesian network construction algorithm, probability dependency relationships between key parameters are refined by using a statistical learning method, probability analysis is performed, and the step of generating a motor operation probability network specifically includes:

s102: based on a Bayesian network structure, analyzing the mutual information quantity among parameters by using a mutual information calculation method, revealing the mutual dependency relationship by calculating the mutual information value of each pair of parameters, and constructing a parameter dependency relationship graph;

s103: based on the parameter dependency graph, a Bayesian network parameter learning algorithm is applied to quantitatively analyze the parameter relation in the network, the conditional probability of a plurality of nodes is estimated, a probability dependency model among a plurality of parameters is constructed, and a Bayesian network probability model is generated;

s104: based on a Bayesian network probability model, a network optimization and calibration method is adopted, connection is added or removed to a network structure, and the performance of the model on various data sets is tested by using a cross verification method, so that the model precision is optimized, and a motor operation probability network is further generated.

In the step S101, firstly, a network structure is constructed by using the existing motor operation data and adopting a bayesian network structure learning algorithm, the data format involved in the process is usually multidimensional, various parameters of motor operation such as temperature, current, rotation speed and the like are included, and the data exist in a time sequence or discrete numerical form. The core of the bayesian network structure learning algorithm is to evaluate the interrelationship between these parameters in order to determine the manner of connection of the nodes in the network. Specifically, the algorithm first performs preprocessing of the data, including normalization and missing value processing of the data, to ensure accuracy of the subsequent learning process. Next, the algorithm learns the dependency between the parameters using, for example, K2, PC algorithm, etc. The algorithms gradually build a network structure by evaluating the conditional independence of the parameters in the data. During the network construction process, the algorithm will select key parameters, which are important factors affecting the motor operation state. By means of iteration, the algorithm gradually optimizes the network structure until the network form which can represent the data relationship most is found. The result of this process is a bayesian network structure that reveals the dependency between the parameters of the motor operation, providing the basis for subsequent analysis.

In the step S102, the mutual information amount between the parameters is analyzed by using a mutual information calculation method based on the bayesian network structure. Mutual information calculation is a method for measuring the degree of interdependence between two variables. In this step, a calculation formula for the mutual information value between each pair of parameters is first defined, which usually involves the calculation of a joint probability distribution and an edge probability distribution. The algorithm then traverses each pair of nodes in the network, calculating mutual information values between each other. In this way, the dependency between the various parameters in the network can be revealed, the key to the steps being the accurate calculation of the mutual information values, which requires efficient algorithms to handle large amounts of data. After the calculation is completed, the algorithm constructs a parameter dependency graph which intuitively shows the dependency among the parameters and provides a visual tool for further analysis.

In the step S103, based on the parameter dependency graph, a bayesian network parameter learning algorithm is applied to quantitatively analyze the parameter relationship in the network. The goal of this step is to estimate the conditional probabilities of multiple nodes in the network, thereby constructing a probability-dependent model between multiple parameters. The algorithm first needs to determine the form of the probability distribution for each node in the network, either discrete or continuous. The algorithm then uses methods such as maximum likelihood estimation, expectation maximization, etc. to estimate the parameters of these distributions. The method enables the network model to be best fit with actual data by iteratively optimizing parameters, a Bayesian network probability model is generated as a result of the steps, the model can describe probability relations among various parameters in motor operation, and an important tool is provided for prediction and decision.

Finally, in the step S104, based on the bayesian network probability model, the accuracy of the model is improved by adopting a network optimization and calibration method, and the steps involve adjusting the network structure, including adding or removing connections, and adjusting parameters. The optimization method comprises a greedy algorithm, a simulated annealing algorithm and the like, and the algorithm finds the optimal model configuration by continuously adjusting the network structure. Meanwhile, the cross-validation method is used for testing the performances of the model on various data sets, which is helpful for evaluating the generalization capability of the model, and the result of the steps is to generate an optimized and calibrated motor operation probability network, which can more accurately predict the operation state of the motor and provide powerful support for maintenance and fault prevention of the motor.

Firstly, collecting motor operation data including multidimensional parameters such as temperature, rotating speed, current, vibration frequency and the like, and recording according to minutes. Based on these data, step S101 performs bayesian network structure learning using a K2 algorithm, and constructs a network structure by evaluating correlations between parameters. In this process, the algorithm normalizes the data, ensures accuracy, and identifies key parameters such as temperature and current, which are critical to motor fault prediction. Then, in step S102, the mutual information amount between the parameters is analyzed by using a mutual information calculation method, the mutual information value of each pair of parameters, such as the relationship between temperature and current, is calculated to reveal the mutual dependence, and a parameter dependence graph is constructed to intuitively show the mutual influence intensity between the parameters. Then, in step S103, a bayesian network parameter learning algorithm, such as an Expectation Maximization (EM) algorithm, is used to quantitatively analyze the parameter relationships in the network, estimate the conditional probability, and construct a probability dependency model, so as to obtain a model describing the probability relationships among the parameters in the motor operation. Finally, in step S104, a network optimization and calibration method, such as a greedy algorithm, is used to adjust the network structure and perform cross-validation to optimize the model accuracy.

Referring to fig. 3, based on a motor operation probability network, a structural equation model is adopted to perform causal relationship inference analysis, causal relationship among various parameters is analyzed through a data driving method, reliability of the relationship is verified, and the step of generating a motor causal relationship model specifically comprises the following steps:

s201: based on a motor operation probability network, a Bayesian network analysis method is adopted, nodes in the network are selected to represent differentiated operation parameters of the motor, probability relations among the nodes are defined through probability distribution estimation, potential connection among the nodes is detected and established through dependency relation learning, and then a network model structure is optimized, adjusted and confirmed through a network structure, so that a motor operation preliminary probability model is generated;

s202: based on a motor operation preliminary probability model, adopting path analysis, setting an assumed path and distributing path coefficients, then using a structural equation model to analyze the path, simultaneously verifying the strength of causal relation of the assumed path through model fitting, evaluating direct and indirect influences among a plurality of variables in the model, and generating a potential causal relation graph among motor parameters;

s203: based on potential causal relation graphs among motor parameters, cross verification is applied, data are divided into a plurality of subsets, model training and verification are respectively carried out by the subsets, stability and reliability of the model on a plurality of data sets are verified, meanwhile statistical hypothesis tests are used, and the verified causal relation graphs are generated by determining statistics of horizontal evaluation causal relation;

S204: based on the verified causal relationship graph, the model is comprehensively optimized, the model is reconfigured, the model efficiency is improved, parameter fine adjustment is carried out, and the causal relationship in the motor operation is reflected by adjusting the weight and the deviation, so that a motor causal relationship model is generated.

In the sub-step S201, the motor operation probability network is further analyzed and constructed by bayesian network analysis. Nodes in the network are first selected, which represent differentiated operating parameters of the motor, such as current, temperature, rotational speed, etc. The data format is typically continuous or discrete numerical data reflecting the real-time operating conditions of the motor. After defining the nodes, the probability relationships between the nodes are defined by probability distribution estimation. This requires calculation of a conditional probability distribution, for example using bayesian rules and probability chain rules to estimate the probability of one node given the state of other nodes. Then, potential connections between nodes are detected and established using a dependency learning method, such as a mutual information method, in which an algorithm evaluates statistical correlations between different combinations of nodes to determine if there are direct dependencies. Through these analyses, the structure of the network is progressively clarified and optimized. Next, network structure optimization adjustments are made, such as adding or removing edges in the network using structure learning algorithms, to improve interpretation and predictive power of the model. These optimization measures ensure that the network model structure more accurately reflects the actual condition of motor operation. After the steps are completed, the generated motor operation preliminary probability model not only provides a comprehensive view of the motor operation state, but also reveals complex interactions between different operation parameters.

In the sub-step S202, path analysis is performed based on the motor operation preliminary probability model to reveal potential causal relationships between motor parameters. First, hypothetical paths are set and path coefficients are assigned, which represent causal relationships between variables in the model. The data format is still numerical, reflecting the operating parameters of the motor. These paths are analyzed using a structural equation model, which analysis involves building a system of equations to describe the linear relationship between the variables and estimating the path coefficients, typically by least squares or maximum likelihood estimation. In this way, both direct and indirect effects in the model can be quantified. Then, model fitting verification is carried out, statistical methods such as chi-square test and goodness-of-fit index are used for evaluating the effectiveness of the hypothesized paths, potential causal relation among motor parameters is revealed in the process, a causal relation graph is generated, deep holes for interaction among the motor parameters are provided, and basis is provided for subsequent adjustment strategies.

In the sub-step S203, a cross-validation method is applied to validate the causal relationship graph between motor parameters, and the step is to train and validate the model by dividing the data into a plurality of subsets, respectively, so as to verify the stability and reliability of the model on different data sets. The method ensures the generalization capability of the model and improves the credibility of the result. The data format used is the same as the previous step, but the segmentation process is performed. Furthermore, the statistical significance of causal relationships is assessed by deterministic levels using statistical hypothesis tests, such as t-test or F-test. The result of this step is a validated causal graph that demonstrates causal relationships between motor operating parameters with a higher degree of confidence, providing powerful data support for motor fault diagnosis and performance optimization.

In the S204 substep, model comprehensive optimization is performed to further improve efficiency and accuracy of the motor causal relationship model. This involves reconfiguring the model structure, optimizing parameters in the model, such as adjusting weights and biases, to more accurately reflect causal relationships in motor operation, and employing algorithms such as gradient descent to fine tune model parameters to ensure that the output of the final model is as close as possible to the actual observations. After the operations are completed, the finally generated motor causal relation model not only has a solid foundation in theory, but also shows high precision and applicability in practical application, and provides a powerful analysis tool for intelligent monitoring and predictive maintenance of the motor.

First, motor operation data including temperature, rotational speed, current, vibration frequency, etc. are collected and recorded in a time series format. In step S201, a preliminary probability model of motor operation is constructed using bayesian network analysis based on the data. By selecting nodes reflecting motor operation parameters, calculating conditional probability distribution, and detecting and establishing potential connection between the nodes by using a mutual information method and the like. And then, optimizing the network structure to ensure that the model accurately reflects the real-time running state of the motor. Subsequently, in step S202, hypothetical paths are set up and analyzed by path analysis and structural equation models to reveal potential causal relationships between motor parameters, and a causal relationship graph is generated by model fit verification methods, such as chi-square test. In step S203, a cross-validation method is applied to divide the data into a plurality of subsets, and model training and validation are performed to verify the stability and reliability of the model, respectively. Meanwhile, the statistical significance of the causal relationship is evaluated by using a statistical hypothesis test method, and a verified causal relationship graph is generated. Finally, in step S204, the model is comprehensively optimized, and a final motor causal relationship model is generated by adjusting the model structure and fine-tuning parameters, such as by adopting a gradient descent method. The model accurately reflects the causal relationship in the operation of the motor, and provides a powerful analysis tool for intelligent monitoring, prediction and maintenance of the motor. Through the series of steps, a complete flow from data collection to causal relation model generation is formed, and scientific and systematic methodology is provided for self-adaptive adjustment of the running state of the motor.

Referring to fig. 4, based on a causal relationship model of a motor, a multivariate regression analysis method is adopted to analyze root causes of performance deviation, analyze key parameters, identify key factors affecting the performance of the motor, and the step of generating a root cause analysis result of the performance deviation specifically includes:

s301: based on a motor causality model, calculating standard deviation and mean value of each pair of variables by adopting a Pearson correlation coefficient algorithm, simultaneously carrying out product and subtraction operation on each data point of each pair of variables, accumulating results, dividing the accumulated results by the number of samples to obtain a correlation coefficient, identifying parameters related to motor performance deviation, and further generating a key parameter correlation analysis result;

s302: based on the key parameter correlation analysis result, carrying out centering treatment on parameter data by using a principal component analysis technology, eliminating a mean value, constructing a covariance matrix of parameters, extracting characteristic values and characteristic vectors of the covariance matrix, selecting the characteristic vectors according to the characteristic values in descending order as principal components, and projecting original data onto the principal components to generate a parameter principal component analysis result;

s303: based on the analysis result of the principal components of the parameters, a multiple linear regression analysis technology is applied, a plurality of principal components are selected as independent variables, the motor performance deviation is taken as a dependent variable, meanwhile, coefficients in a regression equation are calculated through a least square method, the products and sums of the independent variables and the dependent variable are calculated, and a regression model is built by using the coefficients to generate a multiple linear regression model;

S304: based on a multiple linear regression model, diagnosing and verifying the model, checking whether the distribution of residual errors accords with normal distribution, simultaneously evaluating the analysis capability of the model to the data variability by using a decision coefficient, and performing F test to generate a performance deviation root cause analysis result;

the pearson correlation coefficient algorithm uses the formula:

；

In the S301 substep, correlation analysis is performed on each variable in the motor causal relationship model by a pearson correlation coefficient algorithm. Motor operating data such as temperature, rotational speed, current, etc., are typically present as time series numerical data. The step first calculates the standard deviation and the mean of each pair of variables, for which the data points of each variable are summed, the mean is calculated, and then the sum of squares of the differences between each data point and the mean is calculated, finally the standard deviation is obtained. Then, the product and subtraction operation is performed on each data point of each pair of variables, and the results are accumulated and divided by the number of samples by one, thereby obtaining the correlation coefficient. In the process, the algorithm calculates each step accurately, and the accuracy of the result is ensured. Identifying parameters associated with motor performance deviations is critical to this step, and by analyzing the magnitude of the correlation coefficient, it is possible to ascertain which parameters have a strong correlation with motor performance deviations. The analysis result is helpful to determine key factors influencing motor performance, and provides basis for subsequent optimization and adjustment.

In the sub-step S302, based on the key parameter correlation analysis result, the parameter data is further analyzed using Principal Component Analysis (PCA) technique. Firstly, the parameter data is subjected to centering treatment, namely the average value of each parameter is subtracted, and the influence of the average value is eliminated. Then, a covariance matrix of the parameters is constructed, which involves calculating covariance between each pair of parameters, reflecting the degree of linear correlation between the parameters. Then, eigenvalues and eigenvectors of the covariance matrix are extracted, and the eigenvalues and eigenvectors are calculated by a numerical method such as a power iteration or jacobian method. According to the magnitude of the feature values, feature vectors are selected in descending order as the main components. Finally, the original data are projected onto the principal components, so that principal component analysis results of parameters are obtained, the step aims to reduce the dimension of the data, the most important information is reserved, and a clear view angle is provided for solving key factors of the running state of the motor.

In the sub-step S303, a model is built using a multiple linear regression analysis technique based on the principal component analysis results of the parameters. A plurality of principal components are selected as independent variables, and motor performance deviation is selected as the independent variable. Coefficients in the regression equation are calculated by least squares, which involves constructing a design matrix, calculating the inverse of the design matrix, and then multiplying it with the dependent variable vector to obtain the coefficients. Then, the products and sums of the independent and dependent variables are used to construct a regression model, the key to the steps is to calculate the coefficients accurately and establish an accurate linear relationship for the quantitative analysis of the relationship between the motor performance bias and the principal components.

Finally, in a substep S304, the multiple linear regression model is diagnosed and validated. First it is checked whether the distribution of residuals corresponds to a normal distribution, which is typically done by plotting a residual map and performing a normal check. The interpretation ability of the model for data variability is then evaluated using a decision coefficient (R), the closer the decision coefficient is to 1, the stronger the interpretation ability of the model. And then F, checking to evaluate whether the overall influence of the independent variable on the dependent variable in the model is obvious. The diagnosis and verification steps ensure the reliability and the effectiveness of the model, and the generated performance deviation root cause analysis results can provide accurate guidance for the optimization of the motor performance.

First, motor operation data including temperature, rotational speed, current, vibration frequency, etc. are collected, recorded once per minute. These data are present in the form of time-series numerical data for subsequent analysis. In step S301, the relationship between the motor operating parameters is analyzed using a pearson correlation coefficient algorithm. And calculating the correlation coefficient of each pair of variables (such as temperature and current), determining the correlation strength through calculation of standard deviation and mean value, identifying the parameter most correlated with the motor performance deviation, and generating a key parameter correlation analysis result. This provides an important basis for identifying the main factors affecting motor performance.

Next, in step S302, principal Component Analysis (PCA) is performed based on the key parameter correlation analysis result. Firstly, data is subjected to centering treatment, then, a covariance matrix of parameters is constructed, and characteristic values and characteristic vectors are extracted from the covariance matrix. According to the descending order of the characteristic values, main characteristic vectors are selected as main components, original data are projected onto the main components, and main component analysis results of parameters are generated, so that the step reduces the dimension of the data, highlights the most important information, and provides clear directions for further analysis.

In step S303, a multiple linear regression analysis technique is applied based on the principal component analysis result. And selecting a main component as an independent variable, taking the motor performance deviation as a dependent variable, calculating coefficients in a regression equation through a least square method, and constructing a multiple linear regression model by using the coefficients. The model can quantitatively describe the relation between the main component and the motor performance deviation, and provides a new view for understanding the motor operation state.

Finally, in step S304, diagnosis and verification of the multiple linear regression model are performed. Checking whether the model residuals conform to normal distribution, and evaluating the interpretation ability and significance of the model using decision coefficients and F-tests. The diagnosis and verification ensure the accuracy and the reliability of the model, and the generated performance deviation root cause analysis result can provide accurate guidance for motor maintenance and performance optimization.

Referring to fig. 5, based on the analysis result of the performance deviation root cause, a differential geometry method in a control theory is adopted to design a dynamic load response mechanism of a motor, and the steps of generating the differential geometry load response control strategy by optimizing the control strategy and adjusting the response parameters and matching the differential load conditions are specifically as follows:

s402: based on a state space performance deviation model, a differential geometry method is adopted, the internal structure of motor performance change is revealed through analyzing the geometric characteristics of a state space, and meanwhile, the behavior of the model at a differential operation point is researched by using a differential geometry tool, so that a differential geometric characteristic analysis result is generated;

S404: and based on the optimized load response control strategy, performing system parameter adjustment and comprehensive test, and generating a differential geometric load response control strategy by adjusting the control parameters and testing the performance under various load conditions.

In the S401 substep, a dynamic model of motor performance is built by adopting a state space modeling method based on the performance deviation root cause analysis result. State space modeling is a mathematical method of describing the behavior of a dynamic system by defining the relationship between system state variables and inputs and outputs to express the dynamics of the system. First, a state variable and an input-output variable are determined according to operation data and performance deviation analysis results of the motor. These variables include speed, current, load, etc. of the motor, with the data format being time series data. Next, dynamic equations are set up to describe the changes in these state variables over time, and how they are affected by motor operating conditions. Dynamic equations typically comprise a set of differential equations describing the rate of change of state variables versus current state and input. In constructing these equations, linear or nonlinear models are used, depending on the complexity of the motor operation. Through the dynamic equations, the performance change of the motor under various operation conditions can be expressed in a mathematical form, and a state space performance deviation model is generated. This model provides a powerful tool for understanding and predicting motor behavior under different operating conditions.

In the sub-step S402, the internal structure of the motor performance is further analyzed using a differential geometry method based on the state space performance bias model. Differential geometry is a mathematical branch that studies the local properties of geometric objects and reveals the inherent structure and properties of complex systems, where the geometric properties of the state space, such as curvature, connectivity, etc., are first analyzed, which requires precise mathematical description and computation of the state space. The differential geometry tool is then used to study the behavior of the model at different operating points (e.g., under different load or speed conditions). This involves calculating the length, curvature, and how the particular path in the state space changes with the change in the operating point. Through the analysis, the internal structure of the motor performance change can be revealed, the motor behavior under specific conditions can be predicted, and the differential geometric characteristic analysis result can be generated. These results are important for understanding the nature of motor performance and for designing a more efficient control strategy.

In the step S403, a feedback control optimization method is adopted to adjust the control strategy of the motor based on the differential geometry analysis result. Feedback control is a common control method that adjusts the control strategy to bring the system output to a desired value. In the step, firstly, a feedback loop and controller parameters are designed according to the performance characteristics and the differential geometry analysis result of the motor so as to optimize the response of the motor to the load change. This includes adjusting the gain, integration and differentiation parameters to accommodate different load conditions. The effectiveness of these control strategies is then verified through simulation and actual testing, ensuring that the motor operates stably and efficiently under a variety of conditions. Through this optimization, the resulting load response control strategy can improve motor performance and reliability while reducing energy consumption and wear.

Finally, in the sub-step S404, based on the optimized load response control strategy, adjustment and comprehensive test of system parameters are performed, and the purpose of the step is to ensure the effectiveness and stability of the control strategy in actual operation. First, parameters of the motor system, such as parameters of the PID controller, are adjusted according to the requirements of the control strategy. Then, testing is performed under various load conditions to verify the effect of the control strategy on motor performance. These tests include long run tests, abrupt load tests, etc., to evaluate the performance of the control strategy under various extreme and conventional conditions. Through the adjustment and the test, the finally generated differential geometric load response control strategy not only improves the predictability and the stability of the motor performance, but also provides important data support and strategy guidance for the optimization and the maintenance of the motor operation.

First, motor operation data including temperature, rotational speed, current, vibration frequency, etc. are collected, recorded once per minute. These data are presented in time series for subsequent analysis. In step S401, a state space modeling method is used to construct a performance deviation model of the motor based on these data. State space models describe the time-dependent changes of state variables (e.g., rotational speed, current) and how these variables are affected by different operating conditions by defining the variables and creating dynamic equations. Next, in step S402, geometrical properties of the state space model, such as curvature and connectivity, and behavior changes of the model at different operating points are analyzed using a differential geometry method. These analyses help to understand in depth the intrinsic structure and law of variation of the motor performance.

Subsequently, in step S403, a feedback control strategy of the motor is designed and optimized based on the differential geometry analysis result. This involves adjusting the feedback loop and controller parameters to optimize the motor's response to different load conditions. Through simulation and actual test, the effectiveness of the control strategy is verified, and the motor can be ensured to run stably and efficiently under various conditions. Finally, in step S404, adjustment of system parameters and comprehensive test are performed. And adjusting the control parameters to adapt to the optimized control strategy, and performing performance test under different load conditions to verify the stability and efficiency of the control strategy.

Referring to fig. 6, based on a differential geometric load response control strategy, the interaction relationship of multiple groups of motors is analyzed and optimized by adopting a graph neural network algorithm, the interaction efficiency among the motors is enhanced by network learning and pattern recognition, and the steps of generating the motor interaction network optimization model are specifically as follows:

s501: based on a differential geometric load response control strategy, a graph neural network algorithm is adopted, a plurality of components in a motor system are defined as nodes, interaction among the components is used as edges, initial characteristics including temperature and current are given to each node, and a graph neural network initial structure is generated;

S502: based on an initial structure of a graph neural network, a graph convolution algorithm is adopted to aggregate the characteristics of each node and the characteristics of adjacent nodes, the characteristics of the current node are updated and the network is trained through accumulation of weighted adjacent node characteristics, an interaction mode in the system is assisted to be captured, and a network model after graph convolution is generated;

s503: based on a network model after graph convolution, mapping analysis is carried out on graph structure data by using a graph embedding technology, and topological relation and characteristic information among nodes in an original graph are reserved through optimizing node representation, so that low-dimensional graph embedding representation is generated;

s504: based on the embedded representation of the low-dimensional graph, the graph neural network is subjected to learning rate adjustment, network layer number and node number optimization, and the generalization and prediction precision of the model are evaluated by applying a cross-validation technology, so that a motor interaction network optimization model is generated.

In S501 substep, complex interactions within the motor system are analyzed using a Graph Neural Network (GNN) algorithm based on a differential geometry load response control strategy. First, a plurality of components in the motor system are defined as nodes, such as individual sensors, control units, etc. of the motor, and interactions between the components are defined as edges. In order to build an initial structure of the graph neural network, each node is endowed with initial characteristics, wherein the characteristics comprise data points such as temperature, current and the like, the data format is usually multidimensional numerical data, and the key of the steps is to accurately define the relation between the node and the edge so as to ensure that the network can reflect the real interaction mode in the motor system. By this means, an initial graphical neural network structure can be generated that captures key features of the motor system and interactions between components, providing a basis for subsequent analysis.

In the substep S502, based on the initial structure of the graph neural network, the graph convolution algorithm is used to further train and optimize the network. The graph convolution algorithm is an efficient algorithm that can process graph structure data and capture complex relationships between nodes. In this process, the features of each node are aggregated with the features of its neighbors, typically by weighting the accumulation of neighbor node features. Thus, the characteristics of the current node can be updated to more accurately reflect the roles and importance of the node in the whole network. The core of this step is to train the network effectively so that it can assist in capturing complex interaction patterns inside the motor system. In this way, a graph-convolved network model can be generated that provides deeper holes that aid in understanding the internal operating mechanisms of the motor system.

In the step S503, map analysis is performed on the map structure data using the map embedding technique based on the network model after the map convolution. The graph embedding technology is a powerful method, can map graph structure data into a low-dimensional space, and simultaneously reserves topological relation and characteristic information among nodes in an original graph, and the key of the steps is to optimize node representation so as to reserve key information and reduce complexity of the data. In this way, a low-dimensional graph-embedded representation can be generated that makes analysis of the motor system more efficient and intuitive, facilitating further data analysis and decision-making.

Finally, in a substep S504, a graph neural network based on the low-dimensional graph embedded representation is further optimized. The method comprises the steps of adjusting key parameters such as learning rate, optimizing network layer number and node number, and the like, so as to improve the performance of the network model. Meanwhile, the generalization and prediction precision of the model are evaluated by applying a cross-validation technology, so that the model can show good performance on different data sets, the steps aim is to generate an optimized motor interaction network model, and the model not only has a solid foundation in theory, but also shows high precision and applicability in practical application. By the method, a powerful analysis tool can be provided for intelligent monitoring and predictive maintenance of the motor, and the running efficiency and stability of the motor are improved.

First, a Graph Neural Network (GNN) algorithm is adopted to deeply analyze the complex interactions inside the motor system. In step S501, the components of the motor system are defined as nodes of the network, and interactions between the components act as edge connection nodes. Each node is assigned an initial characteristic, such as temperature, current, etc., to reflect the operating state of the motor. In this way, an initial structure of the neural network is constructed, which captures the key features of the motor system and interactions between the components.

Then, in step S502, the graph neural network is trained and optimized by the graph convolution algorithm, in this process, the feature of each node is aggregated with the feature of its neighboring node, and the feature of the current node is updated by weighting the accumulation of the features of the neighboring nodes. Thus, the graphic neural network can more accurately capture and represent the complex interaction mode in the motor system, and a network model after the graphic convolution is generated.

Next, in step S503, the network model after the graph rolling is further analyzed by using the graph embedding technology. The graph embedding technology maps the graph structure data into a low-dimensional space, meanwhile, the topological relation and characteristic information among nodes are reserved, the steps generate the low-dimensional graph embedding representation, the complexity of a motor system is effectively simplified, meanwhile, key information is reserved, and convenience is provided for subsequent analysis and decision.

Finally, in step S504, the neural network of the graph based on the embedded representation of the low-dimensional graph is optimized, the learning rate is adjusted, the number of layers and nodes of the network are optimized, and the generalization and the prediction accuracy of the model are evaluated by applying a cross-validation technique. The optimization ensures that the model can show good performance on different data sets, and the generated motor interaction network optimization model not only has a solid foundation in theory, but also shows high precision and applicability in practical application.

Referring to fig. 7, based on a motor interaction network optimization model, learning and optimization of adaptive load distribution are performed by adopting a depth deterministic strategy gradient algorithm, and a load distribution strategy is adjusted by combining motor performance parameters and real-time feedback data, so that the steps of generating the adaptive load optimization strategy are specifically as follows:

s601: based on a motor interaction network optimization model, adopting a graph neural network algorithm, identifying key network connection and characteristics affecting system performance by learning characteristics of a plurality of nodes in a network and relationships among the nodes, and simultaneously analyzing an interaction structure among motors to generate network characteristic comprehensive analysis;

s602: based on network characteristic comprehensive analysis, a depth deterministic strategy gradient algorithm is applied, a network structure is analyzed by combining deep learning, meanwhile, loads of a motor system are distributed by utilizing a strategy gradient method, influence of a differentiated load distribution scheme on system performance is evaluated, automatic learning is performed, an initial load distribution strategy is built, and a preliminary self-adaptive load distribution model is generated;

s603: based on the preliminary self-adaptive load distribution model, performing iterative optimization of reinforcement learning, performing multiple tests in a simulation environment, adjusting a load distribution strategy according to test results, and simultaneously referring to real-time data feedback and multiple dynamic changes, perfecting the load distribution strategy through loop tests and adjustment, and further generating an optimally-adjusted load distribution model;

S604: based on the load distribution model with optimized adjustment, a performance test and evaluation method is adopted to test the model under different working conditions and load changes, evaluate the performance and stability, and simultaneously perform iterative adjustment of strategy parameters, including adjustment of the key parameters of learning rate and return discount rate in an algorithm, optimize the load distribution capacity of the model, and generate a self-adaptive load optimization strategy.

In the step S601, by adopting the graph neural network algorithm based on the motor interaction network optimization model, specific operations include first constructing a data format of the motor interaction network, which generally includes characteristic data of each node (representing a motor) in the network, such as current, voltage, temperature, rotation speed, and the like, and connection relations between the nodes, that is, interaction relations between the motors. A Graph Neural Network (GNN) algorithm is used to process these data, learning the characteristics of the nodes and the relationships between the nodes through the graph structure. In GNN, the characteristics of each node are updated by aggregating the characteristics of its neighboring nodes, so that the interaction relationship between motors can be effectively captured. Through multi-layer feature aggregation, GNNs can learn key network connections and features that affect system performance. Key parameters in the GNN model include the dimensions of node features, the number of layers, neighbor aggregation functions, etc. The GNN model, trained, can identify motors and their connections that have significant impact on system performance, and this information is used to generate a network characteristics analysis report detailing key factors that affect motor system performance and used to guide subsequent load distribution strategies.

In S602 substep, a Depth Deterministic Policy Gradient (DDPG) algorithm is applied to operate based on the network characteristic comprehensive analysis. DDPG is a reinforcement learning method combining deep learning and strategy gradient, and is suitable for the problem of continuous action space. In the step, DDPG is used to analyze the motor network structure and perform load distribution. The algorithm first initializes a policy network for generating load allocation decisions and a value function network for evaluating the merits of the decisions. By interacting with the environment, i.e. applying a load distribution scheme in the simulated motor network, the algorithm can collect the effect of load distribution and adjust the policy network and the value function network accordingly. Key parameters in DDPG include learning rate, exploring noise, etc. Through continuous iteration, the DDPG can automatically learn and build an initial load distribution strategy, and generate a preliminary self-adaptive load distribution model. The model can dynamically adjust load distribution according to the real-time condition of the motor network so as to optimize the overall performance of the system.

In the sub-step S603, based on the preliminary adaptive load distribution model, iterative optimization of reinforcement learning is performed, and the process includes performing a plurality of experiments in a simulation environment, and adjusting the load distribution strategy according to the experimental results. In the step, the DDPG algorithm is continued to be used, but more environmental variables and real-time data are introduced, such as real-time load of the motor, temperature change, etc. And after each test, adjusting the strategy network according to the test result and the real-time data so as to optimize load distribution. In this process, the parameter adjustment of the algorithm is particularly critical, including adjustment of learning rate, fine adjustment of exploration strategy, and the like. Through repeated cyclic tests and adjustment, the load distribution strategy is gradually perfected, and an optimally-adjusted load distribution model is generated. The model can be more accurately adapted to dynamic changes in a motor network, and the performance and stability of the whole system are optimized.

In the sub-step S604, a performance test and evaluation method is adopted based on the load distribution model of the optimization adjustment. In the step, the model is tested under different working conditions and load changes at first, and the performance and stability of the model are evaluated. This includes applying the model to different motor network configurations and load conditions, observing the response and tuning capabilities of the model. And simultaneously, carrying out iterative adjustment on strategy parameters, wherein key parameters comprise learning rate, return discount rate and the like. Through these tests and adjustments, the load distribution capabilities of the model are further optimized, generating an adaptive load optimization strategy. The strategy can dynamically adjust load distribution according to the real-time state and performance index of the network so as to realize the optimal operation efficiency and stability of the motor system.

In the sub-step S601, it is assumed that a motor interaction network contains 10 motors, each motor having characteristic data: current (a), voltage (V), temperature (°c), and rotational speed (RPM). For example, the motor has data: current 10A, voltage 220V, temperature 50 ℃, rotational speed 1500RPM. In the constructed Graph Neural Network (GNN) model, these features are input as the features of the nodes. GNN updates the feature representation of each motor by aggregating features of neighboring motors, such as interactions between motor 1 and motor 2. During model training, for example, current and rotational speed are found to be particularly critical to network performance. By training, GNN identified that the connection of motor 3 and motor 7 had significant impact on overall network performance, and generated a network characteristic analysis report containing these key features and connections.

In S602 substep, based on the analysis described above, the initial policy network randomly gives a load distribution scheme using a Depth Deterministic Policy Gradient (DDPG) algorithm, distributing higher loads to more performance stable motors. For example, the model initially distributes higher loads to the motor 3 and the motor 7. By interacting with the simulated environment, DDPG constantly adjusts the strategy, finding that distributing more load to motor 5 and motor 6 can improve overall efficiency. After multiple iterations, a preliminary adaptive load distribution model is formed.

In a sub-step S603, the model is further optimized. For example, considering the temperature variation and real-time load of the motor in the simulation environment, it was found that the performance of the motor 3 is unstable under high temperature conditions. Therefore, the model reduces the load distribution to the motor 3 and increases the loads to the motor 2 and the motor 4 by reinforcement learning adjustment strategies. The dynamic adjustment is perfected after multiple experiments, and an optimally adjusted load distribution model is generated.

In a substep S604, the model is tested under different operating conditions. For example, when motors in the network face both high temperatures and high loads, the model can quickly adjust strategies, such as reducing load sharing to motor 3 and motor 7, increasing the load to motor 1 and motor 4. Through testing, the model is found to maintain good performance and stability in most cases. On the basis, the learning rate and the return discount rate are further adjusted so as to improve the response speed and the adaptability of the model. Finally, an adaptive load optimization strategy capable of dynamically adjusting load distribution according to real-time data and working conditions is generated, and the running efficiency and stability of the whole motor network are improved.

Referring to fig. 8, based on the adaptive load optimization strategy, a real-time feedback adjustment algorithm is adopted to dynamically adjust a control system of a motor, operation data is monitored in real time, control parameters are optimized according to feedback information, and the step of generating a dynamic adjustment control command specifically includes:

s701: based on a self-adaptive load optimization strategy, a time sequence analysis method is adopted to analyze the time sequence of motor operation data, identify trend and periodical change, and simultaneously analyze abnormal points in the data by using an isolated forest algorithm to monitor the operation condition of the motor in real time so as to generate real-time performance indexes;

s702: based on real-time performance indexes, a state estimation method is adopted, a Kalman filter is utilized to process noise and uncertainty factors in motor state data, the dynamic state of the motor is continuously estimated, a Bayesian network is simultaneously utilized to carry out probabilistic reasoning, the motor state change is analyzed, and a state estimation analysis result is generated;

S704: based on the optimized control parameters, a predictive control method is adopted to analyze the current state of the motor, simultaneously predict the performance in the future short period, analyze the batch historical data and the instant feedback information, identify the potential running trend and the performance change, and then adjust the control parameters by combining a proportional-integral-differential adjustment method, and dynamically adjust the control parameters by circularly comparing the difference between the current output and the expected target in the adjustment process to generate a dynamic adjustment control instruction.

In the step S701, motor operation data is processed by a time series analysis method and an isolated forest algorithm. The time series of motor operation data includes readings of current, voltage, temperature and rotational speed, which are recorded in the form of time stamps, forming a continuous time series. Using time series analysis methods, the data is first denoised and normalized, and then a moving average or exponential smoothing technique is applied to identify trends and periodic changes in the data. For example, by calculating a 7-day moving average of motor current, a trend of current fluctuation over time can be observed. Next, outliers are detected using an isolated forest algorithm. The isolated forest algorithm isolates each data point by building multiple decision trees, outliers are typically more easily isolated due to the nature of the minority and the difference from most data points. By setting a threshold, data points that differ significantly from the normal operating mode, such as an abnormally high current for a day, are identified. These analysis results are used to monitor the operation of the motor in real time, generating real-time performance metrics such as average current, maximum and minimum temperatures, etc., which are critical to assessing the health of the motor in real time.

In the step S702, the motor state data is processed using a state estimation method and a kalman filter based on the real-time performance index. The state estimation method is used for estimating the real state of the motor from the observation data with noise. The kalman filter is an efficient linear dynamic system state estimation method to predict the next state by taking into account the previous state and the current observed data. Here, the kalman filter is used to process motor state data, such as temperature and current, reducing noise due to measurement errors and external disturbances. Meanwhile, a Bayesian network is used for carrying out probability reasoning and analyzing the change of the state of the motor. A bayesian network is a probabilistic based graph model that can handle uncertainty and conditional dependencies. Through the Bayesian network, the change trend of the motor state can be deduced, for example, the current increase caused by the rise of the motor temperature. These analysis results are integrated to generate state estimation analysis results, which are important for understanding the current operating conditions of the motor and predicting future state changes.

In the step S703, the motor is adjusted by an adaptive control method based on the state estimation analysis result. The adaptive control method adjusts the control strategy according to the difference between the actual and expected performances of the motor to cope with the environmental and the changes of the performances of the motor itself. This includes adjusting neural network parameters based on real-time feedback of motor status using adaptive neural network control techniques. For example, if the actual speed of the motor is less than the desired speed, the adaptive control system may adjust an input parameter (e.g., current) to attempt to increase the speed. The neural network is constantly learned and adjusted in the process to better adapt to the dynamic change of the motor. This adaptive mechanism enables the control system to flexibly cope with various operating conditions and state changes, generating optimized control parameters.

In the sub-step S704, the motor is managed using a predictive control method based on the optimized control parameters. The predictive control method optimizes the control strategy by analyzing the current state of the motor and predicting performance in the future short term. This involves analysis of batch historical data and immediate feedback information to identify potential operational trends and performance changes. For example, by analyzing temperature and current data for the past week, the predictive control system may anticipate motor overheating under high load conditions. And the predictive control system is combined with a proportional-integral-derivative (PID) regulation method to accurately regulate control parameters. The PID controller adjusts the motor input (e.g., current) to reduce the difference between the current output (e.g., rotational speed) and the desired target. The method for dynamically adjusting the control parameters enables the motor to reach the expected running state more accurately, generates a dynamically adjusted control instruction, and ensures that the motor can maintain the optimal performance and stability under various working conditions.

In the sub-step S701, motor operation data is contemplated to include current, voltage, temperature and speed readings per minute. Taking motor current as an example, the data format is a current reading of several consecutive days per minute, e.g., the first day of reading is 10A, 11A, 10.5A, etc. The data is first processed using a moving average method, for example, an average of 7 minutes is calculated to smooth the data. Then, outliers were detected by the isolated forest algorithm, such as in the data with a normal range of 10A-12A, with one 30A reading marked as outliers. These analyses help identify trends in motor operation, such as gradual increases in current, and abnormal conditions, such as sudden current peaks. The generated real-time performance indicators, such as average current and maximum current anomalies, are critical to assessing the health of the motor and preventing potential faults.

In the S702 substep, it is assumed that the state data of the motor includes noise measurements of temperature and current. These data are processed using a kalman filter, for example, if the temperature sensor measurement fluctuates around 50 ℃, but there are occasional peaks or drops, the kalman filter will smooth out these noise, providing a more accurate temperature estimate. Meanwhile, a bayesian network is used to analyze the probability relation of motor states, such as the trend of current increase caused by temperature rise. The analysis results generate a state estimation analysis report, which provides basis for understanding the current state of the motor and predicting future changes.

In a sub-step S703, the motor is adjusted using an adaptive control method based on the state estimation analysis. For example, if the actual rotational speed of the motor continues to be lower than the set point 1500 RPM, the adaptive control system adjusts the input current based on this difference. The adaptive neural network control technique works here, and network parameters are continuously adjusted through real-time feedback of motor states, so that more accurate control is realized. This adaptive adjustment generates optimized control parameters, such as current adjustment to 11A to increase rotational speed, enhancing the performance stability of the motor in a constantly changing environment.

In the sub-step S704, the motor is adjusted using a predictive control method based on the optimized control parameters. For example, the predictive control system analyzes the current motor temperature and current data and, in combination with the historical data, predicts the predicted performance of the motor over the next few hours. If the prediction shows motor = overheat, the system adjusts the input parameters, such as reducing the current, by PID control to prevent overheating. This predictive and responsive mechanism allows for more efficient adaptation to anticipated operating conditions, improved operating efficiency, and reduced failure rates. By constantly comparing actual output with the desired target, the control system dynamically adjusts the control parameters to generate dynamically adjusted control commands to ensure that the motor maintains optimal performance under a variety of conditions.

Referring to fig. 9, the motor running state adaptive adjustment system is configured to execute the above motor running state adaptive adjustment method, where the system includes a probability network construction module, a causal reasoning module, a performance deviation analysis module, a load response policy module, an interactive network optimization module, and an adaptive load adjustment module;

The causal reasoning module is used for analyzing causal relation among motor operation parameters by using a structural equation model based on a motor operation probability network, and generating a motor causal relation model by establishing a statistical model and carrying out hypothesis verification and analyzing causal relation among parameters;

the load response strategy module designs a load response strategy by using a differential geometry control method based on the performance deviation root cause analysis result, and simultaneously analyzes the geometrical characteristics of a state space and adjusts the control strategy to match differential load conditions so as to generate a differential geometry load response control strategy;

the interaction network optimization module analyzes interaction of internal components of the motor system by adopting a graph neural network algorithm, and optimizes interaction efficiency among motors by learning characteristics and relations of nodes in the network to generate a motor interaction network optimization model;

The probability network constructed by the Bayesian network effectively reveals the complex interrelation between the motor operation parameters, and greatly improves the accuracy and maintenance efficiency of fault prediction. The causal reasoning module further deepens understanding of causal links among the parameters, and provides a solid scientific basis for formulating more targeted maintenance strategies and fault preventive measures. The multivariate regression analysis method of the performance deviation analysis module enables key influencing factors to be accurately identified and quantified, and therefore overall performance and reliability of the motor are improved. The differential geometric control strategy designed by the load response strategy module provides highly adaptive control for the motor under different load conditions, and ensures the optimal performance of the motor under various working conditions. The graph neural network algorithm of the interaction network optimization module optimizes the interaction efficiency inside the motor system, and enhances the stability and the energy utilization efficiency of the system. The depth deterministic strategy gradient algorithm of the self-adaptive load adjusting module enables motor load distribution to be more intelligent, dynamic adjustment is carried out according to real-time data, energy efficiency is further improved, energy consumption is reduced, and service life of the motor is prolonged.

The present invention is not limited to the above embodiments, and any equivalent embodiments which can be changed or modified by the technical disclosure described above can be applied to other fields, but any simple modification, equivalent changes and modification made to the above embodiments according to the technical matter of the present invention will still fall within the scope of the technical disclosure.

Claims

1. The self-adaptive motor running state adjusting method is characterized by comprising the following steps of:

modeling the probability relation among parameters by adopting a Bayesian network construction algorithm based on the existing motor operation data, refining the probability dependency relation among key parameters by a statistical learning method, and carrying out probability analysis to generate a motor operation probability network;

based on the motor operation probability network, carrying out causal relationship inference analysis by adopting a structural equation model, analyzing causal relationship among various parameters by a data driving method, verifying the reliability of the relationship, and generating a motor causal relationship model;

analyzing the root cause of the performance deviation by adopting a multivariate regression analysis method based on the motor causal relationship model, analyzing key parameters, identifying key factors influencing the motor performance, and generating a performance deviation root cause analysis result;

Based on the analysis result of the performance deviation root cause, adopting a differential geometric method in a control theory to design a dynamic load response mechanism of the motor, and matching differential load conditions through optimizing a control strategy and adjusting response parameters to generate a differential geometric load response control strategy;

based on the differential geometric load response control strategy, adopting a graph neural network algorithm to analyze and optimize the interaction relation of a plurality of groups of motors, and strengthening the interaction efficiency among the motors through network learning and pattern recognition to generate a motor interaction network optimization model;

based on the motor interaction network optimization model, a depth deterministic strategy gradient algorithm is adopted to learn and optimize self-adaptive load distribution, and a load distribution strategy is adjusted by combining motor performance parameters and real-time feedback data to generate a self-adaptive load optimization strategy;

based on the self-adaptive load optimization strategy, a real-time feedback adjustment algorithm is adopted to dynamically adjust a control system of the motor, operation data are monitored in real time, control parameters are optimized according to feedback information, and a dynamic adjustment control instruction is generated.

2. The method of claim 1, wherein the motor operation probability network comprises probability distribution data of load rate, temperature change and current fluctuation of the motor, the motor causal relationship model comprises direct causal relationship between factors of temperature change and current fluctuation and motor performance, the performance deviation root cause analysis result comprises a root cause list of key parameter deviation, operation condition abnormality and environmental factors, the differential geometric load response control strategy comprises load change response adjustment, performance optimization path and control parameter setting, the motor interaction network optimization model comprises synchronous efficiency, load sharing strategy and fault response coordination among the motors, the adaptive load optimization strategy comprises a load distribution algorithm, response time optimization and energy consumption management strategy, and the dynamic adjustment control instruction comprises motor start-stop strategy, operation mode selection and fault prevention adjustment.

3. The method for adaptively adjusting the running state of a motor according to claim 1, wherein modeling of probability relationships between parameters is performed by using a bayesian network construction algorithm based on existing motor running data, probability dependency relationships between key parameters are refined by using a statistical learning method, probability analysis is performed, and the step of generating a motor running probability network is specifically as follows:

based on the existing motor operation data, a Bayesian network structure learning algorithm is adopted to construct a network structure, the connection mode of a plurality of nodes in a network is determined by evaluating the relation among a plurality of parameters, and meanwhile, key parameters are selected to generate a Bayesian network structure;

based on the Bayesian network structure, analyzing the mutual information quantity among parameters by using a mutual information calculation method, revealing the mutual dependency relationship by calculating the mutual information value of each pair of parameters, and constructing a parameter dependency relationship graph;

based on the parameter dependency graph, a Bayesian network parameter learning algorithm is applied to quantitatively analyze the parameter relationship in the network, the conditional probabilities of a plurality of nodes are estimated, a probability dependency model among a plurality of parameters is constructed, and a Bayesian network probability model is generated;

Based on the Bayesian network probability model, a network optimization and calibration method is adopted, connection is added or removed to a network structure, the performance of the model on various data sets is tested by using a cross verification method, the model precision is optimized, and then a motor operation probability network is generated.

4. The method for adaptively adjusting the running state of a motor according to claim 1, wherein based on the running probability network of the motor, a causal relationship is inferred and analyzed by a structural equation model, causal relationships among various parameters are analyzed by a data driving method, reliability of the relationships is verified, and the step of generating the causal relationship model of the motor is specifically as follows:

based on the motor operation probability network, a Bayesian network analysis method is adopted, nodes in the network are selected to represent the differentiated operation parameters of the motor, meanwhile, probability relations among the nodes are defined through probability distribution estimation, potential connection among the nodes is detected and established through dependency relation learning, and then a network model structure is optimized, adjusted and confirmed through a network structure, so that a motor operation preliminary probability model is generated;

based on the motor operation preliminary probability model, adopting path analysis, setting an assumed path and distributing path coefficients, then using a structural equation model to analyze the path, simultaneously verifying the strength of causal relation of the assumed path through model fitting, evaluating direct and indirect influences among a plurality of variables in the model, and generating a potential causal relation graph among motor parameters;

Based on potential causal relation graphs among motor parameters, cross verification is applied, data are divided into a plurality of subsets, training and verification of a model are respectively carried out by the subsets, stability and reliability of the model on the plurality of data sets are verified, meanwhile, statistical hypothesis verification is used, and the verified causal relation graphs are generated by determining statistics of horizontal evaluation causal relation;

based on the verified causal relationship graph, model comprehensive optimization is adopted, the model is reconfigured, model efficiency is improved, parameter fine adjustment is carried out, and the causal relationship in motor operation is reflected through adjustment of weights and deviations, so that a motor causal relationship model is generated.

5. The method for adaptively adjusting the running state of a motor according to claim 1, wherein the method for analyzing the root cause of the performance deviation by adopting a multivariate regression analysis method based on the causal relation model of the motor is characterized in that the key parameters are analyzed, the key factors influencing the performance of the motor are identified, and the step of generating the root cause analysis result of the performance deviation is specifically as follows:

based on the motor causality model, calculating standard deviation and mean value of each pair of variables by adopting a Pearson correlation coefficient algorithm, simultaneously carrying out product and subtraction operation on each data point of each pair of variables, accumulating results, dividing the accumulated results by the number of samples, subtracting one to obtain a correlation coefficient, identifying parameters related to motor performance deviation, and further generating a key parameter correlation analysis result;

Based on the key parameter correlation analysis result, carrying out centering treatment on parameter data by using a principal component analysis technology, eliminating a mean value, constructing a covariance matrix of parameters, extracting characteristic values and characteristic vectors of the covariance matrix, selecting the characteristic vectors according to a descending order of the characteristic values as principal components, and projecting original data onto the principal components to generate a parameter principal component analysis result;

based on the analysis result of the parameter principal components, a multiple linear regression analysis technology is applied, a plurality of principal components are selected as independent variables, the motor performance deviation is used as a dependent variable, meanwhile, the coefficient in a regression equation is calculated through a least square method, the product sum of the independent variables and the dependent variable is summed, and a regression model is built by using the coefficient, so that a multiple linear regression model is generated;

based on the multiple linear regression model, diagnosing and verifying the model, checking whether the distribution of residual errors accords with normal distribution, simultaneously evaluating the analysis capability of the model to the data variability by using a decision coefficient, and performing F test to generate a performance deviation root cause analysis result;

the pearson correlation coefficient algorithm adopts the formula:

；

wherein,and->For paired data points >And->For the mean value of the corresponding variables>Is the correlation coefficient that is calculated.

6. The method for adaptively adjusting the running state of a motor according to claim 1, wherein the step of designing a dynamic load response mechanism of the motor by adopting a differential geometry method in a control theory based on the analysis result of the performance deviation root causes, and matching differential load conditions by optimizing a control strategy and adjusting response parameters, and generating the differential geometry load response control strategy is specifically as follows:

based on the analysis result of the performance deviation root cause, a state space modeling method is adopted, a dynamic equation is established by defining the relation between the motor performance parameter and the state variable, and the performance change of the motor under various operation conditions is expressed in a mathematical form, so that a state space performance deviation model is generated;

based on the state space performance deviation model, a differential geometry method is adopted, the internal structure of motor performance change is revealed through analyzing the geometric characteristics of a state space, and meanwhile, the behavior of the model at a differential operation point is researched by using a differential geometry tool, so that a differential geometric characteristic analysis result is generated;

based on the analysis result of the differential geometric characteristics, adopting a feedback control optimization method to adjust a control strategy to match various load conditions, and simultaneously designing a feedback loop and adjusting parameters of a controller to optimize the response of a motor to load change to generate an optimized load response control strategy;

And based on the optimized load response control strategy, performing system parameter adjustment and comprehensive test, and generating a differential geometric load response control strategy by adjusting control parameters and testing performance under various load conditions.

7. The method for adaptively adjusting the running states of motors according to claim 1, wherein based on the differential geometric load response control strategy, the interactive relation among a plurality of groups of motors is analyzed and optimized by adopting a graph neural network algorithm, the interactive efficiency among the motors is enhanced by network learning and pattern recognition, and the step of generating a motor interactive network optimization model is specifically as follows:

based on the differential geometric load response control strategy, a graph neural network algorithm is adopted, a plurality of components in a motor system are defined as nodes, interaction among the components is used as edges, initial characteristics including temperature and current are given to each node, and a graph neural network initial structure is generated;

based on the initial structure of the graph neural network, the features of each node and the features of adjacent nodes are aggregated by adopting a graph convolution algorithm, the features of the current node are updated and the network is trained by adding up the features of weighted adjacent nodes, and an interaction mode in the system is assisted to be captured, so that a network model after graph convolution is generated;

Based on the network model after graph convolution, mapping analysis is carried out on graph structure data by utilizing a graph embedding technology, and topological relation and characteristic information among nodes in an original graph are reserved through optimizing node representation, so that low-dimensional graph embedding representation is generated;

based on the low-dimensional graph embedded representation, the graph neural network is subjected to learning rate adjustment, network layer number and node number optimization, and the generalization and prediction precision of the model are evaluated by applying a cross-validation technology, so that a motor interaction network optimization model is generated.

8. The method for adaptively adjusting the running state of a motor according to claim 1, wherein the step of adaptively adjusting the load distribution strategy by adopting a depth deterministic strategy gradient algorithm to learn and optimize the self-adaptive load distribution based on the motor interaction network optimization model and combining the motor performance parameter and real-time feedback data to generate the self-adaptive load optimization strategy is specifically as follows:

based on the motor interaction network optimization model, adopting a graph neural network algorithm, identifying key network connection and characteristics affecting system performance by learning characteristics of a plurality of nodes in a network and relationships among the nodes, and simultaneously analyzing an interaction structure among motors to generate network characteristic comprehensive analysis;

Based on the network characteristic comprehensive analysis, a depth deterministic strategy gradient algorithm is applied, a network structure is analyzed by combining deep learning, meanwhile, loads of a motor system are distributed by utilizing a strategy gradient method, influence of a differentiated load distribution scheme on system performance is evaluated, automatic learning is performed, an initial load distribution strategy is built, and a preliminary self-adaptive load distribution model is generated;

based on the preliminary self-adaptive load distribution model, performing iterative optimization of reinforcement learning, performing multiple tests in a simulation environment, adjusting a load distribution strategy according to test results, and simultaneously referring to real-time data feedback and multiple dynamic changes, perfecting the load distribution strategy through cyclic tests and adjustment, and further generating an optimally-adjusted load distribution model;

based on the load distribution model of optimization adjustment, a performance test and evaluation method is adopted to test the model under differential working conditions and load change, evaluate performance and stability, and simultaneously perform iterative adjustment of strategy parameters, including adjustment of key parameters of learning rate and return discount rate in algorithm, optimize the load distribution capacity of the model, and generate a self-adaptive load optimization strategy.

9. The method for adaptively adjusting the running state of a motor according to claim 1, wherein the step of dynamically adjusting a control system of the motor by using a real-time feedback adjustment algorithm based on the adaptive load optimization strategy, monitoring running data in real time and optimizing control parameters according to feedback information, and generating a dynamic adjustment control command comprises the following steps:

based on the self-adaptive load optimization strategy, a time sequence analysis method is adopted to analyze the time sequence of the motor operation data, the trend and the periodic variation are identified, and meanwhile, abnormal points in the data are analyzed by using an isolated forest algorithm, so that the operation condition of the motor is monitored in real time, and a real-time performance index is generated;

based on the real-time performance index, a state estimation method is adopted, a Kalman filter is utilized to process noise and uncertainty factors in motor state data, the dynamic state of the motor is continuously estimated, a Bayesian network is simultaneously utilized to carry out probabilistic reasoning, the motor state change is analyzed, and a state estimation analysis result is generated;

based on the state estimation analysis result, adopting an adaptive control method, adjusting a control strategy according to the difference between the actual performance and the expected performance of the motor by referring to the adaptive control, and simultaneously adjusting a neural network parameter according to the real-time feedback of the state of the motor by using the adaptive neural network control to generate an optimized control parameter;

Based on the optimized control parameters, a predictive control method is adopted to analyze the current state of the motor, simultaneously predict the performance in the future short period, analyze the batch historical data and the instant feedback information, identify the potential running trend and the performance change, and then adjust the control parameters by combining a proportional-integral-differential adjustment method, wherein the difference between the current output and the expected target is compared in a circulating manner in the adjustment process, dynamically adjust the control parameters, and generate a dynamic adjustment control instruction.

10. The motor running state self-adaptive adjusting system is characterized in that the motor running state self-adaptive adjusting method according to any one of claims 1-9 comprises a probability network construction module, a causal reasoning module, a performance deviation analysis module, a load response strategy module, an interactive network optimization module and a self-adaptive load adjustment module;