CN114417501A - Airborne deployment-oriented health management predictive modeling method - Google Patents

Abstract

The embodiment of the invention discloses an airborne deployment-oriented health management predictive modeling method, which comprises the following steps: step 1, performing data processing and model training on airborne data and ground data on a ground platform to obtain a plurality of training models; step 2, evaluating, selecting and deciding the training model obtained in the step 1 in a principle prototype to obtain a training model meeting the task requirement; step 3, deploying the training model meeting the task requirement obtained in the step 2 in an airborne vehicle; obtaining an airborne full-period PHM modeling strategy through the steps 1-3; further comprising: and (3) constructing an onboard PHM operator library, and combing according to the stages from the step 1 to the step 3 to obtain the PHM operator library. The technical scheme provided by the embodiment of the invention solves the problem that the complete fault prediction and health management (PHM for short) predictive maintenance specification of the aviation airborne and ground are not uniform.

Description

Airborne deployment-oriented health management predictive modeling method

Technical Field

The invention relates to the field of aviation system reliability, in particular to ground/airborne health management, and particularly relates to an airborne deployment-oriented health management predictive modeling method.

Background

In the aviation health management and fault diagnosis system, the airborne and ground platforms have different computing resources and constraint conditions, and different stages of Prediction and Health Management (PHM) processing need to operate under different hardware environments; specifically, the ground platform has unlimited computing performance, storage space, abundant data analysis toolkits and the like, and the airborne environment has limited computing resources, low required power consumption and small storage space.

At present, a complete set of PHM modeling method suitable for aviation airborne deployment and a corresponding operator library do not exist. Therefore, a set of complete modeling strategies needs to be established for the PHM of the aviation system, standard data acquisition and analysis are achieved, the advantages of the ground platform are utilized to perform data processing and model training, airborne operation time and accuracy are guaranteed during model evaluation and model decision, and finally the selected optimal model is deployed at an airborne terminal.

Disclosure of Invention

The purpose of the invention is as follows: the invention provides a health management predictive modeling method facing airborne deployment, aiming at the airborne PHM deployment requirement, considering the difference between the ground and the airborne environment, and finally deploying in an airborne full-period PHM modeling strategy by performing data processing and model training at the ground stage, performing evaluation selection and decision in a principle prototype.

The technical scheme of the invention is as follows: the embodiment of the invention provides an airborne deployment-oriented health management predictive modeling method, which comprises the following steps:

step 1, performing data processing and model training on airborne data and ground data on a ground platform to obtain a plurality of training models;

step 2, evaluating, selecting and deciding the training model obtained in the step 1 in a principle prototype to obtain a training model meeting the task requirement;

step 3, deploying the training model meeting the task requirement obtained in the step 2 in an airborne vehicle;

and (4) obtaining an airborne full-period PHM modeling strategy through the steps 1-3.

Optionally, in the method for predictively modeling health management for on-board deployment as described above, the step 1 includes: a data importing stage, a data processing stage and a model training stage which operate on a ground platform; the method specifically comprises the following steps:

step 1.1, a data import phase, comprising: importing various structured data and unstructured data in a ground high-performance platform;

step 1.2, a data processing stage, comprising: a data exploration and analysis sub-stage, a data preprocessing sub-stage and a characteristic engineering sub-stage;

step 1.3, a model training phase, comprising: through model training based on physical or empirical models and model training based on data driving, a plurality of algorithm models for specific PHM tasks are obtained through training.

Optionally, in the onboard deployment-oriented health management predictive modeling method as described above, in step 1.2,

a data exploration analysis sub-phase comprising: visually exploring and simply mining the imported data to obtain the number of samples, the number of characteristics, data distribution characteristics, characteristic trend and correlation;

the data preprocessing sub-stage is used for improving the data quality and comprises the following steps: data cleaning, data denoising and data standardization;

a feature engineering sub-phase comprising: feature mining, feature selection and feature extraction are carried out on the data to extract features which are most useful for specific tasks (such as fault diagnosis and life prediction).

Optionally, in the method for predictively modeling health management for on-board deployment as described above, the PHM task in step 1.3 includes: state monitoring, fault diagnosis, fault prediction and residual life prediction;

for a specific PHM task type, the method for analyzing and deconstructing the problem is as follows:

in the method 1, according to cognition on an analysis object, when the object has a physical or empirical model which is easy to solve, a corresponding physical model is established for the object aiming at a specific task so as to carry out model training and solution;

in the mode 2, when the object structure is complex and failure or degradation mechanisms are difficult to obtain, a data-driven method is adopted, and algorithms are selected and model training is carried out by using methods such as corresponding machine learning and statistical analysis;

in the model training stage of step 1.3, multiple algorithms are adopted to train multiple models, so as to obtain multiple trained algorithm models.

Optionally, in the method for predictively modeling health management for on-board deployment as described above, the step 2 includes: a model evaluation stage and a model decision stage of a principle prototype which is similar to an airborne software and hardware environment; the method specifically comprises the following steps:

step 2.1, in a model evaluation stage, evaluating a plurality of trained algorithm models in a principle prototype, selecting a plurality of appropriate model evaluation indexes according to the types of the algorithm models and the task requirements, and operating a plurality of algorithm models trained on the ground in the principle prototype to obtain a model evaluation table;

and 2.2, in the model decision stage, making a model decision rule according to the specified task requirement, and carrying out final decision on the model by combining the obtained model evaluation table to select the optimal model for the specific task.

Optionally, in the method for predictively modeling health management for on-board deployment as described above, the step 3 includes:

and in the model deployment stage, packaging the selected optimal model by adopting a language supported by airborne hardware, and deploying the packaged optimal model to an airborne terminal.

Optionally, in the method for predictively modeling health management for on-board deployment, the method further includes:

constructing an airborne PHM operator library, comprising: combing to obtain a PHM operator library according to each stage from the step 1 to the step 3;

wherein, the airborne PHM operator library is classified according to a data processing flow and comprises: a flow module and a support module;

the airborne PHM operator library is classified according to the function of operators and comprises the following components: the system comprises a data processing general operator, a PHM task special operator and an integrated modular airplane component level/system level/full-airplane level PHM operator.

Alternatively, in the onboard deployment-oriented health management predictive modeling approach described above,

the data processing generic operator comprises: the system comprises a data import operator unit, a data basic operation operator unit, a data preprocessing operator unit, a data exploration analysis operator unit, a characteristic engineering operator unit, a machine learning operator unit and a hyper-parameter optimization operator unit;

the PHM task dedicated operator comprises: the PHM task special operators comprise an expert system operator unit, an abnormality monitoring operator unit, a fault diagnosis operator unit, a service life prediction operator unit and the like; the expert system operator unit comprises expert knowledge, an experience-based model, a physical failure mechanism-based model and the like.

The integrated modular PHM operator at the aircraft component level/system level/full aircraft level is a high-integration state monitoring, fault diagnosis and service life prediction operator for a specific component, a specific system or a full aircraft of the aircraft; the airplane component level/system level/full-airplane level PHM operator is a modular operator of a component or system, which is constructed by selecting a specific data processing general operator and a PHM special operator from corresponding functional operators according to a full-flow airborne PHM algorithm processing strategy.

Alternatively, in the onboard deployment-oriented health management predictive modeling approach described above,

the flow module comprises: the system comprises a data import operator unit, a data exploration analysis operator unit, a data preprocessing operator unit, a characteristic engineering operator unit, a model training operator unit, a model evaluation operator unit, a model decision operator unit and a model deployment operator unit;

the supportive module comprises: the system comprises a data basic operation operator unit, a machine learning operator unit, an expert system operator unit and a hyper-parameter optimization operator unit;

the data basic operation operator unit is used as a basic operator for supporting the whole process module; the machine learning operator unit, the expert system operator unit and the hyper-parameter optimization operator unit are used as supporting operators for supporting the model training operator unit.

The invention has the beneficial effects that: on one hand, the health management predictive modeling method facing the airborne deployment, provided by the embodiment of the invention, realizes analysis, problem modeling and model decision by using massive heterogeneous data by means of a high-performance computer of a ground platform and a principle model machine which is the same as the airborne environment, and finally selects an optimal model facing a specific task to perform model deployment, thereby realizing a full-period PHM modeling strategy facing the airborne deployment; on the other hand, the airborne deployment-oriented full-period PHM modeling strategy is embodied into a specific operator, a PHM operator library is provided, and the airborne-oriented PHM modeling process is completely supported.

Drawings

The accompanying drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the example serve to explain the principles of the invention and not to limit the invention.

Fig. 1 is a schematic diagram illustrating a health management predictive modeling method for airborne deployment according to an embodiment of the present invention;

FIG. 2 is a schematic diagram illustrating a model evaluation phase and a model decision phase according to an embodiment of the present invention;

fig. 3 is a schematic diagram of an onboard PHM operator library in an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the embodiments and features of the embodiments in the present application may be arbitrarily combined with each other without conflict.

The following specific embodiments of the present invention may be combined, and the same or similar concepts or processes may not be described in detail in some embodiments.

The embodiment of the invention aims to establish a PHM modeling method facing airborne deployment and build a PHM operator library, which can completely support the analysis and modeling process of the PHM data on the aviation ground/airplane, quickly and conveniently realize the fault diagnosis and the service life analysis of members at the airplane component level/system level/whole airplane level, and evaluate the health state of the airplane.

Fig. 1 is a schematic diagram of a health management predictive modeling method for airborne deployment according to an embodiment of the present invention. The health management predictive modeling method facing airborne deployment provided by the embodiment of the invention is realized by adopting the following technical scheme; the method specifically comprises two parts of contents, namely, an airborne PHM algorithm processing strategy and a PHM operator library are formed.

A first part: forming an airborne PHM algorithm processing strategy:

embodiments of this section include:

step 1, performing data processing and model training on airborne data and ground data on a ground platform to obtain a plurality of training models;

and 2, evaluating, selecting and deciding the training model obtained in the step 1 in a principle prototype to obtain the training model meeting the task requirement.

And 3, deploying the training model meeting the task requirement obtained in the step 2 in an airborne vehicle.

The first part specifically comprises the following steps: a data importing stage, a data processing stage and a model training stage which operate on a ground platform; a model evaluation stage and a model decision stage which operate in a principle prototype, and a model deployment stage which operates in an airborne state; the above-described stages of operation are illustrated in fig. 1.

For each stage in the step 1, namely the data importing stage, the data processing stage and the model training stage, all operate on a ground high-performance computing platform (which may be referred to as a ground platform for short), in the step 1, deep mining and analysis of data are realized by utilizing a rich data analysis toolkit of the ground platform and by means of a high-performance computer of the ground platform aiming at massive heterogeneous data such as airborne operation historical data, test data generated by an experimental bench, simulation data generated by a simulation model platform, maintenance and repair data and the like; the model training aiming at specific PHM tasks (such as fault diagnosis, life prediction and the like) provides a trained model for the specific PHM tasks on the machine, and provides a basis for realizing the PHM tasks on the machine. Specific embodiments of each step in step 1 will be described below.

Step 1.1: and (5) a data importing stage.

The data import stage runs on a ground high-performance platform and can include the import of various structured data (such as data files in txt, csv and the like) and unstructured data (such as files in tables, images and the like).

Step 1.2: and (5) a data processing stage.

The data processing stage operates on a ground high-performance platform and comprises a data exploration and analysis sub-stage, a data preprocessing sub-stage and a characteristic engineering sub-stage.

The data exploration and analysis sub-stage can be realized by a plurality of mature algorithms, for example, the correlation analysis of the data can be realized by calculating the Pearson correlation coefficient, the Kendel correlation coefficient and the like between every two characteristics of the data, and the linear correlation between every two characteristics of the data can be visually displayed by drawing a scattered point pair diagram, a correlation hot area diagram and the like; the trend analysis can be visualized by plotting a line graph for each feature. The initial detection of the distribution condition of the data can be realized by calculating various statistical characteristics (mean, variance, mode, kurtosis, skewness, root mean square and the like) of the data, and the data distribution condition can also be visually displayed by drawing a data box line graph, a frequency histogram and the like.

The data preprocessing sub-stage can realize the improvement of data quality, and comprises the following steps: data cleaning, data denoising, data standardization and the like. The data cleaning is mainly carried out from the following three aspects: consistency check, invalid value and missing value processing and repeated value processing. For invalid values and missing values, three processing methods are adopted according to different situations: when the missing values are few and the importance degree of the attributes is low, if the attributes are numerical data, simply filling the data by adopting an average value, a median and the like according to the distribution condition of the data; if the missing rate is high and the importance degree of the attribute is low, the attribute can be directly deleted; if the deficiency rate is high and the attribute importance degree is high, an interpolation method and a modeling method are adopted. For the judgment of the repeated value, the records in the data set are firstly sorted according to a certain rule, and then whether the records are similar or not is judged by comparing adjacent records, so that whether the records are repeated or not is detected. Common methods for data normalization are min-max normalization (normalizing data for different features to within the [0,1] interval) and z-score normalization (normalizing data for different features to be uniform to data that obey a standard normal distribution).

The feature engineering sub-phase may employ a number of algorithms, specifically classified into an empirical and physical model-based method and a data-driven-based method. The method based on experience and physical models aims at specific objects, utilizes the existing experience knowledge or establishes corresponding physical models for the specific objects, and therefore the most useful characteristics for target tasks are selected or constructed; the data-driven method only starts from the angle of data, and selects or constructs the most significant features of the target by analyzing the trend of different feature data, the relevance between the feature data and the target label and the like, and the specific feature selection method comprises the following steps: the feature extraction method comprises Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Independent Component Analysis (ICA), Local Linear Embedding (LLE), equidistant mapping (Isomap) and the like.

Step 1.3: and (5) a model training stage.

The model training phase operates on a ground high-performance platform and comprises the following steps: through model training based on physical or empirical models and model training based on data driving, a plurality of algorithm models for specific PHM tasks are obtained through training.

The PHM tasks in this step include, for example: condition monitoring, fault diagnosis (identifying whether a fault exists, identifying the type of fault, identifying the location of the fault), fault prediction (predicting when a fault will occur), remaining life prediction (predicting component degradation trends, predicting when a component will fail), and the like.

For a specific PHM task type, the way to analyze and deconstruct the problem may be: firstly, according to cognition on an analysis object, if the object has a physical or empirical model which is easy to solve, a corresponding physical model is established for the object aiming at a specific task so as to carry out model training and solution; if the object structure is complex and failure or degradation mechanisms are difficult to obtain, a data-driven method is adopted, and algorithms are selected and model training is carried out by utilizing methods such as corresponding machine learning and statistical analysis.

It should be noted that, in the model training stage of step 1.3, the embodiment of the present invention may adopt multiple algorithms to train multiple models, so as to obtain multiple trained models.

For the above stages in the above step 2, namely the model evaluation stage and the model decision stage, the principle prototype similar to the airborne software and hardware environment is operated. After the model training phase of step 1 is completed, several trained models are obtained, and the several models need to be decided to obtain the optimal model finally deployed on the aircraft. Because the concept of the 'optimal model' is closely related to practical application and the purpose of airborne deployment is taken, the model is evaluated to designate the hardware environment for the model to operate, and a principle prototype which is the same as the airborne software/hardware environment is constructed to simulate the airborne operating environment. Specific embodiments of each step in step 2 will be described below.

Step 2.1: and (5) a model evaluation stage.

The model evaluation is operated on a principle prototype, and for different task types, various model evaluation indexes reflecting the model prediction accuracy can be adopted for model evaluation. The specific implementation mode is as follows: for the regression problem model, common model evaluation indexes include root mean square error RMSE, mean absolute error MAE, goodness of fit R2 and the like; for the classification problem model, the commonly used model evaluation indexes include accuracy acc, precision, recall, F1 score, ROC curve and the like. Besides the index reflecting the model prediction accuracy, the index also reflects the model operation efficiency: the model predicts the time, an index that is essential for a strong real-time task. A plurality of trained models are evaluated in a principle prototype, and a model evaluation table is obtained, as shown in fig. 2, which is a schematic diagram of a model evaluation stage and a model decision stage in the embodiment of the present invention.

Step 2.1: and (5) a model decision stage.

The model decision phase runs on a principle prototype. Firstly, a model decision rule is made according to specific task requirements, the model is finally decided by combining an obtained model evaluation table, and an 'optimal model' for the specific task, namely a training model meeting the task requirements, is selected.

The process of the model evaluation phase and the model decision phase is shown in fig. 2.

Step 3 of the embodiment of the present invention is specifically a model deployment phase.

In the model deployment stage, the "optimal model" selected in the step 2 is encapsulated by adopting a language supported by airborne hardware, and the encapsulated "optimal model" is deployed at an airborne terminal.

And (4) obtaining the airborne full-period PHM modeling strategy in the first part through the steps 1-3.

A second part: constructing an airborne PHM operator library;

according to each stage and each sub-stage of the onboard PHM algorithm processing strategy in the first part, corresponding contents are combed to obtain a PHM operator library, and as shown in FIG. 3, the PHM operator library is a schematic diagram of the onboard PHM operator library in the embodiment of the invention.

In the embodiment of the present invention, the airborne PHM operator library is classified according to a data processing procedure, and may include: a flow module and a support module.

In the embodiment of the present invention, the classification of the airborne PHM operator library according to the operator function may include: the system comprises a data processing general operator, a PHM task special operator and an integrated modular airplane component level/system level/full-airplane level PHM operator.

As shown in fig. 3, the onboard PHM operator library contains: the system comprises four support modules, namely a data basic operation operator unit, a machine learning operator unit, an expert system operator unit and a hyper-parameter optimization operator unit.

And the data basic operation operator unit is used as a basic operator for supporting the whole process. Including logical operations, operations by element, selecting/replacing rows/columns, etc.

The machine learning operator unit, the expert system operator unit and the hyper-parameter optimization operator unit are used as supporting operators for supporting the model training operator unit. The machine learning operator unit provides various classification, regression and artificial neural network operators for subsequent fault diagnosis and life prediction, which can be flexibly selected. The expert system operator units contain empirical solutions for specific components. The hyper-parameter optimization is a parameter tuning method provided for hyper-parameters in model training, and common model hyper-parameters comprise the number of network layers and the number of nodes in a neural network; maximum tree depth in the decision tree model, regularization coefficients and the like; the super-parameter optimization method comprises particle swarm optimization, a genetic algorithm, a simulated annealing algorithm and the like.

According to the strategy of the airborne PHM algorithm processing method, a specific data processing general operator and a PHM special operator are selected from an operator library according to the sequence of flow modules in the operator library, and the whole-flow state monitoring, fault diagnosis and service life prediction process for a specific component or a specific system is built to be used as an integrated modular operator of the component or the system.

The technical scheme provided by the embodiment of the invention aims at aviation predictive maintenance, solves the problem that the complete fault prediction and health management (PHM for short) predictive maintenance specifications of aviation airborne and ground are not uniform, and comprises a ground platform, a principle prototype and a full-period PHM algorithm strategy of airborne deployment. And forming a PHM operator library supporting data analysis and modeling, wherein the PHM operator library comprises a data import stage, a data processing stage, a model training stage, a model evaluation stage, a model decision stage and a model deployment stage and is realized as a specific PHM operator library. The PHM operator library completely supports the PHM modeling and data analysis processes on the computer, and meanwhile, the flexibility and the integration modularization are considered. Through theoretical analysis and tests, the operator library framework developed under the patent of the invention can meet the requirements of aviation PHM, and the PHM processing of airplane component level/system level/full-airplane level member oriented airborne deployment can be realized based on the operator library framework.

On one hand, the health management predictive modeling method facing the airborne deployment, provided by the embodiment of the invention, realizes analysis, problem modeling and model decision by using massive heterogeneous data by means of a high-performance computer of a ground platform and a principle model machine which is the same as the airborne environment, and finally selects an optimal model facing a specific task to perform model deployment, thereby realizing a full-period PHM modeling strategy facing the airborne deployment; on the other hand, the airborne deployment-oriented full-period PHM modeling strategy is embodied into a specific operator, a PHM operator library is provided, and the airborne-oriented PHM modeling process is completely supported.

The following schematically illustrates an implementation of the onboard deployment-oriented health management predictive modeling method provided by an embodiment of the present invention by a specific embodiment.

The health state of the airplane components can be quickly and conveniently evaluated by using integrated modularized airplane component-level operators aiming at specific parts of the airplane, such as an engine, a lubricating oil module, rotating parts and the like.

Aiming at certain infrequent data and faults on the airplane, according to the onboard PHM algorithm processing strategy provided by the embodiment of the invention, abundant data processing general operators and PHM task special operators in an operator library are supported to be used by the airplane, the data is explored and analyzed according to the sequence of flow modules in the operator library, and the whole diagnosis or prediction flow aiming at the data is flexibly established. The method comprises the following specific steps:

firstly, data import, data processing and model training are carried out in a ground high-performance platform. Take the windows platform using python language as an example. And the operators in the PHM operator library are called for data analysis and model training in the ground platform. The numerical data collected by the test bed are exported to be in the file forms of txt, excel, csv and the like, the collected file data are imported to the python environment, and the data are converted into data in the dataframe format by using a data import operator.

And performing data cleaning on the original data. And (3) checking whether missing data and repeated data exist or not and whether the problem of data inconsistency exists or not by using a data cleaning operator of the data preprocessing part in the PHM operator library, and processing the data problems according to a cleaning algorithm to improve the data quality. And (3) performing preliminary exploration on the data after quality improvement by using an operator for data exploration and analysis in the PHM operator library, acquiring data characteristics such as data scale, distribution characteristics and correlation, and providing ideas and guidance for the following steps. The operators of the data exploration analysis part comprise: correlation thermodynamic diagrams, line graphs, box graphs, data statistics, and the like.

After the step of data exploration and analysis, the analysis modeling thought after preliminary clarification is carried out. First, with task oriented, it is determined what model to build. According to the actual situation of the task, the method can be divided into a regression problem, a classification problem, a clustering problem, a prediction problem and the like, and each problem needs to be established with a different model. Regression and classification belong to the category of supervised learning, and refer to that the features to be solved (sample labels) in the training set data are known, wherein the features to be solved of the regression problem are continuous values, and the features to be solved of the classification problem are discrete values. Clustering belongs to the category of unsupervised learning, namely, the features (sample labels) to be solved in training set data are unknown, and a clustering model is commonly used for the problems of anomaly detection and the like. The prediction problem refers to predicting the value of the characteristic at a future moment by giving the characteristic value trend of a period of time, or predicting the failure moment of the equipment according to the value of some characteristics in a period of time.

And (4) preprocessing and performing characteristic engineering operation on the data by combining the conclusion obtained by data exploration and analysis. If the model working condition is unknown in the actual situation, the data merging operator in the operator library is used in the preprocessing step, all input working condition data files are merged, and the working condition is not used as the model input. If the selected algorithm model is sensitive to the range difference of different input characteristics, the data characteristics are normalized into data which obeys standard normal distribution by using a data normalization operator in a preprocessing operator library. In the feature engineering step, the input features of the model need to be selected or constructed from the raw data. If the variance of a certain feature is found to be small and has almost no change trend in the exploration analysis step, the feature is deleted.

In the model training stage, a proper algorithm is selected for training the model, and the operator library of the part comprises: and algorithms such as linear regression, decision trees, support vector machines, gradient lifting trees, random forests, artificial neural networks and the like support the construction of various regression, classification and prediction models. Selecting a plurality of proper algorithms to construct a model according to the thought provided in the data exploration analysis; dividing all samples into a training set and a testing set according to a proper proportion, and sending the training set samples into the constructed models for model training to respectively obtain a plurality of specific trained models.

At this point, the data analysis and model training phase using the high performance server at the ground platform ends. Because of the limited on-board computational resources, consider only the on-board predictive part of the model: and carrying out lightweight packaging on the model trained on the ground by using C language, and deploying the model in an embedded platform on the machine to carry out real-time diagnosis and prediction on the machine. In order to simulate an airborne embedded environment, a proper embedded development board is selected or a proper development board is constructed by the self to be used as a principle prototype, firstly, a plurality of trained models are encapsulated in a light weight mode by using C language, the obtained trained models are respectively burnt into the principle prototype, test set data are input, model prediction is carried out in the board, and the step is used as the environment simulation of on-board prediction to realize the verification link of the algorithm.

And evaluating results obtained by predicting several models in the board by using various evaluation indexes. Common regression model accuracy assessment indicators are: RMSE, MAE, R2; common classification model accuracy assessment indicators are: precision, recall, accuracy, F1 value, etc. Besides the model accuracy evaluation index, indexes such as model training time, model testing time and the like are also needed for evaluating the prediction real-time performance of the model. And (3) respectively inputting the test set data into a training model in the principle prototype, predicting in a plate, and calculating the numerical values of several model evaluation indexes to obtain a model evaluation table shown in fig. 2. The evaluation table is used as a reference for selecting the optimal model and deploying on board, and the optimal model which best meets the actual requirements is selected according to the requirements of the actual situation, so that the model can be deployed on the board.

Although the embodiments of the present invention have been described above, the above description is only for the convenience of understanding the present invention, and is not intended to limit the present invention. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (9)

1. An airborne deployment-oriented health management predictive modeling method, comprising:

step 1, performing data processing and model training on airborne data and ground data on a ground platform to obtain a plurality of training models;

step 2, evaluating, selecting and deciding the training model obtained in the step 1 in a principle prototype to obtain a training model meeting the task requirement;

step 3, deploying the training model meeting the task requirement obtained in the step 2 in an airborne vehicle;

and (4) obtaining an airborne full-period PHM modeling strategy through the steps 1-3.

2. The on-board deployment oriented predictive modeling of health management of claim 1, wherein said step 1 comprises: a data importing stage, a data processing stage and a model training stage which operate on a ground platform; the method specifically comprises the following steps:

step 1.1, a data import phase, comprising: importing various structured data and unstructured data in a ground high-performance platform;

step 1.2, a data processing stage, comprising: a data exploration and analysis sub-stage, a data preprocessing sub-stage and a characteristic engineering sub-stage;

step 1.3, a model training phase, comprising: through model training based on physical or empirical models and model training based on data driving, a plurality of algorithm models for specific PHM tasks are obtained through training.

3. The on-board deployment oriented health management predictive modeling method of claim 2, characterized in that in said step 1.2,

a data exploration analysis sub-phase comprising: visually exploring and simply mining the imported data to obtain the number of samples, the number of characteristics, data distribution characteristics, characteristic trend and correlation;

the data preprocessing sub-stage is used for improving the data quality and comprises the following steps: data cleaning, data denoising and data standardization;

a feature engineering sub-phase comprising: feature mining, feature selection and feature extraction are carried out on the data to extract features which are most useful for specific tasks (such as fault diagnosis and life prediction).

4. The on-board deployment oriented predictive modeling of health management of claim 3, wherein said PHM task of step 1.3 comprises: state monitoring, fault diagnosis, fault prediction and residual life prediction;

for a specific PHM task type, the method for analyzing and deconstructing the problem is as follows:

in the method 1, according to cognition on an analysis object, when the object has a physical or empirical model which is easy to solve, a corresponding physical model is established for the object aiming at a specific task so as to carry out model training and solution;

in the mode 2, when the object structure is complex and failure or degradation mechanisms are difficult to obtain, a data-driven method is adopted, and algorithms are selected and model training is carried out by using methods such as corresponding machine learning and statistical analysis;

in the model training stage of step 1.3, multiple algorithms are adopted to train multiple models, so as to obtain multiple trained algorithm models.

5. The on-board deployment oriented health management predictive modeling method of claim 4, wherein said step 2 comprises: a model evaluation stage and a model decision stage of a principle prototype which is similar to an airborne software and hardware environment; the method specifically comprises the following steps:

step 2.1, in a model evaluation stage, evaluating a plurality of trained algorithm models in a principle prototype, selecting a plurality of appropriate model evaluation indexes according to the types of the algorithm models and the task requirements, and operating a plurality of algorithm models trained on the ground in the principle prototype to obtain a model evaluation table;

and 2.2, in the model decision stage, making a model decision rule according to the specified task requirement, and carrying out final decision on the model by combining the obtained model evaluation table to select the optimal model for the specific task.

6. The on-board deployment oriented health management predictive modeling method of claim 5, wherein said step 3 comprises:

and in the model deployment stage, packaging the selected optimal model by adopting a language supported by airborne hardware, and deploying the packaged optimal model to an airborne terminal.

7. The airborne deployment-oriented predictive health management modeling method of any of claims 1-6, further comprising:

constructing an airborne PHM operator library, comprising: combing to obtain a PHM operator library according to each stage from the step 1 to the step 3;

wherein, the airborne PHM operator library is classified according to a data processing flow and comprises: a flow module and a support module;

the airborne PHM operator library is classified according to the function of operators and comprises the following components: the system comprises a data processing general operator, a PHM task special operator and an integrated modular airplane component level/system level/full-airplane level PHM operator.

8. The on-board deployment oriented health management predictive modeling method of claim 7,

the data processing generic operator comprises: the system comprises a data import operator unit, a data basic operation operator unit, a data preprocessing operator unit, a data exploration analysis operator unit, a characteristic engineering operator unit, a machine learning operator unit and a hyper-parameter optimization operator unit;

the PHM task dedicated operator comprises: the PHM task special operators comprise an expert system operator unit, an abnormality monitoring operator unit, a fault diagnosis operator unit, a service life prediction operator unit and the like; the expert system operator unit comprises expert knowledge, an experience-based model, a physical failure mechanism-based model and the like.

The integrated modular PHM operator at the aircraft component level/system level/full aircraft level is a high-integration state monitoring, fault diagnosis and service life prediction operator for a specific component, a specific system or a full aircraft of the aircraft; the airplane component level/system level/full-airplane level PHM operator is a modular operator of a component or system, which is constructed by selecting a specific data processing general operator and a PHM special operator from corresponding functional operators according to a full-flow airborne PHM algorithm processing strategy.

9. The on-board deployment oriented health management predictive modeling method of claim 7,

the flow module comprises: the system comprises a data import operator unit, a data exploration analysis operator unit, a data preprocessing operator unit, a characteristic engineering operator unit, a model training operator unit, a model evaluation operator unit, a model decision operator unit and a model deployment operator unit;

the supportive module comprises: the system comprises a data basic operation operator unit, a machine learning operator unit, an expert system operator unit and a hyper-parameter optimization operator unit;

the data basic operation operator unit is used as a basic operator for supporting the whole process module; the machine learning operator unit, the expert system operator unit and the hyper-parameter optimization operator unit are used as supporting operators for supporting the model training operator unit.

Images