CN106156401B - Multi-combination classifier based data driving system state model online identification method - Google Patents

Multi-combination classifier based data driving system state model online identification method Download PDF

Info

Publication number
CN106156401B
CN106156401B CN201610397845.8A CN201610397845A CN106156401B CN 106156401 B CN106156401 B CN 106156401B CN 201610397845 A CN201610397845 A CN 201610397845A CN 106156401 B CN106156401 B CN 106156401B
Authority
CN
China
Prior art keywords
classifier
model
classification
training
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201610397845.8A
Other languages
Chinese (zh)
Other versions
CN106156401A (en
Inventor
吕梅柏
朱丹
杨天社
李浩宇
郭小红
韩治国
姜海旭
李肖瑛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northwestern Polytechnical University
China Xian Satellite Control Center
Original Assignee
Northwestern Polytechnical University
China Xian Satellite Control Center
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Northwestern Polytechnical University, China Xian Satellite Control Center filed Critical Northwestern Polytechnical University
Priority to CN201610397845.8A priority Critical patent/CN106156401B/en
Publication of CN106156401A publication Critical patent/CN106156401A/en
Application granted granted Critical
Publication of CN106156401B publication Critical patent/CN106156401B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/30—Circuit design
    • G06F30/36—Circuit design at the analogue level
    • G06F30/367—Design verification, e.g. using simulation, simulation program with integrated circuit emphasis [SPICE], direct methods or relaxation methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/28—Databases characterised by their database models, e.g. relational or object models
    • G06F16/284—Relational databases
    • G06F16/285—Clustering or classification

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Hardware Design (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Evolutionary Computation (AREA)
  • Geometry (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a data driving system state model online identification method based on a multi-combination classifier, which is used for solving the technical problem that the existing data driving system state model online identification method is difficult to model. The technical scheme includes that analysis of historical monitoring data and system operation characteristics of the spacecraft is combined, real-time measurement data are analyzed and classified through design and training of a plurality of combined classifiers, and an operation state model of each system of the spacecraft is obtained through identification. The technical problem that a mathematical model cannot be built by directly utilizing a physical model due to high system complexity is solved. In the identification process, the classifier is designed and trained in advance, so that the identification time of the real-time model is short, the online identification of the system state model can be realized, the model reflects whether a fault occurs in the running spacecraft and the type and degree of the fault, and the method has important significance for the research of the real-time state monitoring, the fault-tolerant control system design, the fault repair and the like of the system.

Description

Multi-combination classifier based data driving system state model online identification method
Technical Field
The invention relates to an online identification method of a data driving system state model, in particular to an online identification method of a data driving system state model based on a multi-combination classifier.
Background
The document "equivalent replacement-based fault modeling method research, journal of northwest industrial university, volume 33, No. 1 of 2015, 2/month" discloses an equivalent model replacement-based fault modeling method, which can realize effective simulation of faults and form a general fault equivalent model replacement implementation step on the premise of utilizing expert experience knowledge and avoiding building a complex model as far as possible, namely, on the basis of analyzing the system working mechanism, building a normal model, dividing functional modules, then analyzing a fault mode by combining a function layering thought and a fault tree analysis method, and classifying the faults in a grading manner. The method described in the literature requires system module division according to a physical model and a working mechanism of a modeling object, then forms a fault equivalent model library through fault equivalent modeling, and compares an object running state with information in the model library, thereby establishing a fault model. In practical application, however, a spacecraft system is complex, a physical model and a working mechanism of the system cannot be accurately analyzed, and a fault equivalent model library is difficult to establish; the on-orbit operation environment of the spacecraft is also complex and changeable, and the equivalent subsystem model in the model base may have a larger difference from the actual system, which may cause inaccurate model establishment.
Disclosure of Invention
In order to overcome the defect of difficult modeling of the existing data-driven system state model online identification method, the invention provides a data-driven system state model online identification method based on a multi-combination classifier. The method is combined with analysis of historical monitoring data and system operation characteristics of the spacecraft, real-time measurement data are analyzed and classified through design and training of a plurality of combined classifiers, and an operation state model of each system of the spacecraft is obtained through identification. The problem of because the system complexity is higher, can't directly utilize the physical model to build the technique of mathematical model, carry on the inaccurate problem of model that does not consider the model operational environment and cause in modelling according to the system physics operation mechanism is solved, have improved feasibility and accuracy of setting up the spacecraft on orbit running state model. In the identification process, the classifier is designed and trained in advance, so that the identification time of the real-time model is short, the online identification of the system state model can be realized, the model reflects whether a fault occurs in the running spacecraft and the type and degree of the fault, and the method has important significance for the research of the real-time state monitoring, the fault-tolerant control system design, the fault repair and the like of the system.
The technical scheme adopted by the invention for solving the technical problems is as follows: a data driving system state model online identification method based on a multi-combination classifier is characterized by comprising the following steps:
step one, historical data analysis and processing.
And analyzing and processing the historical monitoring data, extracting a characteristic vector from the historical monitoring data, setting a characteristic label of each state, and acquiring the information of the system running state.
And step two, designing a multi-combination classifier.
The method adopts a multi-combination classifier joint decision method, combines a naive Bayes classifier and a support vector machine classifier, inspects the classification capability of each single classifier and the state characteristics of a system to be classified, determines the weight of the classifier, and determines the final classification result by a weighted summation method according to the primary classification results of the two classifiers. And when the weight of the classifier is determined, optimizing by adopting a particle swarm optimization algorithm, and designing a multi-combination classifier according to a classification method. The classification method comprises a fixed weight joint classification method based on the performance of the classifier and a classification-based weight joint classification method based on the characteristic of the characteristic vector of each category of the system to be classified.
And step three, training a multi-combination classifier.
Training and learning are respectively carried out on a naive Bayes classifier and a support vector machine classifier, then training and learning are carried out on an output result of each classifier, and through training of a particle swarm optimization algorithm, a classifier combination weight is obtained, and training of a multi-combination classifier is completed. And according to different angles, adopting different criteria to distribute the weights.
The first distribution method comprises the following steps: and (5) classifying by a fixed weight joint classifier.
After classification decision of the target to be classified is carried out through the NBS classifier and the SVM classifier respectively, weighted summation is carried out on classification results output by the classifiers, and the class corresponding to the maximum value of the weighted summation is used as the class of the target to be classified.
Selecting m groups of sample data S { (X)1,Y1),...,(Xj,Yj),...,(Xc,Yc) Using the obtained data as a training set of a classifier, wherein c is the number of system state classes, XjN-dimensional feature vector set X prepared for dataj＝{xj1,xj2,...,xjn}，YjFor training the labeled classes of the samples, the labeled classes corresponding to the system states of the same class are the same. The NBS classifier and SVM classifier are trained separately, denoted CBayesAnd CSVM. In order to select the optimal weight, the feature vectors in the training set S are collected to be { X }1,X2...,XcAnd the data are respectively input into an NBS classifier and an SVM classifier as test data.
In the NBS classifier, the output result of the measurement layer is a posterior probability set corresponding to each type of the feature vector set to be tested:
Figure BDA0001011634540000021
in the SVM classifier, 1 to 1 multi-class support vector machine classifier is adopted, the classification result of each binary classifier is taken as a voter, and the voting method is adopted to obtain the number of votes corresponding to each type of the feature vector set to be tested:
normalizing the set according to rows to obtain a probability set output by a 1-to-1 multi-class support vector machine classifier:
Figure BDA0001011634540000032
as shown in the formulas (1) and (2), the same test data { X is inputted1,X2...,XcIn the time of the classification, the dimensionalities of the probability sets output by the two classifiers are equal and correspond to one another. Continuously training the classification effect of each classifier by adopting a training set, and assigning a weight omegaNBSAnd ωSVMAnd weighting and adding the output probability sets of the two classifiers to establish a combined classifier output probability set:
Figure BDA0001011634540000033
the label class corresponding to the maximum probability value of each row of the probability set is the label class corresponding to each group of test feature vectors, namely the classification result of the joint classifier, namely the label class set output by the joint classifier:
Figure BDA0001011634540000034
defining the serial number of each training sample data group as i, i is 1, 2.. m, the class serial number of the sample data group is j, j is 1, 2.. c, and the class label corresponding to the sample data group is YiThe class of the sample data classified by the joint classifier is Li：
Li＝arg max{(ωNBS·Pij+ωBVM·Pij)|j＝1,2,...c} (6)
In order to obtain a better weight coefficient, the output results of the NBS classifier and the SVM classifier in the measurement layer are learned, so that joint scores are obtainedThe classification accuracy of the classifier is highest, namely the optimal weight solving process is converted into a combined classifier output label class set L (L)1,L2,...,Lm) And a test sample data mark class set Y (Y)1,Y2,...,Ym) The number of elements corresponding to each element in the method is maximized:
searching for weight coefficient omega (omega) by adopting particle swarm optimization algorithmNBS,ωSVM) The solving process is as follows:
(a) particle initialization: randomly generating D particles in W-dimensional space, and their positions
Figure BDA0001011634540000042
And velocity
Figure BDA0001011634540000043
Figure BDA0001011634540000044
Memo
Figure BDA0001011634540000045
For the best solution, p, currently searched for by the particle dg＝(pg1,pg2,...,pgc) And searching the optimal solution currently searched for the whole particle swarm.
(b) And (3) calculating a fitness value: the performance of each particle is evaluated using equation (8) as a fitness function, and for each particle, its fitness value is compared with its historically optimal fitness value.
(c) Updating the particles according to the size of the fitness value: in [0,1 ]]In the random generation of coefficient r1And r2According to the formula:
Vd(t+1)＝Vd(t)+C1×r1×(pd(t)-Xd(t))+C2×r2×(pgd(t)) (9)
Xd(t+1)＝Xd(t)+Vd(t+1) (10)
updating the speed and the position of the particles to obtain the newly updated optimal position p of each particle ddAnd the optimal position p of the particle swarmgAnd the best value is taken as the historical optimum value.
(d) And (5) calculating the fitness value again, if the condition of finishing is met, namely the classification accuracy reaches the requirement, finishing, otherwise, turning to the step (b), and updating the speed and the position of the particles again.
And a second distribution method: and combining classifier classification according to classification weight.
Due to the different characteristics of each fault, the classification effect in a single classifier is different, namely different weights can be assigned to different fault types when weights are assigned to the NBS classifier and the SVM classifier.
For training sample data S { (X)1,Y1),...,(Xj,Yj),...,(Xc,Yc) C fault types in the training set S, and feature vectors in the training set S are collected into a set (X)1,X2...,XcAnd the data are respectively input into an NBS classifier and an SVM classifier as test data. And (3) outputting a posterior probability set corresponding to each fault type in the feature vector set to be classified through the classification of the NBS classifier:
PNBS＝[P1P2… Pc](11)
wherein P isiAnd the set of the posterior probabilities corresponding to each group of training samples in the ith type of fault.
In the SVM classifier, 1 to 1 multi-class support vector machine classifier is adopted, a voting method is adopted to obtain a classification result, and the output probability set of the normalized classifier is as follows:
PSVM＝[P1P2… Pc](12)
wherein P isiAnd obtaining a probability set corresponding to each group of training samples in the ith fault type.
Distributing weights according to the characteristics of each fault type, weighting and adding the output probability sets of the two classifiers, and establishing a combined classifier according to class weight:
Figure BDA0001011634540000051
the label class corresponding to the maximum probability value of each row of the probability set is the label class corresponding to each group of test feature vectors, namely the classification result of the joint classifier, namely the label class set output by the joint classifier:
Figure BDA0001011634540000052
seeking an optimal weight set through a particle swarm optimization algorithm, enabling a fitness function to be similar to a unified weighting joint classifier, and converting the optimal weight set solving process into a joint classifier output label class set L (L)1,L2,...,Lm) And a test sample data mark class set Y (Y)1,Y2,...,Ym) The number of elements corresponding to each element in the group is maximized, the position and the speed of the particle swarm are continuously updated according to the fitness function, and the optimal weight coefficient set omega (omega) with the fitness function meeting the conditions is searchedNBS1,ωNBS2,...,ωNBSc,ωSVM1,ωSVM2,...,ωSVMc)。
And step four, diagnosing and classifying the real-time measurement data.
In order to realize real-time detection and identification of the fault model of the in-orbit spacecraft system, historical data needs to be analyzed, and a corresponding model structure is determined when the system state is classified, namely, a mapping relation of the classification result and the model structure in one-to-one correspondence is established, so that the structure of the model can be determined after the test data is classified by a classifier. Therefore, when a large amount of simulation data is used for training the classifier, the structure of the model needs to be trained, a system state-structure mapping table is established, and the running state of the system corresponds to the model structure one by one.
When the model structure is known, the model structure of the system in different states can be obtained according to the physical performance analysis of the model, and a system state-structure mapping table is directly established. When the model structure is unknown, a large number of training data sample sets are adopted to identify the model structure of the known fault type, and the AIC information criterion is adopted to identify the structure of the training sample data in consideration of the complexity of the model.
And step five, identifying the state model on line.
After the model structure of the system is determined through the system state classification and structure identification in the first to fourth steps, a recursion method is adopted to carry out online estimation on the measured data acquired in real time, and a new estimated value of the model is obtained. In order to optimize the identification efficiency, the model parameters are identified by adopting a classical least square parameter identification method, so that the online identification of the model parameters is realized.
The invention has the beneficial effects that: the method is combined with analysis of historical monitoring data and system operation characteristics of the spacecraft, real-time measurement data are analyzed and classified through design and training of a plurality of combined classifiers, and an operation state model of each system of the spacecraft is obtained through identification. The problem of because the system complexity is higher, can't directly utilize the physical model to build the technique of mathematical model, carry on the inaccurate problem of model that does not consider the model operational environment and cause in modelling according to the system physics operation mechanism is solved, have improved feasibility and accuracy of setting up the spacecraft on orbit running state model. In the identification process, the classifier is designed and trained in advance, so that the identification time of the real-time model is short, the online identification of the system state model can be realized, the model reflects whether a fault occurs in the running spacecraft and the type and degree of the fault, and the method has important significance for the research of the real-time state monitoring, the fault-tolerant control system design, the fault repair and the like of the system.
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
Drawings
FIG. 1 is a flow chart of the method for on-line identification of a multi-combination classifier-based data-driven system state model according to the present invention.
FIG. 2 is a flow chart of an algorithm of a combined classifier composed of a naive Bayes classifier and a support vector machine classifier in accordance with an embodiment.
Fig. 3 is a curve of the variation of the fitness value in the process of obtaining the optimal weight by combining multiple classifiers through the particle swarm optimization algorithm in the specific embodiment.
FIG. 4 illustrates the recognition result of the model system according to an embodiment.
FIG. 5 illustrates an error in identifying a normal state model according to an embodiment.
FIG. 6 illustrates the system identification of a fault model in the event of a single axis stuck fault in an embodiment.
FIG. 7 illustrates fault model identification errors in the event of a single axis stuck fault in an embodiment.
FIG. 8 illustrates the system identification of a two-axis stuck-time fault model in accordance with an embodiment.
FIG. 9 illustrates the fault model identification error for a two-axis stuck condition in an embodiment.
FIG. 10 is a fault model system identification when friction torque is increased in an embodiment.
FIG. 11 illustrates fault model identification errors when friction torque increases according to an embodiment.
FIG. 12 illustrates the fault model system identification for a constant gain change in a sensor, according to an embodiment.
FIG. 13 illustrates fault model identification errors for constant gain changes in a sensor, according to an embodiment.
Detailed Description
Reference is made to fig. 1-13. The invention discloses a data driving system state model online identification method based on a multi-combination classifier, which comprises the following specific steps:
step 1: and (5) analyzing and processing historical data.
Analyzing and researching the in-orbit satellite attitude control system, establishing a corresponding mathematical model, acquiring a large amount of in-orbit satellite attitude control system simulation data in different states through fault state analysis and fault injection, and selecting simulation data in a certain initial state from the simulation data as measured data for testing. Several typical fault modes of the satellite attitude control system are taken as examples, and training data and real-time test data are obtained through fault injection simulation. Wherein the sampling period is set to T ═ 1s, each variable is sampled uniformly, the time sequence is consistent, and the length of the data collected under each condition is 1500 groups, and the selected typical failure mode is as shown in table 1 below:
TABLE 1 satellite attitude control System typical failures
Figure BDA0001011634540000071
Step 2: and (4) designing a multi-combination classifier.
And designing a multi-combination classifier for classifying the states of the satellite attitude control system. Designing a fixed weight combined classifier according to different classification performances of a naive Bayes classifier and a support vector machine classifier; according to the normal operation state of the satellite control system and the operation characteristics of the satellite control system in several typical failure modes, a classification-weight combined classifier is designed.
And step 3: and (5) training a multi-combination classifier.
The learning process of the multi-classifier combination under the double mechanism of naive Bayes and a support vector machine belongs to directed learning, wherein single classifiers firstly carry out respective training learning, and then combined classifiers carry out the learning of training combination and basic classifier output results, namely the learning of a combined algorithm. Aiming at a fixed weight joint classification method based on classification performance, the optimal weight only has two numerical values, and two classifiers are directly combined to form a new classifier; two groups of weight values need to be searched for the classification-weight combined classification method based on different state characteristics of the system, and different weights are selected for different system states including a normal state and different fault states to be combined, so that a new classifier is formed.
And optimizing the weight by adopting a particle swarm optimization algorithm, and continuously updating the positions of the particle swarm until the correct recognition rate of the combined classification method on the training set data set is the highest.
In the process of optimizing the weight combined by the multiple classifiers by using the particle swarm optimization algorithm, the adaptability values of the fixed weight combined classification method and the classified weight combined classification method can be rapidly converged, which shows that the method has higher efficiency and can realize the rapid optimization of the combined weight of the multiple classifiers.
In the training process, the recognition rates of the different classification methods are shown in the following table:
TABLE 2 training set data test result comparison
Based on the feature vectors, the method for classifying the single classifiers by adopting the training data comprises the steps of training an NBS classifier, an SVM multi-classifier of 1 to 1 and two NBS and SVM combined classifiers, and testing the classification effect of the classifiers by adopting the test data. The performance and results of the four classification methods are compared as shown in the following table:
TABLE 3 comparison of Performance and results for four classification methods
Figure BDA0001011634540000082
As shown in the table, the two single classifiers have good effect on the classification of various faults, and the classification accuracy of the measured data set reaches more than eighty percent. The NBS classifier is simple in structure, the training time of the classifier and the testing time required by classifying each group of test data are short, the 1-to-1 support vector machine classifier is composed of a plurality of support vector machine two classifiers, each state class in the training data needs to be trained pairwise, and the required time is relatively long. The correct recognition rate of the two multi-classifiers to the test data set is higher than that of a single classifier, and the application value is higher. Although the training time is greatly improved compared with that of a single classifier because the particle swarm optimization algorithm needs to carry out multiple iterations when seeking the optimal weight during the training of multiple classifiers, the classification time of each group of test data is still short, and the time limit in practical application can be met.
And 4, step 4: and diagnosing and classifying the real-time measurement data.
The relevant variables of the satellite attitude control system comprise satellite attitude angles in three directions, satellite attitude angular velocities and electromagnetic driving moments of a flywheel, wherein the satellite attitude angles are output variables, and the satellite attitude angular velocities and the electromagnetic driving moments of the flywheel are input variables. In the case where the physical model is unknown, the system can be represented by the following canonical difference equation:
a(z-1)Y(k)＝B(z-1)U(k)+ξ(k)
wherein Y (k) is 3-dimensional output, U (k) is 6-dimensional input, ξ (k) is 3-dimensional noise, a and B are parameters to be identified, and:
a(z-1)＝1+a1z-1+…+anz-n
B(z-1)＝B+B1z-1+…+Bnz-n
wherein the number of parameters to be identified is na+18·(nb+1) of which n isa、nbThe model order for the output variables and the input variables, respectively.
Through model structure identification based on the AIC criterion, system model orders and model structure parameters in each state can be obtained, so as to establish a system state-structure mapping table of a satellite operation normal state and a typical fault type, as shown in table 4:
table 4 satellite attitude control system state-structure mapping table
Figure BDA0001011634540000091
Through fault diagnosis and classification, a model structure corresponding to the measured data can be obtained according to the table above and used for online parameter identification.
And 5: and (4) online identification of the state model.
Fig. 4 shows the system identification result of the satellite attitude control system in a normal state. In the whole simulation process, the satellite attitude is gradually stabilized at a target attitude under the regulation and control of a flywheel, and the output of a mathematical model obtained through state classification and online identification is matched with the change of actual measurement data.
FIG. 5 shows the system identification error of the satellite attitude control system in a normal state. Although the identification error is different at different moments, the error is less than 0.005 percent in general, the identification precision is higher, and the identification result has higher credibility and stability.
FIGS. 6-13 illustrate system identification results and identification errors for several exemplary failure modes of a satellite attitude control system, respectively. When the flywheel breaks down, the satellite attitude cannot reach the controlled target and deviates from the normal operation state. And errors between the identification result and the actual operation state are less than 0.1%, and the operation state of the satellite can be accurately tracked through the fault model obtained by identification. Through the model, the change condition and the fault condition of the satellite attitude control system can be mastered, and the identification result has higher accuracy and credibility. In the simulation process, the total time used for fault classification and model identification of each group of measurement data is not more than 1 millisecond, the requirement of real-time online model identification can be met, and the method has high practical application value
The above examples illustrate that the method for on-line identification of the on-orbit state model of the spacecraft under the data-driven condition by adopting the multiple classifiers is accurate and effective, and can be applied to the field of fault diagnosis, prediction and control development of the spacecraft to a certain extent. Meanwhile, the technology is not limited to identification of the spacecraft state model, can also be applied to online identification of system state models in other fields, and has good popularization and application values.

Claims (1)

1. A data driving system state model online identification method based on a multi-combination classifier is characterized by comprising the following steps:
step one, analyzing and processing historical data;
analyzing and processing historical monitoring data, extracting a characteristic vector from the historical monitoring data, setting a characteristic label of each state, and acquiring information of the system running state;
step two, designing a multi-combination classifier;
combining a naive Bayes classifier and a support vector machine classifier by adopting a multi-combination classifier joint decision method, inspecting the classification capability of each single classifier and the state characteristics of a system to be classified, determining the weight of the classifiers, and determining the final classification result of the primary classification results of the two classifiers by a weighted summation method; when the weight of the classifier is determined, optimizing by adopting a particle swarm optimization algorithm, and designing a multi-combination classifier according to a classification method; the classification method comprises a fixed weight joint classification method based on the performance of a classifier and a classification-based weight joint classification method based on the characteristic of the characteristic vector of each category of the system to be classified;
step three, training a multi-combination classifier;
training and learning a naive Bayes classifier and a support vector machine classifier respectively, then training and learning an output result of each classifier, acquiring a classifier combination weight through training of a particle swarm optimization algorithm, and finishing the training of a multi-combination classifier; according to different angles, adopting different criteria to distribute weights;
the first distribution method comprises the following steps: classifying by a fixed weight joint classifier;
after classification decision of the target to be classified is carried out through an NBS classifier and an SVM classifier respectively, weighted summation is carried out on classification results output by the classifiers, and the class corresponding to the maximum value of the weighted summation is taken as the class of the target to be classified;
selecting m groups of sample data S { (X)1,Y1),...,(Xj,Yj),...,(Xc,Yc) Using the obtained data as a training set of a classifier, wherein c is the number of system state classes, XjN-dimensional feature vector set X prepared for dataj＝{xj1,xj2,...,xjn}，YjFor the label class of the training sample, the label classes corresponding to the system states of the same class are the same; the NBS classifier and SVM classifier are trained separately, denoted as
Figure FDA0002273754590000011
And
Figure FDA0002273754590000012
in order to select the optimal weight, the feature vectors in the training set S are collected to be { X }1,X2...,XcInputting the test data into an NBS classifier and an SVM classifier respectively;
in the NBS classifier, the output result of the measurement layer is a posterior probability set corresponding to each type of the feature vector set to be tested:
in the SVM classifier, 1 to 1 multi-class support vector machine classifier is adopted, the classification result of each binary classifier is taken as a voter, and the voting method is adopted to obtain the number of votes corresponding to each type of the feature vector set to be tested:
Figure FDA0002273754590000022
normalizing the set according to rows to obtain a probability set output by a 1-to-1 multi-class support vector machine classifier:
as shown in the formulas (1) and (2), the same test data { X is inputted1,X2...,XcIn the time of the classification, the dimensionalities of the probability sets output by the two classifiers are equal and correspond to one another; continuously training the classification effect of each classifier by adopting a training set, and assigning a weight omegaNBSAnd ωSVMAnd weighting and adding the output probability sets of the two classifiers to establish a combined classifier output probability set:
Figure FDA0002273754590000024
the label class corresponding to the maximum probability value of each row of the probability set is the label class corresponding to each group of test feature vectors, namely the classification result of the joint classifier, namely the label class set output by the joint classifier:
Figure FDA0002273754590000025
defining the serial number of each training sample data group as i, i is 1, 2.. m, the class serial number of the sample data group is j, j is 1, 2.. c, and the class label corresponding to the sample data group is YiThe class of the sample data classified by the joint classifier is Li：
Figure FDA0002273754590000031
In order to obtain a better weight coefficient, the output results of the NBS classifier and the SVM classifier in the measurement layer are learned, so that the classification accuracy of the combined classifier is highest, namely, the optimal weight solving process is converted into a combined classifier output label class set L (L)1,L2,...,Lm) And a test sample data mark class set Y (Y)1,Y2,...,Ym) The number of elements corresponding to each element in the method is maximized:
Figure FDA0002273754590000032
searching for weight coefficient omega (omega) by adopting particle swarm optimization algorithmNBS,ωSVM) The solving process is as follows:
(a) particle initialization: randomly generating D particles in W-dimensional space, and their positions
Figure FDA0002273754590000033
And velocity
Figure FDA0002273754590000035
And note pd＝(pd1,pd2) For the best solution, p, currently searched for by the particle dg＝(pg1,pg2) Searching the optimal solution for the whole particle swarm currently;
(b) and (3) calculating a fitness value: evaluating the performance of each particle by taking the formula (7) as a fitness function, and comparing the fitness value of each particle with the historical optimal fitness value of each particle;
(c) updating the particles according to the size of the fitness value: in [0,1 ]]In the random generation of coefficient r1And r2According to the formula:
Vd(t+1)＝Vd(t)+C1×r1×(pd(t)-Xd(t))+C2×r2×(pgd(t)) (9)
Xd(t+1)＝Xd(t)+Vd(t+1) (10)
updating the speed and the position of the particles to obtain the newly updated optimal position p of each particle ddAnd the optimal position p of the particle swarmgTaking the optimal value as a historical optimal value;
(d) calculating the fitness value again, if the condition of finishing is met, namely the classification accuracy reaches the requirement, finishing, otherwise, turning to the step (b), and updating the speed and the position of the particles again;
and a second distribution method: classifying by a classification weight combined classifier;
because the characteristics of each fault are different, the classification effect in a single classifier is different, namely different weights are assigned to different fault types when weights are assigned to an NBS classifier and an SVM classifier;
for training sample data S { (X)1,Y1),...,(Xj,Yj),...,(Xc,Yc) C fault types in the training set S, and feature vectors in the training set S are collected into a set (X)1,X2...,XcInputting the test data into an NBS classifier and an SVM classifier respectively; classified by an NBS classifier, and the feature to be classified is outputAnd measuring a posterior probability set corresponding to each fault type in the set:
wherein
Figure FDA0002273754590000045
A set of posterior probabilities corresponding to each group of training samples obtained by an NBS classifier for the ith fault type;
in the SVM classifier, 1 to 1 multi-class support vector machine classifier is adopted, a voting method is adopted to obtain a classification result, and the output probability set of the normalized classifier is as follows:
Figure FDA0002273754590000042
wherein
Figure FDA0002273754590000046
A set of posterior probabilities corresponding to each group of training samples obtained by an SVM classifier for the ith fault type;
distributing weights according to the characteristics of each fault type, weighting and adding the output probability sets of the two classifiers, and establishing a combined classifier according to class weight:
Figure FDA0002273754590000043
the label class corresponding to the maximum probability value of each row of the probability set is the label class corresponding to each group of test feature vectors, namely the classification result of the joint classifier, namely the label class set output by the joint classifier:
Figure FDA0002273754590000044
seeking an optimal weight set, fitness function and through a particle swarm optimization algorithmThe uniform weighted joint classifiers are the same, and the optimal weight set solving process is converted into a joint classifier output mark class set
Figure FDA0002273754590000047
And a test sample data mark class set Y (Y)1,Y2,...,Ym) The number of elements corresponding to each element in the group is maximized, the position and the speed of the particle swarm are continuously updated according to the fitness function, and the optimal weight coefficient set with the fitness function meeting the conditions is searched
Step four, diagnosing and classifying real-time measurement data;
in order to realize real-time detection and identification of a fault model of the in-orbit spacecraft system, historical data needs to be analyzed, and a corresponding model structure is determined when the system state is classified, namely, a mapping relation of the classification result and the model structure in one-to-one correspondence is established, so that the structure of the model can be determined after test data are classified by a classifier; therefore, when a large amount of simulation data are used for training the classifier, the structure of the model needs to be trained, a system state-structure mapping table is established, and the running state of the system corresponds to the model structure one by one;
when the model structure is known, the model structure of the system in different states can be obtained according to the physical performance analysis of the model, and a system state-structure mapping table is directly established; when the model structure is unknown, identifying the model structure of the known fault type by adopting a large number of training data sample sets, and identifying the structure of training sample data by adopting an AIC information criterion in consideration of the complexity of the model;
step five, identifying the state model on line;
after the model structure of the system is determined through the system state classification and structure identification in the first step to the fourth step, online estimation is carried out on the measurement data acquired in real time by adopting a recursion method to obtain a new estimation value of the model; in order to optimize the identification efficiency, the model parameters are identified by adopting a classical least square parameter identification method, so that the online identification of the model parameters is realized.
CN201610397845.8A 2016-06-07 2016-06-07 Multi-combination classifier based data driving system state model online identification method Expired - Fee Related CN106156401B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610397845.8A CN106156401B (en) 2016-06-07 2016-06-07 Multi-combination classifier based data driving system state model online identification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610397845.8A CN106156401B (en) 2016-06-07 2016-06-07 Multi-combination classifier based data driving system state model online identification method

Publications (2)

Publication Number Publication Date
CN106156401A CN106156401A (en) 2016-11-23
CN106156401B true CN106156401B (en) 2020-02-18

Family

ID=57352873

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610397845.8A Expired - Fee Related CN106156401B (en) 2016-06-07 2016-06-07 Multi-combination classifier based data driving system state model online identification method

Country Status (1)

Country Link
CN (1) CN106156401B (en)

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108255522A (en) * 2016-12-27 2018-07-06 北京金山云网络技术有限公司 A kind of application program sorting technique and device
CN106844775B (en) * 2017-03-02 2021-03-16 上海交通大学 Spacecraft fault rapid detection system
CN108791954A (en) * 2018-05-30 2018-11-13 兰州空间技术物理研究所 A kind of fault early warning method based on the in-orbit interior charged effect dynamic base table of spacecraft
CN110046256A (en) * 2019-04-22 2019-07-23 成都四方伟业软件股份有限公司 The prediction technique and device of case differentiation result
CN110162013A (en) * 2019-05-29 2019-08-23 上海理工大学 A kind of breakdown of refrigeration system diagnostic method
CN110162014A (en) * 2019-05-29 2019-08-23 上海理工大学 A kind of breakdown of refrigeration system diagnostic method of integrated multi-intelligence algorithm
CN110837769B (en) * 2019-08-13 2023-08-29 中山市三卓智能科技有限公司 Image processing and deep learning embedded far infrared pedestrian detection method
CN111259604A (en) * 2020-01-16 2020-06-09 中国科学院空间应用工程与技术中心 High orbit satellite light pressure model identification method and system based on machine learning
CN111289892B (en) * 2020-04-07 2022-07-01 南京机电职业技术学院 High-voltage circuit breaker fault diagnosis method based on PCA-SVM
CN112116598A (en) * 2020-08-04 2020-12-22 北京农业信息技术研究中心 Flower type identification method and system
CN114611743A (en) * 2020-12-08 2022-06-10 中兴通讯股份有限公司 System performance optimization method and device, electronic equipment and readable medium thereof
CN113569424B (en) * 2021-08-05 2022-06-07 北京微纳星空科技有限公司 Satellite operation simulation method and device, electronic equipment and storage medium

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104134010A (en) * 2014-08-12 2014-11-05 哈尔滨工业大学 Satellite fault diagnosis method for discrete type data based on Naive Bayes
CN104899456A (en) * 2015-06-15 2015-09-09 西北工业大学 Real-time correction method for on-orbit spacecraft model

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104134010A (en) * 2014-08-12 2014-11-05 哈尔滨工业大学 Satellite fault diagnosis method for discrete type data based on Naive Bayes
CN104899456A (en) * 2015-06-15 2015-09-09 西北工业大学 Real-time correction method for on-orbit spacecraft model

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
《Transformer fault portfolio diagnosis based on the combination of the multiple Bayesian classifier and SVM》;Zhongli Wu等;《2009 International Conference on Electronic Computer Technology》;20090227;第379-382页 *
《基于多分类器组合的水电机组故障诊断》;田源等;《水力发电》;20130430;第39卷(第4期);第51-54页 *
《基于多种贝叶斯分类器和SVM相结合的变压器故障组合诊断新思想》;章斌等;《2009年度电力行业高压测试及电气测量技术研讨会》;20100506;第53-58页 *
《基于组合多分类器的变压器故障诊断方法研究》;汪凤;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20140515(第5期);第C042-148页 *

Also Published As

Publication number Publication date
CN106156401A (en) 2016-11-23

Similar Documents

Publication Publication Date Title
CN106156401B (en) Multi-combination classifier based data driving system state model online identification method
Miao et al. Joint learning of degradation assessment and RUL prediction for aeroengines via dual-task deep LSTM networks
CN108594788B (en) Airplane actuator fault detection and diagnosis method based on deep random forest algorithm
CN101907681B (en) Analog circuit dynamic online failure diagnosing method based on GSD-SVDD
CN109766583A (en) Based on no label, unbalanced, initial value uncertain data aero-engine service life prediction technique
CN111860982A (en) Wind power plant short-term wind power prediction method based on VMD-FCM-GRU
Ayodeji et al. Causal augmented ConvNet: A temporal memory dilated convolution model for long-sequence time series prediction
CN105938116A (en) Gas sensor array concentration detection method based on fuzzy division and model integration
CN112765890B (en) Dynamic domain adaptive network-based multi-working-condition rotating machine residual life prediction method
CN116448419A (en) Zero sample bearing fault diagnosis method based on depth model high-dimensional parameter multi-target efficient optimization
CN110880369A (en) Gas marker detection method based on radial basis function neural network and application
CN111638034B (en) Strain balance temperature gradient error compensation method and system based on deep learning
Ali et al. Clustering sparse sensor placement identification and deep learning based forecasting for wind turbine wakes
CN111680875A (en) Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model
CN112507479B (en) Oil drilling machine health state assessment method based on manifold learning and softmax
Zhang et al. Fault detection and diagnosis for data incomplete industrial systems with new Bayesian network approach
CN115757103A (en) Neural network test case generation method based on tree structure
Tamilselvan et al. Multi-sensor health diagnosis using deep belief network based state classification
Farahani et al. A novel method for designing transferable soft sensors and its application
Zhang et al. State-of-health prediction for reaction wheel of on-orbit satellite based on fourier broad learning system
Elashmawi et al. Neural network monitoring model for industrial gas turbine
Liu et al. Industrial Big Data Analytical System in Industrial Cyber-Physical Systems Based on Coarse-to-Fine Deep Network
Yu et al. Fault detection based on a combined approach of FA-CP-ELM with application to wind turbine system
Yu et al. Fault diagnosis of analog circuit based CS_SVM algorithm
Cui et al. Prediction of Aeroengine Remaining Useful Life Based on SE-BiLSTM

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20200218

CF01 Termination of patent right due to non-payment of annual fee