CN113821012A - Fault diagnosis method for variable working condition satellite attitude control system - Google Patents

Fault diagnosis method for variable working condition satellite attitude control system Download PDF

Info

Publication number
CN113821012A
CN113821012A CN202111002442.6A CN202111002442A CN113821012A CN 113821012 A CN113821012 A CN 113821012A CN 202111002442 A CN202111002442 A CN 202111002442A CN 113821012 A CN113821012 A CN 113821012A
Authority
CN
China
Prior art keywords
domain
label
sample
fault
working condition
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.)
Granted
Application number
CN202111002442.6A
Other languages
Chinese (zh)
Other versions
CN113821012B (en
Inventor
冒泽慧
顾彧行
马亚杰
姜斌
李文博
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.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
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 Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to CN202111002442.6A priority Critical patent/CN113821012B/en
Publication of CN113821012A publication Critical patent/CN113821012A/en
Application granted granted Critical
Publication of CN113821012B publication Critical patent/CN113821012B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
    • G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
    • G05B23/0275—Fault isolation and identification, e.g. classify fault; estimate cause or root of failure

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Control Of Position, Course, Altitude, Or Attitude Of Moving Bodies (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The embodiment of the invention discloses a fault diagnosis method for a variable working condition satellite attitude control system, relates to the field of satellite fault diagnosis, and can improve the robustness and accuracy of a fault diagnosis system. The invention comprises the following steps: acquiring multi-dimensional signals of a satellite attitude control system before (source domain) and after (target domain) working condition changes, preprocessing the acquired signals to create a multi-dimensional off-line training database, constructing a domain distributed countermeasure adaptive deep neural network model through the multi-dimensional data, increasing the countermeasure training of each type of samples on the basis of full-local countermeasure, performing dynamic weight adjustment, and obtaining a variable working condition fault diagnosis model on the database. According to the technical scheme provided by the invention, the unsupervised variable working condition self-adaptive fault diagnosis can be realized.

Description

Fault diagnosis method for variable working condition satellite attitude control system
Technical Field
The invention relates to the field of satellite fault diagnosis, in particular to a fault diagnosis method for a variable working condition satellite attitude control system.
Background
The satellite attitude control system is a key component used for controlling the satellite motion in the satellite, and ensures the stable attitude control of the satellite. As a large-scale complex system, the main system has strong relevance with the environment. The traditional satellite fault diagnosis method is usually effective for a single model under the condition that the working condition is not changed, and once the working condition is changed due to the change of the environment or the task, the original diagnosis method is usually failed. Therefore, the method for diagnosing the variable working condition fault of the satellite attitude control system has great significance for improving the reliability and effectiveness of the system and ensuring the safe operation of the satellite.
The existing satellite domain adaptive variable-working-condition-resistant fault diagnosis method based on data driving only considers the resistant training on edge distribution, but does not consider the condition distribution among different faults, lacks the weight distribution of samples under different condition distributions, finally causes the fault diagnosis effect to be limited, and is difficult to further promote. Therefore, further improvement of the fault diagnosis means of the satellite attitude control system under the variable working condition is needed.
Disclosure of Invention
The embodiment of the invention provides a fault diagnosis method for a variable working condition satellite attitude control system, which can improve the robustness and accuracy of the fault diagnosis system.
In order to achieve the above purpose, the embodiment of the invention adopts the following technical scheme:
collecting operating condition data from a satellite attitude control system, wherein the collected operating condition data comprises: a source domain signal before the working condition changes and a target domain signal after the working condition changes;
generating sample sections by using the acquired working condition data, wherein the lengths of signals in the sample sections are the same, and a fault class label and the belonging working condition information are also recorded in the sample sections;
constructing a domain distributed countermeasure adaptive deep neural network model, and performing countermeasure training on each type of sample on the basis of full-local countermeasure by using the generated sample segment to obtain a variable working condition fault diagnosis model;
and after the satellite attitude control system performs variable working condition operation, acquiring the latest working condition data, importing the latest working condition data into a variable working condition fault diagnosis model, and outputting a target domain fault diagnosis prediction result.
The embodiment of the invention discloses a variable working condition fault diagnosis scheme of a satellite attitude control system based on a domain distributed countermeasure adaptive deep neural network, which comprises the steps of collecting multi-dimensional signals of the satellite attitude control system before (source domain) and after (target domain) working condition changes, preprocessing the collected signals to create a multi-dimensional offline training database, constructing a domain distributed countermeasure adaptive deep neural network model through the multi-dimensional data, increasing countermeasure training of each type of samples on the basis of global local countermeasure, performing dynamic weight adjustment, and obtaining a variable working condition fault diagnosis model on the database. According to the technical scheme provided by the invention, unsupervised variable working condition self-adaptive fault diagnosis can be realized, and the robustness and the accuracy of a fault diagnosis system can be improved.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a schematic diagram of a process flow provided by an embodiment of the present invention;
FIG. 2 is a schematic diagram of a model structure according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a model training and application process according to an embodiment of the present invention;
Detailed Description
In order to make the technical solutions of the present invention better understood, the present invention will be described in further detail with reference to the accompanying drawings and specific embodiments. Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention. As used herein, the singular forms "a", "an", "the" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Further, "connected" or "coupled" as used herein may include wirelessly connected or coupled. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
The embodiment of the invention provides a fault diagnosis method for a variable working condition satellite attitude control system, which is realized based on a domain distributed confrontation adaptive deep neural network and is mainly used for fault diagnosis of the variable working condition satellite attitude control system, and comprises the following steps:
and S1, collecting working condition data from the satellite attitude control system.
Wherein, the operating mode data of gathering includes: the source domain signal before the working condition changes and the target domain signal after the working condition changes.
And S2, generating a sample segment by using the collected working condition data.
The lengths of the signals in the sample sections are the same, and a fault class label and the working condition information are also recorded in the sample sections.
S3, constructing a domain distributed countermeasure adaptation deep neural network model, and performing countermeasure training of each type of samples on the basis of full-local countermeasure by using the generated sample segments to obtain a variable working condition fault diagnosis model.
And S4, after the satellite attitude control system performs variable working condition operation, acquiring the latest working condition data, importing the latest working condition data into a variable working condition fault diagnosis model, and outputting a target domain fault diagnosis prediction result.
Specifically, the present embodiment can be divided into four stages, including: the method comprises a data acquisition stage, a data preprocessing stage, a model training stage and an online application stage, wherein:
1) a data acquisition stage:
acquiring multi-dimensional simulation signals of a satellite attitude control system before (source domain) and after (target domain) working condition changes;
2) a data preprocessing stage:
dividing the acquired signals into sample segments with the same length, and reserving the fault class labels and the working condition (field) information of the sample segments to form a training database;
3) a model training stage:
constructing a domain distributed countermeasure adaptive deep neural network model through a multidimensional database, increasing the countermeasure training of each type of sample on the basis of full-local countermeasure, carrying out dynamic weight adjustment, and training the network on a database to obtain a variable working condition fault diagnosis model;
4) and (3) an online application stage:
and acquiring a multidimensional simulation signal of the satellite attitude control system after changing working conditions as target domain online data, preprocessing the data to generate a test sample, and inputting the test sample into the trained domain distributed confrontation adaptive deep neural network model to obtain a target domain fault diagnosis prediction result.
In this embodiment, the signal types in the collected operating condition data include: actuator signals and sensor signals. In the data acquisition stage, multi-dimensional signals generated in the working process of a ground simulation model of the satellite attitude control system are acquired. The signals include actuator signals, sensor signals, and the like.
In this embodiment, the generating a sample segment by using the collected operating condition data includes:
and carrying out time sequence alignment on each characteristic index in the signal and carrying out default filling to obtain m-dimensional original data. And then, performing sliding window cutting with the length of c on the obtained original data along a time sequence to obtain a sample segment with the data type of m.c.
Wherein m is a positive integer, the label of the collected working condition data corresponds to different fault categories under each working condition, and c represents the length of the sample segment. Specifically, in the process of satellite simulation data preprocessing, data of each actuator and sensor index, which is m-dimensional, is obtained through simulation under each working condition and different fault category labels, time sequence alignment and default filling are carried out on each characteristic index in a signal, sliding window cutting with the length of c is carried out on processed m-dimensional original data along a time sequence, a sample segment with the data type of m.c is formed, and the sample segment comprises data points, fault category labels and working condition (field) information.
Further, the method also comprises the following steps:
and creating a training database after obtaining the sample segments, wherein the training database has the same number of sample segments before and after the working condition change, the same fault classes and the same number of sample segments corresponding to each fault class. In the training database, the source domain sample set is
Figure BDA0003235992610000051
Wherein
Figure BDA0003235992610000052
Denotes the source domain, xiFor the ith sample, yiRepresents a sample xiThe corresponding fault class label is used to identify the fault class,
Figure BDA0003235992610000053
the label of the sample is L, and L represents the shared L class of the data to be classified; a sample set of the target domain as
Figure BDA0003235992610000054
Wherein
Figure BDA0003235992610000055
Representing the target domain, N + N 'representing the total number of samples, N representing the number of source domain samples, and N' representing the number of target domain samples.
Specifically, in the process of creating a sample library, the preprocessed sample segments are created to obtain a training database, wherein the number of the sample segments before and after the working condition change is the same, the number of the fault classes is the same, and the number of the sample segments corresponding to each fault class is the same. Wherein the source domain sample set is
Figure BDA0003235992610000056
Wherein
Figure BDA0003235992610000057
Denotes the source domain, xiFor the (i) th sample,
Figure BDA0003235992610000061
for the label of the sample, L indicates that the data to be classified has a common L class. A sample set of the target domain as
Figure BDA0003235992610000062
Wherein
Figure BDA0003235992610000063
Representing the target domain, xiFor the ith sample, the target domain training set contains no label. N + N' represents the total number of samples.
In the embodiment, a domain distributed confrontation adaptive deep neural network model is mainly constructed through a multidimensional database; on the basis of global antagonism, the antagonism training of each type of sample is added, and dynamic weight adjustment is carried out; training a network on a database to obtain a trained variable working condition fault diagnosis model; the model can realize unsupervised fault diagnosis on a target domain. Wherein the constructing of the domain distributed confrontation adaptation deep neural network model comprises:
a model is built based on a deep neural network, a feature extractor, a label classifier, a whole-area classifier and a distributed domain classifier are arranged in the distributed countermeasure adaptation deep neural network model of the built domain, and the built model comprises a gradient inversion layer.
Wherein, the feature extractor GhThe hidden layer features are extracted by carrying out feature extraction on the source domain data sample x
Figure BDA0003235992610000064
Label classifier GyFor characterizing hidden layers
Figure BDA0003235992610000065
Classifying fault labels to obtain labels of different faults
Figure BDA0003235992610000066
Wherein
Figure BDA0003235992610000067
And participates in training as pseudo labels for samples in the domain classifier. Whole-local classifier
Figure BDA0003235992610000068
For applying intermediate features
Figure BDA0003235992610000069
Performing domain classification to obtain predicted labels of different domains
Figure BDA00032359926100000610
Distributed domain classifier
Figure BDA00032359926100000611
For applying intermediate features
Figure BDA00032359926100000612
Classifying the fields according to the label types to obtain predicted labels in different fields
Figure BDA00032359926100000613
Where c ═ { 1., L } represents the class label to which it belongs.
Specifically, by a feature extractor GhAnd a label classifier GyFault label classification section that collectively constitutes a model:
Figure BDA00032359926100000614
and the fault label classification part is used for carrying out supervised label prediction on the labeled data of the source domain. Whole-local classifier
Figure BDA00032359926100000615
Distributed domain classifiers corresponding to each label class
Figure BDA00032359926100000616
Performing weight combination to obtain a domain classifier GdWherein, in the step (A),
Figure BDA0003235992610000071
omega is the distribution weight and is more than or equal to 0 and less than or equal to 1 so as to adjust the weight between the distribution domain classifier and the whole local area classifier. By a feature extractor GhAnd domain classifier GdThe domain classification part that collectively constitutes the model:
Figure BDA0003235992610000072
in the domain classification part, the source domain and the target domain have no fault label, so that the domain of the source domain and the target domain is taken as a label for domain prediction.
Wherein, the feature extractor G of the modelhLabel sorter GyGlobal domain classifier
Figure BDA0003235992610000073
Distributed domain classifier
Figure BDA0003235992610000074
The network parameters in the four parts are respectively thetah、θy、
Figure BDA0003235992610000075
Wherein
Figure BDA0003235992610000076
Parameters constituting a domain classification learning part
Figure BDA0003235992610000077
Feature extractor G in fault label classification learning part and field classification learning part (including whole-area classification and distributed-type field classification)hParameter theta ofhAnd (5) the consistency is achieved.
In this embodiment, the designed specific function model includes:
trouble label classification learning section FyThe loss function of (d) is:
Figure BDA0003235992610000078
the global classification learning partial loss function is:
Figure BDA0003235992610000079
the distributed domain classification learning partial loss function is:
Figure BDA00032359926100000710
model domain classification learning part FdThe loss function of (d) is:
Figure BDA00032359926100000711
wherein, thetahNetwork parameter, theta, representing a feature extractoryNetwork parameter, θ, representing a tag classifieriParameters of the model, y, representing the input of the ith sampleiRepresents a sample xiThe corresponding fault class label is used to identify the fault class,
Figure BDA0003235992610000081
represents the ith sample participating in the global classification training,
Figure BDA0003235992610000082
representing the ith sample participating in the distributed domain classification training,
Figure BDA0003235992610000083
presentation Domain classifier GdThe network parameters of the global classification section in (1),
Figure BDA0003235992610000084
presentation Domain classifier GdNetwork parameters of the distributed classification part corresponding to the medium fault pseudo label c, diA label that represents a domain class is provided,
Figure BDA0003235992610000085
representing the network parameters of the feature extractor and the label classifier in the model after the training is finished,
Figure BDA0003235992610000086
and representing network parameters of the global and distributed domain classifiers in the model after the training is finished. It should be noted that y, h, i, although referred to as subscripts, have no individual meaning and are used only in combination with the main parameters and to indicate a complete definition.
In this embodiment, in the process of performing the countermeasure training of each type of sample on the basis of the global countermeasure by using the generated sample segment, the method includes: in the feature extractor GhAnd domain classifier GdA gradient inversion layer is introduced in between. Multiplying by a constant-k when training back propagation, wherein 0 < k < 1, makes the label classifier GyAnd domain classifier GdThe training directions of the two parts are opposite. Specifically, in order to realize the purpose of the countermeasure of the fault label classification learning part and the field classification learning part, the feature extractor GhAnd domain classifier GdA gradient inversion layer is introduced between the label classification layer and the gradient inversion layer, so that the network needs to weaken the resolution capability to the field while strengthening the accurate classification capability to the label. Multiplying by a constant-k in training back propagation, where 0 < k < 1, makes two partsThe training directions are opposite.
Specifically, the model cost function is:
Figure BDA0003235992610000087
wherein, the fault class label of the target domain sample adopts a prediction pseudo label obtained by a fault label classification learning part
Figure BDA0003235992610000088
The initial value of the distribution weight omega of the domain classification module is 1, and the updating mode after the initial iteration is
Figure BDA0003235992610000089
And k is a gradient inversion layer hyperparameter used for balancing the weight of the two subtasks.
The task optimization goal of the model is as follows:
Figure BDA00032359926100000914
Figure BDA0003235992610000091
and adopting a random gradient descent method to obtain weight parameters of the model
Figure BDA0003235992610000092
The parameter updating rule in back propagation comprises the following steps: thetai+1←θi+δθi,
Figure BDA0003235992610000093
Where λ represents the learning rate, the network training is completed when the parameters converge.
When in use
Figure BDA0003235992610000094
And
Figure BDA0003235992610000095
and after the stability tends to converge, terminating the training. And will be based on the parameters at that time
Figure BDA0003235992610000096
The determined fault label classifier is used for target domain fault diagnosis prediction. And obtaining the trained domain distributed confrontation adaptation deep neural network model. Respectively when the parameters are
Figure BDA0003235992610000097
And
Figure BDA0003235992610000098
after the stability tends to converge, the training is terminated to obtain
Figure BDA0003235992610000099
At this time parameter
Figure BDA00032359926100000910
The determined fault label classifier can implement fault diagnosis of the target domain data.
The online application phase in this embodiment may include: and collecting the multidimensional simulation signal of the satellite attitude control system after the working condition changes as online data, and generating an online application sample. And (3) inputting the on-line application sample into a supervised training part in the trained domain distributed confrontation adaptive deep neural network model for label prediction to obtain a fault diagnosis result. Obtaining a multidimensional simulation signal of the satellite attitude control system after working condition change as online data, and generating a test sample after preprocessing
Figure BDA00032359926100000911
Inputting test samples into a fault label classification learning part F in the trained domain distributed confrontation adaptive deep neural network modelyAnd (3) performing label prediction to obtain a fault diagnosis result:
Figure BDA00032359926100000912
wherein
Figure BDA00032359926100000913
Namely, the trained fault label classification part carries out label prediction on the input test sample to obtain a prediction label.
The embodiment of the invention discloses a variable working condition fault diagnosis scheme of a satellite attitude control system based on a domain distributed countermeasure adaptive deep neural network, which comprises the steps of collecting multi-dimensional signals of the satellite attitude control system before (source domain) and after (target domain) working condition changes, preprocessing the collected signals to create a multi-dimensional offline training database, constructing a domain distributed countermeasure adaptive deep neural network model through the multi-dimensional data, increasing countermeasure training of each type of samples on the basis of global local countermeasure, performing dynamic weight adjustment, and obtaining a variable working condition fault diagnosis model on the database. According to the technical scheme provided by the invention, unsupervised variable working condition self-adaptive fault diagnosis can be realized, and the robustness and the accuracy of a fault diagnosis system can be improved.
The embodiment of the invention can adjust the application mode according to specific scenes, for example, in some application scenes, the scheme can be divided into 2 stages, as shown in fig. 1, wherein the first stage is an off-line modeling stage, and a variable working condition diagnosis model is established. The second stage is an online application stage, and online unsupervised fault diagnosis is realized. The method comprises the following steps:
the method comprises the following steps that 1, multi-dimensional simulation signals of a satellite attitude control system before (a source domain) and after (a target domain) working condition changes are collected; 2, dividing the acquired signals into sample segments with the same length, and reserving the fault class labels and the working condition information of the sample segments to form a training database; 3, constructing a domain distributed confrontation adaptive deep neural network model through a multidimensional database, increasing the confrontation training of each type of sample on the basis of full-local confrontation, performing dynamic weight adjustment, and training the network on the database to obtain a variable working condition fault diagnosis model;
specifically, the acquiring step 1 acquires a multi-dimensional simulation signal of the satellite attitude control system before (source domain) and after (target domain) working condition change, and includes:
and acquiring multi-dimensional simulation signals of the satellite attitude control system before (as a source domain) and after (as a target domain) the working condition changes. And (4) simulating under each working condition and different fault category labels to obtain data m-dimensional signals of the indexes of each actuator and sensor.
Specifically, the step 2 of preprocessing the acquired source domain and target domain signals and creating an offline training database includes:
(1) and (4) preprocessing satellite simulation data. Satellite simulation data preprocessing: and simulating under each working condition and different fault type labels to obtain data of each actuator and sensor index which is m-dimensional, carrying out time sequence alignment and default filling on each characteristic index in the signal, and carrying out sliding window cutting with the length of c on the processed m-dimensional original data along a time sequence to form a sample segment with the data type of m.c, wherein the sample segment comprises data points, fault type labels and working condition (field) information.
(2) A training sample library is created. And establishing the preprocessed sample segments to obtain a training database, wherein the number of the sample segments before and after the working condition change is the same, the fault classes are the same, and the number of the sample segments corresponding to each fault class is the same. Wherein the source domain training set is
Figure BDA0003235992610000111
Wherein
Figure BDA0003235992610000112
Denotes the source domain, xiFor the (i) th sample,
Figure BDA0003235992610000113
for the label of the sample, L indicates that the data to be classified has a common L class. The target domain training set is
Figure BDA0003235992610000114
Wherein
Figure BDA0003235992610000115
Representing the target domain, xiFor the ith sample, the target domain training set contains no label. N + N' represents the total number of samples。
Specifically, as shown in fig. 2, the 3-step method includes constructing a domain distributed confrontation adaptive deep neural network model through a multidimensional database, increasing the confrontation training of each type of sample on the basis of the global confrontation, performing dynamic weight adjustment, and training a network on a database to obtain a variable working condition fault diagnosis model, and includes:
(1) and constructing a domain distributed countermeasure adaptation deep neural network model. The model is based on a deep neural network and is divided into four modules: the system comprises a feature extractor, a label classifier, a whole-area classifier and a distributed domain classifier, and comprises a gradient inversion layer.
The model is divided into four modules:
(i) feature extractor GhExtracting the characteristics of the source domain data sample x to extract the hidden layer characteristics
Figure BDA0003235992610000116
Figure BDA0003235992610000117
(ii) Label classifier GyWill imply layer characteristics
Figure BDA0003235992610000118
Classifying fault labels to obtain labels of different predicted faults
Figure BDA0003235992610000119
Figure BDA0003235992610000121
Wherein
Figure BDA0003235992610000122
Participate in the training as pseudo labels for the samples in the domain classifier.
(iii) Whole-local classifier
Figure BDA0003235992610000123
Intermediate characteristics
Figure BDA0003235992610000124
Performing domain classification to obtain predicted labels of different domains
Figure BDA0003235992610000125
Figure BDA0003235992610000126
(Vi) distributed Domain classifier
Figure BDA0003235992610000127
Intermediate characteristics
Figure BDA0003235992610000128
Classifying the fields according to the label types to obtain predicted labels in different fields
Figure BDA0003235992610000129
Figure BDA00032359926100001210
Where c ═ {1, …, L } represents the class label to which it belongs.
Of the above four parts, the feature extractor GhAnd label classifier GyFault label classification section that collectively constitutes a model:
Figure BDA00032359926100001211
the component performs supervised label prediction on the labeled data of the source domain.
Whole-local classifier
Figure BDA00032359926100001212
Distributed domain classifiers corresponding to each label class
Figure BDA00032359926100001213
Performing weight combination to obtain a domain classifier Gd。
Figure BDA00032359926100001214
Wherein omega is the distribution weight and omega is more than or equal to 0 and less than or equal to 1, and the weight between the distribution domain classifier and the whole local area classifier is adjusted. Feature extractor GhAnd domain classifier GdThe domain classification part that collectively constitutes the model:
Figure BDA00032359926100001215
the source domain and the target domain in the part have no fault label, and the domain of the source domain and the target domain is taken as a label to carry out domain prediction.
Modeled feature extractor GhLabel sorter GyGlobal domain classifier
Figure BDA00032359926100001216
Distributed domain classifier
Figure BDA00032359926100001217
The network parameters in the four parts are respectively thetah、θy、
Figure BDA00032359926100001218
Wherein
Figure BDA00032359926100001219
Parameters constituting a domain classification learning part
Figure BDA0003235992610000131
Fault label classification learning part and field classification learning part (including whole local classification and distributed domain classification)Feature extractor GhParameter theta ofhAnd (5) the consistency is achieved.
(2) Training domain distributed confrontation adapts to a deep neural network model. Fault label classification learning part F in the modelyThe loss function of (d) is:
Figure BDA0003235992610000132
the global classification learning partial loss function is:
Figure BDA0003235992610000133
the distributed domain classification learning partial loss function is:
Figure BDA0003235992610000134
model domain classification learning part FdThe loss function of (d) is:
Figure BDA0003235992610000135
in order to realize the purpose of resisting the learning of the fault label classification learning part and the field classification learning part, the feature extractor GhAnd domain classifier GdA gradient inversion layer is introduced between the label classification layer and the gradient inversion layer, so that the network needs to weaken the resolution capability to the field while strengthening the accurate classification capability to the label. Multiplying by constant-k when training is propagated reversely, wherein k is more than 0 and less than 1, and making the training directions of the two parts opposite. The model cost function is then:
Figure BDA0003235992610000136
wherein, the fault class label of the target domain sample adopts a prediction pseudo label obtained by a fault label classification learning part
Figure BDA0003235992610000137
The initial value of the distribution weight omega of the domain classification module is 1, and the updating mode after the initial iteration is
Figure BDA0003235992610000141
k is a gradient inversion layer hyper-parameter, the weights of the two subtasks are balanced, and the adjustment can be carried out manually. The task optimization goal of the model is as follows:
Figure BDA0003235992610000142
Figure BDA0003235992610000143
solving weight parameters of model by adopting random gradient descent method
Figure BDA0003235992610000144
The parameter updating rule in the back propagation is as follows:
Figure BDA0003235992610000145
where λ represents the learning rate, the network training is completed when the parameters converge.
(3) And obtaining the trained domain distributed confrontation adaptation deep neural network model. Respectively when the parameters are
Figure BDA0003235992610000146
And
Figure BDA0003235992610000147
after the stability tends to converge, the training is terminated to obtain
Figure BDA0003235992610000148
At this time parameter
Figure BDA0003235992610000149
The determined fault label classifier can implement fault diagnosis of the target domain data.
In the second phase, as shown in fig. 3, online fault diagnosis: and 4, acquiring a multidimensional simulation signal of the satellite attitude control system after working condition change as online data, preprocessing the online data to generate a test sample, inputting the test sample into the label classification part of the domain distributed confrontation adaptive deep neural network model obtained in the step 3, and obtaining a fault diagnosis result.
Specifically, the stage of inputting the test sample into the domain distributed confrontation adaptation deep neural network model obtained in the step 3 to obtain the fault diagnosis result includes:
obtaining a multidimensional simulation signal of the satellite attitude control system after working condition change as online data, and generating a test sample after preprocessing
Figure BDA00032359926100001410
Inputting test samples into a fault label classification learning part F in the trained domain distributed confrontation adaptive deep neural network modelyAnd (3) performing label prediction to obtain a fault diagnosis result:
Figure BDA0003235992610000151
wherein
Figure BDA0003235992610000152
Namely, the trained fault label classification part carries out label prediction on the input test sample to obtain a prediction label.
The domain-distributed countermeasure-adaptive deep neural network provided by the embodiment can be used as a fault detection and diagnosis algorithm for realizing fault diagnosis of the satellite attitude control system under the condition of variable working conditions. According to the method, the difference between the source domain and the target domain is reduced, the distinguishing capability of the model on two distribution domains before and after the working condition changes is weakened, and the feature extractor and the fault label classifier can classify the fault data without labels in the target domain through the label prediction function learned by the source domain data. By introducing the gradient inversion layer, the antagonistic learning of the label distribution of the source field and the domain distribution of the source field and the target field is realized. By increasing the field distributed countermeasure, the countermeasure distribution weight can be dynamically adjusted, and the importance of condition distribution is improved. It may help the algorithm to generalize better over the target domain than using only the source data. The method achieves higher fault diagnosis accuracy after the working state changes.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the apparatus embodiment, since it is substantially similar to the method embodiment, it is relatively simple to describe, and reference may be made to some descriptions of the method embodiment for relevant points. The above description is only for the specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. A fault diagnosis method for a variable working condition satellite attitude control system is characterized by comprising the following steps:
collecting operating condition data from a satellite attitude control system, wherein the collected operating condition data comprises: a source domain signal before the working condition changes and a target domain signal after the working condition changes;
generating sample sections by using the acquired working condition data, wherein the lengths of signals in the sample sections are the same, and a fault class label and the belonging working condition information are also recorded in the sample sections;
constructing a domain distributed countermeasure adaptive deep neural network model, and performing countermeasure training on each type of sample on the basis of full-local countermeasure by using the generated sample segment to obtain a variable working condition fault diagnosis model;
and after the satellite attitude control system performs variable working condition operation, acquiring the latest working condition data, importing the latest working condition data into a variable working condition fault diagnosis model, and outputting a target domain fault diagnosis prediction result.
2. The method of claim 1, wherein the signal types in the collected operating condition data comprise: actuator signals and sensor signals.
3. The method of claim 1, wherein generating a sample segment using the collected operating condition data comprises:
carrying out time sequence alignment on various characteristic indexes in the signals and carrying out default filling to obtain m-dimensional original data, wherein m is a positive integer, and labels of the collected working condition data correspond to different fault categories under each working condition;
and c, performing sliding window cutting on the obtained original data along the time sequence to obtain a sample segment with the data type of m & c, wherein c represents the length of the sample segment.
4. The method of claim 3, further comprising:
creating a training database after obtaining the sample segments, wherein the training database has the same number of sample segments before and after the working condition change, the same fault class and the same number of sample segments corresponding to each fault class;
in the training database, the source domain sample set is
Figure FDA0003235992600000011
Wherein
Figure FDA0003235992600000012
Denotes the source domain, xiFor the ith sample, yiRepresents a sample xiThe corresponding fault class label is used to identify the fault class,
Figure FDA0003235992600000021
the label of the sample is L, and L represents the shared L class of the data to be classified; a sample set of the target domain as
Figure FDA0003235992600000022
Wherein
Figure FDA0003235992600000023
Representing the target domain, N + N 'representing the total number of samples, N representing the number of source domain samples, and N' representing the number of target domain samples.
5. The method of claim 3, wherein constructing the domain distributed countermeasure adaptation deep neural network model comprises:
establishing a model based on a deep neural network, and setting a feature extractor, a label classifier, a full-local classifier and a distributed domain classifier in the model, wherein the established model comprises a gradient inversion layer;
wherein, the feature extractor GhThe hidden layer features are extracted by carrying out feature extraction on the source domain data sample x
Figure FDA0003235992600000024
Label classifier GyFor characterizing hidden layers
Figure FDA0003235992600000025
Classifying fault labels to obtain labels of different faults
Figure FDA0003235992600000026
Wherein
Figure FDA0003235992600000027
And the pseudo label is used as a sample in a domain classifier to participate in training;
whole-local classifier
Figure FDA0003235992600000028
For applying intermediate features
Figure FDA0003235992600000029
Performing domain classification to obtain predicted labels of different domains
Figure FDA00032359926000000210
Distributed domain classifier
Figure FDA00032359926000000211
For applying intermediate features
Figure FDA00032359926000000212
Classifying the fields according to the label types to obtain predicted labels in different fields
Figure FDA00032359926000000213
Where c ═ { 1., L } represents the class label to which it belongs.
6. The method of claim 5, characterized by a feature extractor GhAnd a label classifier GyFault label classification section that collectively constitutes a model:
Figure FDA00032359926000000214
the fault label classification part is used for carrying out supervised label prediction on the labeled data of the source domain;
whole-local classifier
Figure FDA00032359926000000215
Distributed domain classifiers corresponding to each label class
Figure FDA00032359926000000216
Performing weight combination to obtain a domain classifier GdWherein, in the step (A),
Figure FDA00032359926000000217
omega is distribution weight and is more than or equal to 0 and less than or equal to 1;
by a feature extractor GhAnd domain classifier GdThe domain classification part that collectively constitutes the model:
Figure FDA0003235992600000031
in the domain classification section, the source domain and the target domain are free of fault labels.
7. The method of claim 6, wherein the fault label classification learning component FyThe loss function of (d) is:
Figure FDA0003235992600000032
the global classification learning partial loss function is:
Figure FDA0003235992600000033
the distributed domain classification learning partial loss function is:
Figure FDA0003235992600000034
model domain classification learning part FdThe loss function of (d) is:
Figure FDA0003235992600000035
wherein, thetahNetwork parameter, theta, representing a feature extractoryNetwork parameters representing tag classifiersNumber, thetaiParameters of the model, y, representing the input of the ith sampleiRepresents a sample xiThe corresponding fault class label is used to identify the fault class,
Figure FDA0003235992600000036
represents the ith sample participating in the global classification training,
Figure FDA0003235992600000037
representing the ith sample participating in the distributed domain classification training,
Figure FDA0003235992600000038
presentation Domain classifier GdThe network parameters of the global classification section in (1),
Figure FDA0003235992600000039
presentation Domain classifier GdNetwork parameters of the distributed classification part corresponding to the medium fault pseudo label c, diA label that represents a domain class is provided,
Figure FDA00032359926000000310
representing the network parameters of the feature extractor and the label classifier in the model after the training is finished,
Figure FDA00032359926000000311
and representing network parameters of the global and distributed domain classifiers in the model after the training is finished.
8. The method according to claim 6 or 7, wherein in the course of performing the confrontation training of each type of sample on the basis of the global confrontation by using the generated sample segment, the method comprises:
in the feature extractor GhAnd domain classifier GdIntroducing a gradient inversion layer in between;
multiplying by a constant-k when training back propagation, wherein 0 < k < 1, makes the label classifier GyAnd domain classifier GdThese two-part training methodThe opposite direction is used.
9. The method of claim 8, wherein the model cost function is:
Figure FDA0003235992600000041
wherein, the fault class label of the target domain sample adopts a prediction pseudo label obtained by a fault label classification learning part
Figure FDA0003235992600000042
The initial value of the distribution weight omega of the domain classification module is 1, and the updating mode after the initial iteration is
Figure FDA0003235992600000043
k is a gradient inversion layer hyper-parameter used for balancing the weight of the two subtasks;
the task optimization goal of the model is as follows:
Figure FDA0003235992600000044
Figure FDA0003235992600000045
and adopting a random gradient descent method to obtain weight parameters of the model
Figure FDA0003235992600000046
The parameter updating rule in back propagation comprises the following steps:
Figure FDA0003235992600000047
where λ represents the learning rate, the network training is completed when the parameters converge.
10. The method of claim 9Is characterized in that when
Figure FDA0003235992600000048
And
Figure FDA0003235992600000049
after the stability tends to convergence, terminating the training;
and will be based on the parameters at that time
Figure FDA0003235992600000051
The determined fault label classifier is used for target domain fault diagnosis prediction.
CN202111002442.6A 2021-08-30 2021-08-30 Fault diagnosis method for variable-working-condition satellite attitude control system Active CN113821012B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111002442.6A CN113821012B (en) 2021-08-30 2021-08-30 Fault diagnosis method for variable-working-condition satellite attitude control system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111002442.6A CN113821012B (en) 2021-08-30 2021-08-30 Fault diagnosis method for variable-working-condition satellite attitude control system

Publications (2)

Publication Number Publication Date
CN113821012A true CN113821012A (en) 2021-12-21
CN113821012B CN113821012B (en) 2023-12-29

Family

ID=78923369

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111002442.6A Active CN113821012B (en) 2021-08-30 2021-08-30 Fault diagnosis method for variable-working-condition satellite attitude control system

Country Status (1)

Country Link
CN (1) CN113821012B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110751207A (en) * 2019-10-18 2020-02-04 四川大学 Fault diagnosis method for anti-migration learning based on deep convolution domain
CN111651937A (en) * 2020-06-03 2020-09-11 苏州大学 Method for diagnosing similar self-adaptive bearing fault under variable working conditions
CN112665852A (en) * 2020-11-30 2021-04-16 南京航空航天大学 Variable working condition planetary gearbox fault diagnosis method and device based on deep learning
CN112784920A (en) * 2021-02-03 2021-05-11 湖南科技大学 Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN113095413A (en) * 2021-04-14 2021-07-09 山东建筑大学 Variable working condition fault diagnosis method, system, storage medium and equipment
CN113158878A (en) * 2021-04-19 2021-07-23 合肥工业大学 Heterogeneous migration fault diagnosis method, system and model based on subspace

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110751207A (en) * 2019-10-18 2020-02-04 四川大学 Fault diagnosis method for anti-migration learning based on deep convolution domain
CN111651937A (en) * 2020-06-03 2020-09-11 苏州大学 Method for diagnosing similar self-adaptive bearing fault under variable working conditions
CN112665852A (en) * 2020-11-30 2021-04-16 南京航空航天大学 Variable working condition planetary gearbox fault diagnosis method and device based on deep learning
CN112784920A (en) * 2021-02-03 2021-05-11 湖南科技大学 Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN113095413A (en) * 2021-04-14 2021-07-09 山东建筑大学 Variable working condition fault diagnosis method, system, storage medium and equipment
CN113158878A (en) * 2021-04-19 2021-07-23 合肥工业大学 Heterogeneous migration fault diagnosis method, system and model based on subspace

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
YUXING GU等: "Fault Diagnosis of Satellites under Variable Conditions based on Domain Adaptive Adversarial Deep Neural Network", 《2021 IEEE 10TH DATA DRIVEN CONTROL AND LEARNING SYSTEM CONFERENCE》 *
YUXING GU等: "Fault Diagnosis of Satellites under Variable Conditions based on Domain Adaptive Adversarial Deep Neural Network", 《2021 IEEE 10TH DATA DRIVEN CONTROL AND LEARNING SYSTEM CONFERENCE》, 31 May 2021 (2021-05-31), pages 1492 - 1497, XP033930557, DOI: 10.1109/DDCLS52934.2021.9455711 *
岳帅旭等: "深度对抗迁移学习的故障诊断方法研究", 《机械科学与技术》 *
岳帅旭等: "深度对抗迁移学习的故障诊断方法研究", 《机械科学与技术》, 15 April 2021 (2021-04-15) *
李霁蒲等: "一种用于主轴轴承故障诊断的深度卷积动态对抗迁移网络", 《振动工程学报》 *
李霁蒲等: "一种用于主轴轴承故障诊断的深度卷积动态对抗迁移网络", 《振动工程学报》, 3 June 2021 (2021-06-03), pages 1 - 8 *

Also Published As

Publication number Publication date
CN113821012B (en) 2023-12-29

Similar Documents

Publication Publication Date Title
Shao et al. Generative adversarial networks for data augmentation in machine fault diagnosis
CN107194433A (en) A kind of Radar range profile&#39;s target identification method based on depth autoencoder network
CN103728551B (en) A kind of analog-circuit fault diagnosis method based on cascade integrated classifier
CN109765333A (en) A kind of Diagnosis Method of Transformer Faults based on GoogleNet model
CN107677472A (en) The bearing state noise diagnostics algorithm that network-oriented Variable Selection merges with Characteristic Entropy
CN108537259A (en) Train control on board equipment failure modes and recognition methods based on Rough Sets Neural Networks model
CN109740057B (en) Knowledge extraction-based enhanced neural network and information recommendation method
CN109214460A (en) Method for diagnosing fault of power transformer based on Relative Transformation Yu nuclear entropy constituent analysis
CN112541572A (en) Residual oil distribution prediction method based on convolutional encoder-decoder network
CN107423747A (en) A kind of conspicuousness object detection method based on depth convolutional network
Zheng et al. Synthetic dynamic PMU data generation: A generative adversarial network approach
CN114692310B (en) Dueling DQN-based virtual-real fusion primary separation model parameter optimization method
CN105678381A (en) Gender classification network training method, gender classification method and related device
CN114038055A (en) Image generation method based on contrast learning and generation countermeasure network
CN112017255A (en) Method for generating food image according to recipe
CN104598925A (en) Multiclass Adaboost integrated studying method based on ELM
CN114169396B (en) Training data generation model construction method and application for aircraft fault diagnosis
CN110598737A (en) Online learning method, device, equipment and medium of deep learning model
Li et al. Adaptive exploration policy for exploration–exploitation tradeoff in continuous action control optimization
CN113869451A (en) Rolling bearing fault diagnosis method under variable working conditions based on improved JGSA algorithm
CN113821012A (en) Fault diagnosis method for variable working condition satellite attitude control system
CN114880767B (en) Aero-engine residual service life prediction method based on attention mechanism Dense-GRU network
CN112085079B (en) Rolling bearing fault diagnosis method based on multi-scale and multi-task learning
CN115952493A (en) Reverse attack method and attack device for black box model and storage medium
CN113688867B (en) Cross-domain image classification method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant