Satellite fault diagnosis method with variable operating conditions based on distributed domain cooperative game adversarial network
By adopting a distributed domain collaborative game adversarial network model in satellite fault diagnosis, the problem of insufficient robustness and accuracy of traditional fault diagnosis models under variable operating conditions is solved, and higher fault diagnosis accuracy is achieved.
Patent Information
- Application Number
- CN202111564153.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-20
- Publication Date
- 2025-03-28
- Estimated Expiration
- 2041-12-20
AI Technical Summary
Under variable operating conditions, traditional deep learning fault diagnosis models show insufficient robustness and accuracy in adaptive fault diagnosis of satellites.
A variable-condition satellite fault diagnosis method based on a distributed domain collaborative game adversarial network is adopted. By collecting multi-dimensional signals under different operating conditions, a database is built, and a distributed domain collaborative game adversarial network model is established. The model is trained through the game adversarial iterative method, as a variable-condition fault diagnosis model.
It improves the robustness and accuracy of the adaptive satellite fault diagnosis system, and can achieve high fault diagnosis accuracy after the operating state changes.
Smart Images
Figure CN114332528B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of satellite fault diagnosis, and in particular to a variable operating condition satellite fault diagnosis method based on a distributed domain collaborative game confrontation network. Background Art
[0002] In recent years, with the continuous development and popularization of deep learning, it has begun to be applied to large-scale and complex systems such as satellites, and its fault diagnosis capabilities have been significantly improved.
[0003] Traditional machine learning and deep learning algorithms assume that the test signals used in their applications come from the same distribution as those used during training. However, the satellite’s operating environment is harsh and changeable, the mission conditions are complex, and there are large changes in operating conditions, which results in a significant drop in the performance of the original fault diagnosis model. Although the deep learning diagnosis method based on transfer learning has achieved certain results in the case of unchanged domains, its simple model structure often leads to poor parameter adjustability and poor results for new target domains. This problem is reflected in the following practical applications: for unsupervised variable operating conditions, the robustness and accuracy of the adaptive satellite fault diagnosis system are poor. Summary of the invention
[0004] The embodiments of the present invention provide a variable-operating-condition satellite fault diagnosis method based on a distributed domain collaborative game adversarial network, which can improve the robustness and accuracy of an adaptive satellite fault diagnosis system.
[0005] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:
[0006] S1. When the working condition of the satellite changes, multi-dimensional signals under different working conditions are collected, and a database is constructed according to the collected multi-dimensional signals;
[0007] S2. Establish a distributed domain collaborative game adversarial network model;
[0008] S3, training the distributed domain collaborative game adversarial network model through the database constructed in S1 by game adversarial iteration, and using the trained distributed domain collaborative game adversarial network model as a variable operating condition fault diagnosis model;
[0009] S4. Using the variable operating condition fault diagnosis model obtained through training, a fault label classifier is established, and fault diagnosis is performed through the fault label classifier.
[0010] The variable operating condition satellite fault diagnosis method based on the distributed domain collaborative game adversarial network provided by the embodiment of the present invention can be used to realize the fault diagnosis of the satellite attitude control system under variable operating conditions. In the scheme of this embodiment, by reducing the difference between the source domain and the target domain, the model's ability to distinguish the two distribution fields before and after the operating condition changes is weakened, so that the feature extractor and the fault label classifier can classify the unlabeled fault data in the target domain through the label prediction function learned by the source domain data. By introducing the gradient reversal layer, adversarial learning of the source domain label distribution and the domain distribution of the source domain and the target domain is realized. By increasing the domain distributed adversarial, the adversarial distribution weight can be dynamically adjusted to increase the importance of the conditional distribution. It can help the algorithm to be better promoted in the target domain than using only the source data. A higher fault diagnosis accuracy is achieved after the working state changes. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0012] Figure 1 A schematic diagram of a process flow of a specific example provided for an embodiment of the present invention;
[0013] Figure 2 A schematic diagram of a model structure provided by an embodiment of the present invention;
[0014] Figure 3 A schematic diagram of a method flow chart provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0015] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention will be described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention. It can be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "one", "said" and "the" used herein may also include plural forms. It should be further understood that the term "including" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used here may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as generally understood by those of ordinary skill in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.
[0016] In view of the distribution differences of data under variable operating conditions and taking into account the conditional distribution differences between fault samples of different types, in order to make up for the deficiencies of the prior art, the design purpose of this embodiment is to: diagnose satellite faults under variable operating conditions based on a distributed domain collaborative game adversarial network, realize unsupervised fault diagnosis in the target domain, dynamically adjust the adversarial weights for samples of different types, and dynamically adjust the parameters of each hidden layer in the feature extraction to achieve synergy between the classification enhancement part and the adversarial part, so as to solve the problem of satellite fault diagnosis under changing operating conditions.
[0017] The general design ideas include: obtaining multi-dimensional signals in the source and target domains before and after the satellite working condition changes, and constructing the signals into a sample-label multi-dimensional database; constructing a distributed domain collaborative game adversarial network, and constructing a collaborative game adversarial framework with each hidden layer feature and each type of source domain sample, adding dynamic adjustment parameters for dynamic weight adjustment, and training on the corresponding multi-dimensional database to obtain a variable working condition fault diagnosis model; the trained model fault label classifier can be used for fault diagnosis of target domain signals. In this way, the dynamic adjustment of each layer parameter in the deep network structure can be realized, and unsupervised variable working condition adaptive fault diagnosis can be achieved, which improves the robustness and accuracy of the fault diagnosis system.
[0018] The embodiment of the present invention provides a method for diagnosing satellite faults under variable working conditions based on a distributed domain cooperative game confrontation network. Figure 3 As shown, including:
[0019] S1. When the working condition of the satellite changes, multi-dimensional signals under different working conditions are collected, and a database is constructed based on the collected multi-dimensional signals.
[0020] S2. Establish a distributed domain collaborative game adversarial network model.
[0021] S3. Using the database constructed in S1, the distributed domain collaborative game adversarial network model is trained by game adversarial iteration, and the trained distributed domain collaborative game adversarial network model is used as a variable operating condition fault diagnosis model.
[0022] Among them, the training is completed when the parameters of the distributed domain collaborative game adversarial network model converge. Variable operating conditions refer to the change of operating conditions from the source domain to the target domain, and the ultimate goal is to act on the target domain. Therefore, for those skilled in the art who understand transfer learning (variable operating condition learning), the "variable operating condition fault diagnosis model" can also be called the "target domain fault diagnosis model."
[0023] S4. Using the variable operating condition fault diagnosis model obtained through training, a fault label classifier is established, and fault diagnosis is performed through the fault label classifier.
[0024] In practical applications, this embodiment can be divided into at least three stages, namely, data collection and preprocessing, model building stage, and model training stage. Among them, data collection and preprocessing refers to: collecting multidimensional signals of satellites in two different working conditions before and after the working condition changes, as source domain and target domain signals, and dividing them into time segments of the same size to obtain a sample-label multidimensional database, where the source domain sample category labels are known and the target domain sample category labels are initially empty; the model building stage refers to: building a distributed domain collaborative game adversarial network model; the model training stage refers to: using the multidimensional database, using game adversarial iterative training model, until the training is completed when the parameters converge to obtain the target domain fault diagnosis model.
[0025] This embodiment can be used to implement fault diagnosis of satellite attitude control systems under variable working conditions. In the scheme of this embodiment, by reducing the difference between the source domain and the target domain, the model's ability to distinguish between the two distribution domains before and after the working condition changes is weakened, so that the feature extractor and the fault label classifier can classify the unlabeled fault data in the target domain through the label prediction function learned by the source domain data. By introducing the gradient reversal layer, adversarial learning of the source domain label distribution and the domain distribution of the source domain and the target domain is achieved. By increasing the domain distribution adversarial, the adversarial distribution weight can be dynamically adjusted to increase the importance of the conditional distribution. It can help the algorithm to generalize better on the target domain than using only the source data. A higher fault diagnosis accuracy is achieved after the working state changes.
[0026] In this embodiment, in S1, the acquisition of multidimensional signals under different working conditions includes: acquiring multidimensional signals of the satellite before and after the working condition changes; dividing the acquired multidimensional signals into time segments of the same size, and using the obtained time segments as samples to construct the database. Among them, the source domain samples include: time segments belonging to the source domain multidimensional signals, and the sample category labels added to each time segment are known, and the target domain samples include: time segments belonging to the target domain multidimensional signals, and the sample category labels added to each time segment are empty. For example: obtain the source domain and target domain multidimensional signals before and after the satellite working condition changes, and construct the signals into a sample-label multidimensional database; collect the multidimensional signals of the satellite in two different working conditions before and after the working condition changes, as the source domain and target domain signals, and divide them into time segments of the same size to obtain a sample-label multidimensional database, wherein the source domain sample category labels are known, and the target domain sample category labels are initially empty. Specifically, in the data acquisition and preprocessing stage: multidimensional signals of the satellite in two different working conditions before and after the working condition changes are collected as source domain and target domain signals, and divided into time segments of the same size to obtain a sample-label multidimensional database, wherein the source domain sample category labels are known, and the target domain sample category labels are empty, and the signals include satellite actuator signals and sensor signals, etc.
[0027] In this embodiment, the multidimensional signals of the satellite before and after the operating condition changes include: collecting m-dimensional simulation signals of the satellite before and after the operating condition changes, and using the m-dimensional simulation signal before the operating condition changes as the source domain multidimensional signal, and using the m-dimensional simulation signal after the operating condition changes as the target domain multidimensional signal.
[0028] Specifically, the method of dividing the collected multidimensional signal into time segments of the same size and using the obtained time segments as samples to construct the database includes: dividing the collected multidimensional signal into time segments of size c, and the constructed database is a sample-label m-dimensional database, wherein:
[0029] The source domain samples are x i is the i-th sample, the sample fault category label y i For known labels, It contains L types of fault class labels, i is a positive integer between 1 and n, and n is the total number of source domain samples. is the source domain; D is the domain subscript, which is used to distinguish parameter symbols and has no special meaning.
[0030] The target domain sample is Sample fault category label Initially empty, N = n + n' represents the total number of samples, n' is the number of samples in the target domain, T D is the target domain, which is updated to the pseudo label obtained by the fault label classifier during training.
[0031] In this embodiment, in the process of establishing the distributed domain collaborative game adversarial network model, it includes: forming a fault discrimination network through a feature extractor and a fault label classifier based on a multi-layer deep neural network. Among them, the deep neural network is a multi-layer neural network, which includes multiple hidden layers. Therefore, it is called "multi-layer deep neural network" in this embodiment, which is also more in line with the wording habits of technicians in this field. The feature extractor and the domain discriminator form a distributed domain discrimination network, and each level of hidden layer in the feature extractor and the domain discriminator form a collaborative domain discrimination network. Among them, the feature extractor extracts features for all sample features that do not contain fault class labels and each type of sample with fault class labels. The feature extractor of this patent refers to a deep neural network. According to the common understanding in the industry, the output obtained in the hidden layer of the network after the original data of the sample passes through the network is called a feature. The domain discriminator performs a binary classification on whether the sample belongs to the source domain or the target domain. The feature extractor adopts a multi-layer deep neural network structure, and each hidden layer and the corresponding layer of the collaborative domain discriminator form a collaborative domain discrimination network.
[0032] In this embodiment, the multidimensional database is used to iteratively train the model using game adversarial training until the training is completed when the parameters converge to obtain the target domain fault diagnosis model; wherein, dynamic adjustment parameters can be added for dynamic weight adjustment, and the variable operating condition fault diagnosis model is trained on the multidimensional database, and the model fault label classifier obtained by training can be used for fault diagnosis of target domain signals.
[0033] The various parts of the above network work together to form a distributed domain collaborative game adversarial network. The model is trained through a multi-dimensional database for game adversarial training. After iterative training until all network parameters converge, the fault discrimination network can achieve fault classification of target domain signals. This requires designing the cost function of the collaborative part of the distributed domain collaborative game adversarial network model, as well as the cost functions of the fault label classifier and the domain discriminator.
[0034] Specifically, in S3, the process of training the distributed domain collaborative game adversarial network model through game adversarial iteration includes: loading the cost functions of the fault label classifier and the domain discriminator respectively, wherein the types of domain discriminators include global domain discriminators, distributed domain discriminators and collaborative domain discriminators.
[0035] Among them, the cost function of the collaborative part in the distributed domain collaborative game adversarial network model is: J task represents the classification part cost function, represents the feature domain classification cost function from layer 1 to r-1, is the global domain (rth layer) classification cost function, represents the domain classification cost function corresponding to each type of fault, μ=(μ1,...,μ r ) is the collaborative game weight parameter, i.e., the feature extractor The weight value of the k-th layer feature, represents the domain classification cost function parameter of the rth hidden layer feature of the feature extractor composed of a deep neural network, μ r represents the weight of the feature when the feature extractor passes the feature to the global domain discriminator, represents the network parameters of the first k layers in the feature extractor, represents the network parameters of the global domain discriminator, represents the parameters of the collaborative domain discriminator corresponding to the output features of the k-th layer of the feature extractor, represents the parameters of the corresponding distributed domain discriminator belonging to the jth type of fault, It represents the collaborative domain discrimination cost function corresponding to the features of the k-th layer of the feature extractor. j and k are both positive integers. j is used to represent the j-th type of fault, k is used to represent the number of layers, and ω represents the weight parameter between the global domain discriminator and the distributed domain discriminator. The adaptive adjustment weakens the domain discrimination effect so that the trained network cannot discriminate the domain as much as possible.
[0036] For the fault label classifier G y , select the softmax classification function, and the cost function J used for model classification task for: in, Represents the loss function for each sample:
[0037] make Represents the classification label predicted by the model, i∈(1,n) Used to train the classification fault discrimination network, i∈(n+1,N) The predicted pseudo-labels for the target domain participate in the domain discrimination network training, i represents the i-th sample, n is a positive integer, represents a feature extractor based on an r-layer deep network, x i represents the i-th sample, y i Represents the corresponding fault class label of the i-th sample.
[0038] The domain discriminator in this embodiment includes: a global domain discriminator, a distributed domain discriminator and a collaborative domain discriminator, and its domain discrimination cost function is: The domain discrimination cost function of the global domain discriminator for: represents the loss function for each sample, d i Represents sample x i The corresponding domain label, that is, the domain to which it belongs.
[0039] The domain discrimination cost function of the distributed domain discriminator for: in, Represents the loss function for each sample:
[0040] in, The i-th sample is classified according to the fault category label and belongs to the j-th category, that is, the sample classified according to the fault category label, and the source domain sample label is its own label y i The target domain sample label is the pseudo label obtained by the fault discriminator
[0041] The domain discrimination cost function in the kth hidden layer of the collaborative domain discriminator is for: in, Represents the loss function for each sample: Where k = {1,...,r-1} represents the kth hidden layer in the feature extractor;
[0042] Specifically, Figure 2 As shown, in the distributed domain collaborative game adversarial network model, the feature extractor based on a multi-layer deep neural network is used There are r hidden layers in total, r is a positive integer, and k∈{1,...,r} uses the softmax classification function; represents the global domain discriminator, that is, the domain discriminator corresponding to the r-th layer feature of the feature extraction layer, represents the distributed domain discriminator corresponding to the j-th fault label, where j∈{1,...,L}; Indicates the domain discriminator corresponding to the k-th feature of the feature extraction layer in the collaborative domain discriminator, where k∈{1,...,r-1}. In this embodiment, the letters h, y, d, etc. marked in the subscripts of the parameter symbols are used to match the subscripts used in parameter naming, and are used to distinguish the parameter symbols without any special meaning.
[0043] In this embodiment, for the distributed domain collaborative game adversarial network model, the cost function J is designed total The adaptive weight parameter μ in r ), where {μ1,...,μ r-1} is the collaborative domain discriminator parameter, which is constrained as And |μ k |≤μ0, when μ k > 0 is equivalent to disabling the gradient reversal layer. k When μ<0, it is equivalent to enabling the gradient reversal layer; r The parameters of the distributed domain discriminator and the global domain discriminator are used, and the gradient reversal layer is used, where ω is the weight hyperparameter of the two, and its initial value is 1. The iterative expression in each training process is Since ω does not participate in the training process, it will not increase the computational complexity of the model. The model optimization goal is:
[0044]
[0045]
[0046] Among them, the update strategy uses the stochastic gradient descent method to update the parameters, the learning rate is set to λ, and the model training is completed when the parameters converge. The trained fault discrimination network model is obtained, and the fault diagnosis results of the target domain signal can be obtained by inputting the target domain sample.
[0047] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The above is only a specific implementation method of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A satellite fault diagnosis method for variable operating conditions based on a distributed domain collaborative game confrontation network, characterized in that: include: S1. When the working condition of the satellite changes, multi-dimensional signals under different working conditions are collected, and a database is constructed according to the collected multi-dimensional signals; S2. Establish a distributed domain collaborative game adversarial network model; S3, training the distributed domain collaborative game adversarial network model through the database constructed in S1 by game adversarial iteration, and using the trained distributed domain collaborative game adversarial network model as a variable operating condition fault diagnosis model; S4, using the variable operating condition fault diagnosis model obtained through training to establish a fault label classifier, and perform fault diagnosis through the fault label classifier; In the process of establishing the distributed domain collaborative game adversarial network model, it includes: forming a fault discrimination network through a feature extractor and a fault label classifier based on a multi-layer deep neural network, wherein the feature extractor and the domain discriminator form a distributed domain discrimination network, and each hidden layer in the feature extractor and the domain discriminator form a collaborative domain discrimination network; In S3, the process of training the distributed domain collaborative game adversarial network model by game adversarial iteration includes: loading the cost functions of the fault label classifier and the domain discriminator, wherein the types of the domain discriminator include a global domain discriminator, a distributed domain discriminator, and a collaborative domain discriminator; The cost function of the collaborative part in the distributed domain collaborative game adversarial network model is: Among them, J task represents the classification part cost function, represents the feature domain classification cost function from layer 1 to r-1, is the classification cost function of the rth layer in the global domain, represents the domain classification cost function corresponding to each type of fault, μ=(μ1,...,μ r ) is the collaborative game weight parameter, i.e., the feature extractor The weight value of the k-th layer feature, represents the domain classification cost function of the rth hidden layer feature of the feature extractor composed of a deep neural network, μ r represents the weight of the feature when the feature extractor passes the feature to the global domain discriminator, represents the network parameters of the first k layers in the feature extractor, are the network parameters of the fault label classifier, represents the network parameters of the global domain discriminator, represents the parameters of the collaborative domain discriminator corresponding to the output features of the k-th layer of the feature extractor, represents the parameters of the corresponding distributed domain discriminator belonging to the jth type of fault, represents the collaborative domain discrimination cost function corresponding to the feature of the k-th layer of the feature extractor, j and k are both positive integers, j is used to represent the j-th type of fault, k is used to represent the number of layers, ω represents the weight parameter between the global domain discriminator and the distributed domain discriminator, μ k is the weight parameter of the k-th layer collaborative game.
2. The method according to claim 1, characterized in that: In S1, the collecting of multi-dimensional signals under different working conditions includes: Collect multi-dimensional signals of satellites before and after working condition changes; The collected multidimensional signal is divided into time segments of the same size, and the obtained time segments are used as samples to construct the database, wherein the source domain samples include: time segments belonging to the source domain multidimensional signal, and the sample category labels added to each time segment are known, and the target domain samples include: time segments belonging to the target domain multidimensional signal, and the sample category labels added to each time segment are empty.
3. The method according to claim 2, characterized in that The method of collecting the satellite's multidimensional signals before and after the working condition changes includes: collecting the satellite's m-dimensional simulation signals before and after the working condition changes, and using the m-dimensional simulation signal before the working condition changes as the source domain multidimensional signal, and using the m-dimensional simulation signal after the working condition changes as the target domain multidimensional signal.
4. The method according to claim 3, characterized in that The step of dividing the collected multi-dimensional signal into time segments of the same size and using the obtained time segments as samples to construct the database comprises: The collected multidimensional signal is divided into time segments of size c, and the constructed database is a sample-label m-dimensional database, where: The source domain samples are x i is the i-th sample, the sample fault category label y i For known labels, It contains L types of fault class labels, i is a positive integer between 1 and n, n is the total number of source domain samples, S D is the source domain; The target domain sample is Sample fault category label Initially empty, N = n + n' represents the total number of samples, n' is the number of samples in the target domain, T D is the target domain, which is updated to the pseudo label obtained by the fault label classifier during training.
5. The method according to claim 1, characterized in that For the fault label classifier G y , the cost function J used for model classification task for: in, Represents the loss function for each sample: make Represents the classification label predicted by the model, i∈(1,n) Used to train the classification fault discrimination network, i∈(n+1,N) The predicted pseudo-labels for the target domain participate in the domain discrimination network training, i represents the i-th sample, n is a positive integer, represents a feature extractor based on an r-layer deep network, x i represents the i-th sample, y i Represents the corresponding fault class label of the i-th sample.