Drag-free control system microthruster fault diagnosis method based on weight embedding and heuristic information

Abstract

The application discloses a kind of based on weight embedding and heuristic information's no-drag control system microthruster fault diagnosis method, first to no-drag and attitude control system dynamics modeling is carried out, and under different fault conditions of microthruster characterization is carried out;Then construct robust residual observer, and collect residual data set from no-drag control system;Then based on the collected multi-channel residual signal, construct the feature processing module based on Pearson coefficient and cross entropy, realize the weight embedding of multi-channel signal;Finally, construct the fusion feature classification model containing heuristic information, the data after using feature processing module processing are used for the training of model, and determine the diagnosis result based on the trained model.The application can highlight fault features in multi-channel signal recognition tasks, and a heuristic structure is designed, which can introduce heuristic information into the network to accelerate convergence and enable the model to focus on important feature information in the early stage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent fault diagnosis, and particularly relates to a fault diagnosis method for micro-thruster of a no-drag control system based on weight embedding and heuristic information. BACKGROUND

[0002] Gravitational waves contain important scientific information such as quantum gravity, space-time structure, origin of matter and origin of the universe. Gravitational wave detection needs an ultra-quiet, extremely low non-gravitational interference environment by capturing the picometer-level changes between free-floating reference masses to infer specific wave source information. The realization of the ultra-quiet, extremely low non-gravitational interference environment needs to use satellite no-drag control technology to shield the multi-source interference noise in the universe. The no-drag control technology controls the satellite body to track the internal mass block through the micro-thruster, so that the mass block always remains in the center of the cavity, to realize the ultra-quiet and ultra-stable environment. The micro-thruster is the core of the no-drag and attitude control system, which generates micro-newton precision thrust to drive the spacecraft to move millimeter level, to achieve nanometer-level control precision. Since the space gravitational wave detection needs to run on-orbit for a long time, the micro-thruster works in the space environment of long-term extreme high and low temperature, strong radiation, etc., and is prone to failure. Considering the high precision requirement brought by the detection task, once the micro-thruster fails, the ultra-stable and ultra-quiet space environment will be difficult to guarantee, which will lead to the failure of the entire detection task. Therefore, the healthy operation of the micro-thruster is crucial to ensure the success of the satellite mission, and it is necessary to ensure its performance and reliability in long-term operation. Obtaining accurate fault information of the thruster is of great significance for the health maintenance and accident prevention of the thruster, and thus helps to ensure the healthy operation of the no-drag and attitude control system. Therefore, it has important practical value to realize accurate thruster fault diagnosis.

[0003] Currently, the number of researches and patents on fault diagnosis is increasing. For example, patent CN118916743A "Equipment fault diagnosis method for self-adaptive updating of model weight" designs a fault diagnosis method that can adaptively update the model weight according to the diagnosis result. Patent CN119004053A "Bearing fault diagnosis method based on generalized time-frequency extraction transformation" designs a time-frequency transformation fault diagnosis method based on the vibration signal of the bearing. In the method for fault diagnosis, although the existing method has made progress in many aspects, few studies have been conducted on complex objects such as micro-thrusters of non-towed spacecraft. For example, the non-towed system has a complex structure and the effect of multiple thrusters is coupled seriously, so it is difficult for the existing method to decouple the fault source. The fault information of the micro-thruster of the non-towed spacecraft is easily submerged by noise, and the signal-to-noise ratio is extremely low; in addition, there are many signal channels carrying the fault information of the micro-thruster, and the requirement for calculation amount is relatively high, and the current research still lacks research on the above problems. At the same time, the existing research can only identify two types of faults, thruster blockage and valve always open, and lacks identification of more common partial failure faults of the thruster, especially the specific failure ratio cannot be determined. Therefore, in view of these difficulties, how to design a reasonable method to realize accurate fault diagnosis of the micro-thruster of the non-towed control system is the core problem to ensure the stable operation of the space gravitational wave detection. SUMMARY

[0004] To solve the above technical problems, the present application provides a non-towed control system micro-thruster fault diagnosis method based on weight embedding and heuristic information, comprising the following steps:

[0005] S1: Build a non-towed and attitude control system dynamics model to characterize different fault conditions of the micro-thruster;

[0006] S2: Construct a robust residual observer to collect residual data sets from the non-towed control system;

[0007] S3: Based on the collected multi-channel residual signals, a feature processing module based on Pearson coefficient and cross-entropy is constructed to realize weight embedding of the multi-channel signals;

[0008] S4: Construct a fusion feature classification model containing heuristic information, use the data processed by the feature processing module to train the model, and determine the diagnosis result based on the trained model.

[0009] The present application has the following effects:

[0010] The application discloses a weight embedding and heuristic information-based fault diagnosis method for microthrusters of a drag-free control system. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 A flowchart of the weight embedding and heuristic information-based fault diagnosis method for microthrusters of a drag-free control system is provided in the application.

[0012] Figure 2 A data sample collected in the embodiment of the application is shown in the figure.

[0013] Figure 3 A fusion feature classification model containing heuristic information in the application is shown in the figure.

[0014] Figure 4 The specific structure of the neural network in the embodiment of the application is shown in the figure, (a) is the structure of a ResNet encoder, and (b) is the structure of a Transformer encoder.

[0015] Figure 5 The training effect under different initial learning rates in the embodiment of the application is shown in the figure.

[0016] Figure 6 The classification results of different neural networks in the embodiment of the application are shown in the figure. DETAILED DESCRIPTION

[0017] In order to make the objectives, technical solutions and advantages of the application clearer, the application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other. In order to achieve the above-mentioned purpose, the application adopts the following technical solutions.

[0018] The application provides a weight embedding and heuristic information-based fault diagnosis method for a micro thruster of a drag-free control system. Figure 1 As shown in the figure, the method comprises the following steps:

[0019] S1: a drag-free and attitude control system dynamics model is built to characterize different fault conditions of the micro thruster;

[0020] S2: a robust residual observer is constructed to collect multi-channel residual signals from the drag-free and attitude control system;

[0021] S3: based on the collected multi-channel residual signals, a feature processing module based on Pearson coefficients and cross-entropy is constructed to realize weight embedding of the multi-channel residual signals;

[0022] S4: a fusion feature classification model containing heuristic information is constructed, the data processed by the feature processing module is used for training of the model, and a diagnosis result is determined based on the trained model.

[0023] The above weight embedding and heuristic information-based fault diagnosis method for the micro thruster of the drag-free control system provided by the application will be described in detail below through a specific embodiment.

[0024] Embodiment 1

[0025] S1: a drag-free and attitude control system dynamics model is built to characterize different fault conditions of the micro thruster, comprising:

[0026] The drag-free and attitude control system dynamics model is built as follows:

[0027] ,

[0028] wherein, , and are control inputs, , and are disturbances, and are stiffness matrices, and are drag-free and suspended state vectors.

[0029] In this embodiment, three common thruster faults are considered, which are thruster blockage (no matter what instruction is received, the thruster does not output any thrust), thruster valve always open (the thruster remains in the state of outputting maximum thrust) and thruster partial failure (the thruster loses part of its effect in proportion). The definitions of different thruster faults are as follows:

[0030] ,

[0031] wherein, , is the input command of the i-th thruster, is the actual output of the i-th thruster, is the maximum thrust that can be output, is the failure ratio of the i-th thruster.

[0032] Step S2: Constructing a robust residual observer, collecting multi-channel residual signals from the drag-free and attitude control system; including:

[0033] Based on the drag-free and attitude control system built in step S1, a linear time-invariant state space model is established as follows:

[0034] ,

[0035] wherein, represents the coefficient matrix of the internal state of the system, represents the control matrix of the input acting on the state, and respectively represent the coefficient matrix of the state and the input acting on the output, and respectively represent the coefficient matrix of the disturbance and the fault acting on the state, and respectively represent the coefficient matrix of the disturbance and the fault acting on the output, is the state vector, is the control input, and respectively represent the disturbance and the fault, represents the measured output, is the control gain matrix.

[0036] Based on the above state space model, a robust residual observer is designed and residual is generated as follows:

[0037] ,

[0038] wherein, is the full-dimensional state estimation value of x, and are the robust residual observer gain matrix and the post-filter matrix, represented as follows:

[0039] ,

[0040] is the stable solution of the Riccati equation and is represented as follows:

[0041] ,

[0042] Then, based on The multi-channel residual signals are collected, as shown in Figure 2 On this basis, residual data sets in different fault states are made and divided into training sets, validation sets and test sets.

[0043] Step S3: Based on the collected multi-channel residual signals, a Weight-Embedding Based on Pearson Correlation Coefficient and Cross-Entropy (WEPC) module is constructed to realize feature selection and weight embedding of the multi-channel residual signals.

[0044] First, taking variables and as examples, the Pearson correlation coefficients between the multi-channel residual signals are calculated:

[0045] ,

[0046] wherein, denotes the Pearson correlation coefficient between and , and are the sample means of and , and are the sample standard deviations.

[0047] Then, a single-channel training cross-entropy calculation strategy is designed: different channels are divided into single channels, which are trained by three-layer perceptrons respectively, and after adjusting to appropriate parameters, a plurality of trained models are obtained; the cross-entropy of the models obtained by training different channels is calculated to predict the probability distribution and the real probability distribution For example, the cross-entropy calculation in the embodiment is as follows:

[0048] ,

[0049] ,

[0050] ,

[0051] wherein, denotes the predicted probability that the sample belongs to the i-th class, is the real probability that the sample belongs to the i-th class, and C denotes the serial number of the correct class.

[0052] Finally, based on the above calculation results, the following multi-channel signal weight distribution mechanism is designed:

[0053] ,

[0054] wherein, is the weight of the i-th channel distribution, represents the cross-entropy obtained by the model trained by the i-th channel signal, represents the sum of the absolute values of the Pearson correlation coefficients between the i-th channel and other channels.

[0055] Based on the above mechanism, each channel signal processed by the feature processing module is assigned a unique weight.

[0056] Step S4: Constructing a fusion feature classification model containing heuristic information (Optimal Features-Guided Fusion Network, OFFN), training the model using the data processed by the feature processing module, and determining the diagnosis result based on the trained model.

[0057] Based on the output data obtained in step S3, a fusion feature classification model containing heuristic information is constructed as shown in Figure 3 . For a given input , it is transmitted to the convolution module and the attention module.

[0058] It should be noted that the two parts are performed simultaneously in the forward propagation of data, and there is no sequence.

[0059] Convolution module:

[0060] First, the input data is subjected to feature extraction in a convolution network with a convolution kernel size of 3, a channel number of 16, and a step of 1, obtaining a 16xHxW-dimensional feature vector; the 16xHxW-dimensional feature vector is input into a ResNet encoder for deep feature extraction, and the structure of the ResNet encoder is shown in Figure 4 (a); the ResNet encoder is composed of eight residual blocks connected in series, and the internal structure of each residual block is in the following order: convolution layer, normalization and ReLU activation function layer, convolution layer, normalization layer, and residual connection and ReLU activation function layer; the convolution kernel size in the ResNet encoder is 3x3, and the channel number is 16, 16, 64, 64, 128, 128, 256, and 256 in the order of residual blocks, and the normalization method is batch normalization; after the above operation, a 256-dimensional feature vector is obtained; Finally, after the global pooling layer and the full connection layer, a high-dimensional feature vector is obtained to represent the output of the convolution module.

[0061] Attention module:

[0062] First, the input data is split into H vectors of length W; then, a mechanism is designed to introduce heuristic information into the neural network:

[0063] The sequence with the highest weight obtained in step S3 is used to construct a heuristic structure. The parameters in the vector are fed into the optimizer to become learnable variables, and then appended to the front of the above H vectors of length W.

[0064] The feature vectors with added heuristic structure are positionally encoded and then input into a Transformer encoder, the structure of which is as follows: Figure 4 As shown in (b) above, the Transformer encoder consists of two main sequentially connected blocks. The first block, in sequence, comprises a multi-head attention layer, a residual connection and normalization layer, a feedforward network layer, and another residual connection and normalization layer. The second block, in sequence, comprises a multi-head attention layer, a residual connection and normalization layer, a feedforward network layer, another residual connection and normalization layer, and a multilayer perceptron layer. Batch normalization is used for all normalization methods in the Transformer encoder. The attention scoring function is calculated as follows:

[0065] ,

[0066] Where d is the sequence length, Indicates a query. Indicates key, Indicates the value.

[0067] After the Transformer encoder described above, a heuristic structure is selected. Corresponding output sequence This represents the output of the attention module.

[0068] The feature vectors output by the two modules mentioned above and The features are added together to obtain the final fused feature sequence. Finally, the sequence is passed through a two-layer fully connected layer and a Softmax function to generate the final probability sequence. The category with the highest probability is taken as the classification result.

[0069] The advantages of the present invention over the prior art will be explained in detail below with reference to this embodiment.

[0070] The embodiment adopts the data set collected by the above-mentioned S2, and each data set has 8 categories of health states. The specific label representing each category is shown in Table 1, wherein failure (x-y) represents the failure rate of the thruster in the interval (x-y).

[0071] Table 1

[0072]

[0073] The data set is divided into a training set, a validation set and a test set in a ratio of 7:1:2. Based on the mechanism of the third step, the data is input into the WEPC module. After the data processing work is completed, the parameter setting of the training is performed. The performance of the OFFN is significantly affected by the initial learning rate (ILR). The ILR is selected on the training set and the validation set, the cross-entropy is used to measure the classification effect, and Adam is used as the optimizer. The training effects under different ILRs are shown in Figure 5 The results show that the OFFN achieves the best effect when the ILR is set to 0.00002.

[0074] The details of the comparative algorithm set in the embodiment are as follows, and the results are shown in Figure 6

[0075] (1) ResNet18: a convolutional neural network containing a residual structure.

[0076] (2) VGG: a deep convolutional neural network proposed by the University of Oxford.

[0077] (3) GoogLeNet: a neural network architecture containing a multi-parallel path convolution structure.

[0078] (4) TE: an encoder containing an attention mechanism.

[0079] (5) OFFN: the classification model disclosed in the present application, which is trained using data that has not been processed by the WEPC module.

[0080] (6) WEPC+OFFN: the classification model disclosed in the present application, which is trained using data processed by the WEPC module.

[0081] The model of the present application achieves an identification accuracy of 99.1% on the collected data set, which shows that the proposed model can accurately diagnose under the constraints of weak signal and strong noise interference, and the diagnosis result meets the needs of actual engineering application.

[0082] ​The application discloses a kind of based on weight embedding and heuristic information's no-drag control system microthruster fault diagnosis method, first to no-drag and attitude control system dynamics modeling is carried out, and under different fault conditions of microthruster characterization is carried out;Then construct robust residual observer, and collect residual data set from no-drag control system;Then based on the collected multi-channel residual signal, construct the feature processing module based on Pearson coefficient and cross entropy, realize the weight embedding of multi-channel signal;Finally, construct the fusion feature classification model containing heuristic information, utilize the data handled after feature processing module to train model, and determine diagnosis result based on the model trained.The above-mentioned method utilizes the residual of no-drag and attitude control system, not original signal (such as attitude angle) to construct data set, to highlight thruster fault feature more effectively;A kind of feature processing module capable of realizing weight embedding is constructed, can complete efficient feature engineering in multi-channel signal identification task;In addition, a heuristic structure is designed, can introduce heuristic information into network, to accelerate convergence and enable model to focus on important feature information in early stage.

[0083] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for fault diagnosis of a micro-thrust in a drag-free control system based on weighted embedding and heuristic information, characterized in that, Includes the following steps: S1: Build a dynamic model of the drag-free and attitude control system and characterize the micro-thruster under different fault conditions; S2: Construct a robust residual observer to collect multi-channel residual signals from the drag-free and attitude control system; S3: Based on the collected multi-channel residual signals, a feature processing module based on Pearson coefficient and cross-entropy is constructed to realize the weight embedding of multi-channel residual signals; S4: Construct a fusion feature classification model containing heuristic information, train the model using the data processed by the feature processing module, and determine the diagnostic results based on the trained model.

2. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 1, characterized in that, Step S1 includes building a dynamic model of the drag-free and attitude control system; considering different thruster faults, and injecting micro-thruster faults into the above dynamic model to complete fault modeling.

3. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 1, characterized in that, Step S2 includes designing a robust residual observer and generating residuals, acquiring multi-channel residual signals; based on this, creating residual datasets under different fault states, and dividing them into training sets, validation sets, and test sets.

4. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 1, characterized in that, Step S3 includes first calculating the Pearson correlation coefficient between the multi-channel signals; Then, a strategy for calculating cross-entropy during single-channel training is designed: different channels are divided into single channels and trained separately using a three-layer perceptron, and multiple trained models are obtained after adjusting appropriate parameters; cross-entropy is calculated for the models trained on different channels; finally, based on the above calculation results, the following multi-channel signal weight allocation mechanism is designed: ， in, It is the weight assigned to the i-th channel. Let represent the cross-entropy obtained by the model trained from the signal of the i-th channel. This represents the sum of the absolute values ​​of the Pearson correlation coefficients between the i-th channel and the other channels; based on the above mechanism, each channel signal after processing by the feature processing module is assigned a unique weight.

5. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 4, characterized in that, Step S4 includes: constructing a fusion feature classification model containing heuristic information based on the output data obtained from the feature processing module; for a given input The data is transmitted to the convolution module and the attention module simultaneously during the forward propagation of the data, without any order of transmission.

6. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 4, characterized in that, The convolution module is specifically as follows: First, the input data undergoes feature extraction in a convolutional network with a kernel size of 3, 16 channels, and a stride of 1, yielding a 16×H×W dimensional feature vector. This 16×H×W dimensional feature vector is then input into a residual encoder for deep feature extraction. After these operations, the desired feature vector is obtained. The feature vectors are first processed by a global pooling layer and then by a fully connected layer to obtain a high-dimensional feature vector representing the output of the convolutional module. .

7. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 6, characterized in that, The attention module specifically consists of: First, the input data is split into H vectors of length W. Then, a mechanism is designed to introduce heuristic information into the neural network: the sequence with the largest weight obtained in step S3 is used to create a heuristic structure. The parameters in the vectors are fed into the optimizer to become learnable variables, and then concatenated to the front of the H vectors of length W mentioned above. The feature vectors with the added heuristic structure are positionally encoded and input into the Transformer encoder. After passing through the Transformer encoder, the heuristic structure is selected. Corresponding output sequence This represents the output of the attention module.

8. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 7, characterized in that, eigenvectors and The features are added together to obtain the final fused feature sequence. Finally, the sequence is passed through a two-layer fully connected layer and a Softmax function to generate the final probability sequence. The category with the highest probability is taken as the classification result.

9. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 8, characterized in that, The residual encoder consists of eight residual blocks connected in series. The internal structure of each residual block, in sequence, is a convolutional layer, a normalization and ReLU activation function layer, a convolutional layer, a normalization layer, and a residual connection and ReLU activation function layer. The convolutional kernel size in the residual encoder is 3×3. The number of channels, in the order of the residual blocks, is 16, 16, 64, 64, 128, 128, 256, and 256, respectively. Batch normalization is selected for all of them.

10. The method for fault diagnosis of a drag-free control system micro-thruster based on weighted embedding and heuristic information as described in claim 9, characterized in that, The Transformer encoder consists of two main sequentially connected blocks. The first block, in sequence, comprises a multi-head attention layer, a residual connection and normalization layer, a feedforward network layer, and another residual connection and normalization layer. The second block, in sequence, comprises another multi-head attention layer, a residual connection and normalization layer, a feedforward network layer, another residual connection and normalization layer, and a multilayer perceptron layer. Batch normalization is used throughout the Transformer encoder. The attention scoring function is calculated as follows: ， Where d is the sequence length, Indicates a query. Indicates key, Indicates the value.

Images