Satellite power system fault diagnosis method based on multi-strategy fusion deep learning

Abstract

The application discloses a satellite power supply system fault diagnosis method based on multi-strategy fusion deep learning, and belongs to the technical field of fault diagnosis. The application solves the defects of poor fault diagnosis model precision and low stability of the existing satellite power supply system caused by complex working conditions. Based on the DET&S4R satellite power supply system topology, the application uses Dymola&Modelica software for modeling and simulation, sets data sampling points as output and input ports of each device, injects faults of each level into the model, records sampling voltage and current during fault simulation, and constructs a fault sample data set. The application uses a convolutional neural network to extract sample features, constructs a low-dimensional sample data set, and reduces the calculation complexity. The application proposes a self-adaptive boosting algorithm combined with enhanced multi-head attention to improve a bidirectional gated recurrent unit, and uses a white shark optimization algorithm WSO to solve hyperparameters and network parameters of the BIGRU, so as to improve the precision and training efficiency of the diagnosis model, and establish a strong learner multi-classification fault diagnosis model based on multi-strategy fusion deep learning. Finally, fault diagnosis is performed, and diagnosis results are output, including fault diagnosis precision, diagnosis time, fault position, fault type and the like. The application is suitable for satellite power supply system fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis technology, specifically relating to a fault diagnosis method for satellite power systems based on deep learning. Background Technology

[0002] Earth-based satellites play a vital role in military, communications, weather, and exploration fields. However, the variable space environment and complex satellite structures pose significant challenges to their safe and stable operation. As a critical system, the satellite's power system, with its exposed photovoltaic arrays in space, is highly susceptible to failure. Furthermore, the internal lithium-ion battery packs and secondary power supplies are also prone to sudden malfunctions. Failure to respond promptly to these sudden failures could lead to catastrophic consequences for the satellite.

[0003] Researchers typically employ data-driven machine learning or deep learning methods to study fault diagnosis in satellite power systems. This involves modeling and simulating data in Simulink based on satellite telemetry data, and then performing fault diagnosis based on this data. However, current fault diagnosis models are not well-developed. Some scholars have used traditional deep learning models for fault diagnosis, but these models rely heavily on the values ​​of network parameters, and their generalization and stability lack validation. In recent years, graph neural network-based fault diagnosis using image data has become more prevalent, but these models are complex, slow, and difficult to implement online.

[0004] Satellite power systems are complex in structure, with highly coupled and dimensional fault features, significantly impacting the accuracy and efficiency of diagnostics. Simultaneously, the performance of the diagnostic algorithm affects the accuracy and efficiency of online fault diagnosis. This case study investigates the correlation of features in satellite power systems, focusing on feature analysis to reduce data complexity and improve fault diagnosis efficiency. Furthermore, using the BIGRU model as the diagnostic base algorithm, hyperparameter optimization using an improved metaheuristic algorithm is conducted, along with a strong learning classifier based on Adaboost, to enhance the stability of the diagnostic model and the accuracy of fault diagnosis.

[0005] To effectively extract data features, reducing feature dimensionality can decrease diagnostic complexity and avoid model overfitting. Therefore, we propose using convolutional neural networks (CNNs) for feature scaling. After extracting features through convolutional layers, the features are input into pooling layers for dimensionality reduction, thereby reducing the impact of high-dimensional features on the model and improving diagnostic performance.

[0006] Bidirectional Gated Recurrent Unit (BIGRU) is a deep learning model improved upon Recurrent Neural Networks (RNNs). It consists of two integrated GRUs and effectively mitigates gradient vanishing, exhibiting strong generalization and low computational complexity, but it suffers from low sensitivity to key features. Therefore, an Enhanced Multi-Head Attention (EMHA) mechanism is proposed to improve the BIGRU model, enhancing the traditional EMHA mechanism based on layer normalization and residual mechanisms. Adding an EMHA layer before the output layer simplifies the feedforward layer and mask matrix while effectively improving the model's sensitivity to key features, thereby enhancing its diagnostic performance. Summary of the Invention

[0007] To address the limitation of BIGRU model performance being highly dependent on parameters, researchers often employ optimization strategies to optimize network parameters and hyperparameters. However, many algorithms suffer from the curse of dimensionality, which negatively impacts optimization performance. This invention proposes a multi-strategy improved White Shark Optimizer (WSO) algorithm to optimize BIGRU (Bidirectional Recurrent Neural Network) parameters. The hybrid improvement methods include adaptive mutation perturbation and elite learning strategies, effectively enhancing the algorithm's optimization accuracy and convergence speed, and enabling rapid and accurate solution of model parameters. Furthermore, to address the weak stability of BIGRU, Adaboost (Adaptive Boosting) is proposed to improve BIGRU by integrating multiple high-weight weak classifiers to construct a strong classifier, thereby increasing model diagnostic stability.

[0008] This invention provides a satellite power system fault diagnosis method based on multi-strategy fusion deep learning. To achieve the above-mentioned objective, this invention provides a satellite power system fault diagnosis method based on multi-strategy fusion deep learning. This method uses CNN (Convolutional Neural Networks) to reduce the fault features of the satellite power system, and proposes to use EMHA (Enhance multi-head attention) to enhance the sensitivity of key features of BIGRU. WSO is used to solve the high-dimensional parameters of the model, and Adaboost is combined to build a strong classifier, thereby improving the diagnostic accuracy and efficiency of the model.

[0009] A fault diagnosis method for satellite power systems based on multi-strategy fusion deep learning, characterized by comprising:

[0010] A fault diagnosis method for satellite power systems based on multi-strategy fusion deep learning, characterized by comprising:

[0011] Step 1: Based on the satellite power system topology of DET & S4R, use Dymola & Modelica software to perform normal power system modeling and fault modeling.

[0012] Step 2: Set the data sampling points to the output and input ports of each device, inject faults at all levels into the model, record the sampled voltage and current during fault simulation, and construct a fault sample dataset;

[0013] Step 3: Use CNN to extract sample features, construct a fused sample dataset, and set up the training set and validation set;

[0014] Step 4: Construct a multi-classification model using a deep learning method based on EMHA and BIGRU, and optimize the hyperparameters and network parameters in the BIGRU model using a multi-strategy improved WSO.

[0015] Step 5: Use Adaboost to fuse IWSO-BIGRU-EMHA and train the model based on the training set data to build a strong classifier based on a multi-strategy fusion deep learning model.

[0016] Step 6: Use the multi-strategy fusion deep learning model to perform fault diagnosis on the test set and obtain fault diagnosis results, including diagnosis accuracy, diagnosis time, fault location, and fault type.

[0017] Beneficial effects of this invention:

[0018] (1) The satellite power system fault diagnosis method based on multi-strategy fusion deep learning of the present invention uses convolutional neural network (CNN) for feature scaling, which reduces the impact of strong coupling and high-dimensional features on the model and improves the diagnostic performance.

[0019] (2) The satellite power system fault diagnosis method based on multi-strategy fusion deep learning proposed in this invention proposes an enhanced multi-head attention mechanism (EMHA) to improve the BIGRU model. Based on layer normalization and residual mechanism, the traditional multi-head attention mechanism is enhanced, which can effectively improve the model’s sensitivity to key features, thereby increasing the model’s diagnostic performance.

[0020] (3) The satellite power system fault diagnosis method based on multi-strategy fusion deep learning proposed in this invention proposes WSO to optimize BIGRU parameters. The hybrid improvement method includes adaptive mutation perturbation strategy and elite learning strategy, which can effectively improve the optimization accuracy and convergence speed of the algorithm and quickly and accurately solve the model parameters.

[0021] (4) The satellite power system fault diagnosis method based on multi-strategy fusion deep learning of the present invention addresses the weakness of BIGRU stability by proposing Adaboost to improve BIGRU, integrating multiple high-weight weak classifiers to construct a strong classifier, which greatly increases the stability of model diagnosis. Attached Figure Description

[0022] Figure 1 This is a topology diagram of the satellite power system;

[0023] Figure 2 The flowchart of the method described in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0027] This invention combines Figure 1 Topology and Figure 2 This embodiment describes a method for fault diagnosis of satellite power systems based on multi-strategy fusion deep learning, which specifically includes the following steps:

[0028] Step 1: Based on the satellite power system topology of DET & S4R, use Dymola & Modelica software to perform normal power system modeling and fault modeling.

[0029] Step 2: Set the data sampling points to the output and input ports of each device, inject faults at various levels into the model, record the sampled voltage and current during fault simulation, and construct a fault sample dataset. The fault parameters are shown in the table below:

[0030] By injecting faults into the simulation model, a total of 20 faults were obtained, covering four working conditions: illumination period-1, shadow period-2, ground shadow period-3, and shadow period-4. Each fault had 10,000 fault samples, for a total of 200,000 fault samples. The fault types and corresponding fault codes are shown in the table below:

[0031] Step 3: Use CNN to extract features from the fault samples obtained in Step 2, construct a fused sample dataset, and set the training set and validation set in a 7:3 ratio;

[0032] Step 4: Construct a multi-classification model using a deep learning method based on EMHA and BIGRU, and optimize the hyperparameters and network parameters in the BIGRU model using a multi-strategy improved WSO.

[0033]

[0034] Step 5: Use Adaboost to fuse IWSO-BIGRU-EMHA and train the model based on the training set data to build a strong classifier based on a multi-strategy fusion deep learning model.

[0035] Step 6: Use a multi-strategy fusion deep learning model to perform fault diagnosis on the test set and obtain fault diagnosis results, including diagnosis accuracy, diagnosis time, fault location, and fault type.

[0036]

[0037] The invention is further characterized in that,

[0038] Based on sample features, the main steps of CNN for sample feature extraction are as follows:

[0039] Step 31: Input the sample data into the input layer;

[0040] Step 32: Extract features from the sample data using convolutional layers;

[0041] Step 33: Perform secondary feature extraction on the sample data using a pooling layer;

[0042] Step 34: Combine the data extracted from the convolutional layer with the data extracted from the pooling layer based on the fully connected layer;

[0043] Furthermore,

[0044] In step 32, the convolution formula is:

[0045]

[0046] Where m is the feature map, k is the convolution kernel sliding row, s is the activation function, l is the convolution kernel sliding column, w is the weight coefficient, and b is the offset coefficient;

[0047] In step 33, the formula for the max pooling process is:

[0048] s i =max{m1, m2, ..., m3}

[0049] In step 34, the formula for the fully connected layer is:

[0050] y = σ(ws + b)

[0051] Step 4, using WSO to optimize the hyperparameters and network parameters in the BIGRU model, includes the following steps:

[0052] Step 41: Initialize the White Shark optimization algorithm, which mainly includes initializing the maximum number of iterations T, dimension D, and population size N;

[0053] Step 42: Calculate the fitness value of the great white shark population and identify the optimal great white shark individual;

[0054] Step 43, White Shark Population Regeneration Speed;

[0055] Step 44: White shark population update location;

[0056] Step 45: The great white shark population performs WSO (Wide Schooling) behavior and updates the position of the best great white shark individual.

[0057] Step 46: Determine if the current iteration number is greater than T. If yes, continue to the next step; otherwise, go to step 32.

[0058] Step 47: Decompose the optimal white shark individual position obtained in Step 45 into dimensions and assign values ​​to the BIGRU parameters to be optimized.

[0059] Furthermore,

[0060] In step 43, the formula for updating the population speed of the great white sharks is:

[0061]

[0062] Let be the velocity of the i-th white shark at iteration number t+1. Let w be the velocity of the i-th white shark at iteration number t. gbestt This represents the optimal individual position of the white shark population at iteration number t. Let c1 and c2 be the optimal position corresponding to the velocity of the great white shark population at iteration number t, where c1 and c2 are random numbers between [0, 1], m is the contraction factor, and p1 and p2 are control coefficients. The formula is as follows:

[0063]

[0064] T is the maximum number of iterations;

[0065] In step 44, the formula for the great white shark swarm to surround the optimal prey is:

[0066]

[0067] Let u and l be the position of the i-th white shark at iteration number t+1, and let a and b be vectors, respectively, representing the upper and lower bounds of the search space. The formula is:

[0068]

[0069] In step 45, the formula for the schooling behavior of the great white shark population is:

[0070]

[0071] The fault coding method used in step 4 above is the multi-strategy improved White Shark optimization algorithm.

[0072] Step 5 uses Adaboost to fuse WSO-BIGRU-EMHA and trains the model based on the training set data to build a strong classifier based on a multi-policy fusion deep learning model. The specific process of fusing the strong classifier based on Adaboost is as follows:

[0073] First, the sample set weights are initialized, and a classifier is obtained based on the training results of WSO-BIGRU-EMHA.

[0074] Then, based on the classification results, increase the weight of misclassified data and decrease the weight of correctly classified data;

[0075] Secondly, train multiple classifiers, increase the weight of classifiers with high classification accuracy, and decrease the weight of classifiers with low classification accuracy.

[0076] Finally, after retraining the classifier until the accuracy metric is met, all classifiers are weighted and voted on.

[0077] In step 4, a multi-classification model is constructed using a deep learning method based on EMHA-improved BIGRU. The EMHA formula is:

[0078] y t =A+h t

[0079] E = Normalization(y t )

[0080] Where A represents the traditional attention mechanism, and h t The output of the classifier, i.e., the input of EMHA, is y. t This is the output of EMHA, where E represents the normalized output.

[0081] In step 41, the maximum number of iterations T ranges from 50 to 100, and the population size N ranges from 30 to 50.

[0082] In step 45, the WSO fish swarm behavior is enhanced with an adaptive mutation perturbation strategy:

[0083]

[0084] Where r and rand are random numbers between (0, 1). The maximum value of the population at the t-th iteration. Let be the minimum population value at the t-th iteration, p be the mutation probability, p = r(1-t / T), t be the current iteration number, and V be the mutation coefficient;

[0085] In step 4, the multi-strategy improved WSO update strategy adds an elite learning strategy:

[0086]

[0087] N is the population size of great white sharks, index(x) represents the value selected in sorting order x, and randn(3) represents a random selection from 1, 2, and 3. Let f be the position of the Nth white shark individual in the (t+1)th iteration. E The elitism factor is calculated using the following formula:

[0088]

[0089] By implementing the above diagnostic methods, fault diagnosis was performed on the satellite power system. Table 1 shows the diagnostic results obtained from the satellite power system fault diagnosis model using a multi-strategy fusion deep learning approach. The fault diagnosis results of Stacked Sparse Autoencoder (SSAE), Long Short-Term Memory (LSTM), and Random Forest (RF) are also compared. Analysis of Table 1 shows that the fault diagnosis evaluation metrics based on the implemented methods are the highest, at 98.6%, 99.1%, 99.8%, and 99.4%, respectively. This comparison demonstrates that the improved methods effectively enhance the model's diagnostic accuracy and exhibit high diagnostic performance for high-dimensional faults in satellite power systems.

[0090] Table 1 Comparison of Diagnostic Results

[0091] Diagnostic model Diagnostic accuracy F1 score Recall rate Specificity SSAE 92.2％ 92.5％ 98.4％ 92％ LSTM 91.1％ 91.3％ 98.2％ 91.7％ RF 88.9％ 91.8％ 98.2％ 91.1％ Implementation Methods 98.6％ 99.1％ 99.8％ 99.4％

Claims

1. A fault diagnosis method for satellite power systems based on multi-strategy fusion deep learning, characterized in that, include: Step 1: Based on the satellite power system topology of DET&S4R, use Dymola & Modelica software to perform normal power system modeling and fault modeling. Step 2: Set the data sampling points to the output and input ports of each device, inject faults at all levels into the model, record the sampled voltage and current during fault simulation, and construct a fault sample dataset; Step 3: Use CNN to extract sample features, construct a fused sample dataset, and set up the training set and validation set; Step 4: Construct a multi-classification model using a deep learning method based on EMHA and BIGRU fusion. The formula is as follows: ； Where A represents the traditional attention mechanism, and h t The output of the classifier, i.e., the input of EMHA, is y. t This is the output of EMHA, where E represents the normalized output; The hyperparameters and network parameters in the BIGRU model were optimized using a multi-strategy improved WSO, including: Step 41: Initialize the White Shark Optimization Algorithm, including initializing the maximum number of iterations T, dimension D, and population size N; Step 42: Calculate the fitness value of the great white shark population and identify the optimal great white shark individual; Step 43, White Shark Population Regeneration Speed; Step 44: White shark population update location; Step 45: The great white shark population performs WSO (Wide Schooling) behavior and updates the position of the best great white shark individual. Step 46: Determine if the current iteration number is greater than T. If yes, continue to the next step; otherwise, go to step 32. Step 47: Decompose the optimal white shark individual position obtained in Step 45 into dimensions and assign values ​​to the BIGRU parameters to be optimized. Step 5: Use Adaboost to fuse IWSO-BIGRU-EMHA and train the model based on the training set data to build a strong classifier based on a multi-strategy fusion deep learning model. Step 6: Use the multi-strategy fusion deep learning model to perform fault diagnosis on the test set and obtain fault diagnosis results, including diagnosis accuracy, diagnosis time, fault location, and fault type.

2. The satellite power system fault diagnosis method based on multi-strategy fusion deep learning according to claim 1, characterized in that, Step 3 uses CNN to extract sample features, including: Step 31: Input the sample data into the input layer; Step 32: Extract features from the sample data using convolutional layers; Step 33: Perform secondary feature extraction on the sample data using a pooling layer; Step 34: Combine the data extracted from the convolutional layer with the data extracted from the pooling layer based on the fully connected layer.

3. The satellite power system fault diagnosis method based on multi-strategy fusion deep learning according to claim 1, characterized in that, The maximum number of iterations T mentioned in step 41 ranges from 50 to 1s00, and the population size N ranges from 30 to 50.

4. The satellite power system fault diagnosis method based on multi-strategy fusion deep learning according to claim 1, characterized in that, Step 45 describes an adaptive mutation perturbation strategy for WSO fish swarm behavior: ； Where w is the position of the individual great white shark, and r and rand are random numbers between (0, 1). The maximum value of the population at the t-th iteration. Let be the minimum population value at the t-th iteration, p be the mutation probability, p = r(1-t / T), t be the current iteration number, and V be the coefficient of mutation. Let be the position of the i-th white shark in the (t+1)-th iteration. Let be the position of the i-th white shark individual at the t-th iteration.

5. The satellite power system fault diagnosis method based on multi-strategy fusion deep learning according to claim 1, characterized in that, The improved WSO with multiple strategies adds an elite learning strategy to the updated strategy: ； N is the population size of the great white sharks, index(x) represents selecting the value with sorting order x, and randn(3) represents randomly selecting from 1, 2, and 3. Let f be the position of the Nth white shark individual in the (t+1)th iteration. E The elitism factor is calculated using the following formula: 。 6. The satellite power system fault diagnosis method based on multi-strategy fusion deep learning according to claim 1, characterized in that, Step 5 uses Adaboost to fuse WSO-BIGRU-EMHA and trains the model based on the training set data to build a strong classifier based on a multi-policy fusion deep learning model, including: Initialize the sample set weights and obtain a classifier based on the WSO-BIGRU-EMHA training results; then increase the weight of misclassified data and decrease the weight of correctly classified data based on the classification results. Train multiple classifiers, increase the weight of classifiers with high classification accuracy, and decrease the weight of classifiers with low classification accuracy; retrain the classifiers until the accuracy metric is met, and then vote on all classifiers using a weighted average.

Images