Satellite power supply system fault prediction method based on multi-model combination

Through the multi-model combination method, the convolutional neural network and the long-term and short-term memory network optimized by the Sand Cat Group, combined with grayscale map degradation and rolling prediction, the problem of insufficient overall state reflection of the satellite power system is solved, and the failure mode of the satellite power system is realized and accurate prediction of lifespan improves prediction accuracy and generalization ability.

CN119089270APending Publication Date: 2024-12-06SUN YAT SEN UNIVERSITY SHENZHEN +1

Patent Information

Application Number
CN202411197429.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing technology is difficult to fully reflect the overall status of satellite power systems. A single model cannot accurately capture the performance degradation of solar panels or batteries, resulting in inaccurate satellite failure prediction.

Method used

Using a multi-model combination method, a convolutional neural network and a long and short-term memory network based on Sand Cat Group optimization is used to construct a prediction model and comprehensively analyze different types of data of the power system to improve the Accuracy of fault prediction.

Benefits of technology

It realizes more accurate failure mode and residual life prediction of satellite power supply systems, improves prediction accuracy and generalization capabilities, and enhances the ability to extract deep battery failure characteristics.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089270A_ABST
    Figure CN119089270A_ABST
Patent Text Reader

Abstract

The invention discloses a satellite power supply system fault prediction method based on multi-model combination, and the method comprises the steps: collecting the data of a storage battery of a satellite power supply system, and obtaining a first data set and a second data set; performing grey-scale map degradation on the first data set to obtain first input data; performing data conversion on the second data set based on a rolling prediction method to obtain second input data; constructing a prediction model based on a convolutional neural network and a long-short-term memory network based on the salat swarm optimization; training a prediction model based on the first input data and the second input data to obtain a trained prediction model; and predicting the state of the storage battery in a period of time in the future based on the trained prediction model to obtain a prediction result. According to the invention, different types of data of the power supply system can be comprehensively analyzed, and the potential fault mode and the residual life can be predicted more accurately. The method can be widely applied to the technical field of satellite fault prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of satellite fault prediction, and in particular to a satellite power system fault prediction method based on multi-model combination. Background Art

[0002] With the development of aerospace technology, satellites are key tools for modern communications, navigation and earth observation. Their stability and reliability are crucial to ensuring the smooth operation of the global information network. The satellite power supply system is responsible for providing a stable and reliable power supply to support the normal operation of various payloads and equipment on the satellite. The satellite power supply system mainly includes solar panels, batteries, power controllers and other parts. These components operate in a long-term space environment and are affected by various adverse factors such as cosmic rays and extreme temperature differences. They often face the risk of failure. Any failure in the power supply system may cause partial or complete loss of satellite functions.

[0003] Fault prediction technology can be used to predict the system status and remaining life in advance, enhance the stability and safety of satellites in orbit, and ensure mission continuity. Batteries will age during repeated charging and discharging, which is manifested as decreased capacity and increased internal resistance. Predicting the capacity change trend and failure threshold of batteries can help to understand the performance and life of batteries in a timely manner.

[0004] In recent years, many scholars have used data-driven methods to predict faults in different power systems. For example, a fault prediction method based on artificial fish swarm algorithm to optimize gray neural network determines whether the system has a fault by predicting the output voltage signal of a certain power combination. Simulation results show that this method works well.

[0005] However, the data source under a single model usually cannot fully reflect the overall status of the satellite power system. If only relying on voltage data, it is impossible to accurately capture the performance degradation of solar panels or batteries. Summary of the invention

[0006] In order to solve the above technical problems, the purpose of the present invention is to provide a satellite power system fault prediction method based on multi-model combination, which can comprehensively analyze different types of data of the power system and more accurately predict potential failure modes and remaining life.

[0007] The technical solution adopted by the present invention is: a satellite power system fault prediction method based on multi-model combination, comprising the following steps:

[0008] Collecting satellite power system battery data to obtain a first data set and a second data set;

[0009] Performing grayscale image degradation on the first data set to obtain first input data;

[0010] Performing data conversion on the second data set based on a rolling prediction method to obtain second input data;

[0011] Constructing prediction models based on convolutional neural networks and long short-term memory networks optimized by sand cat colonies;

[0012] Training the prediction model based on the first input data and the second input data to obtain a trained prediction model;

[0013] Based on the trained prediction model, the battery status in the future is predicted to obtain the prediction result.

[0014] Furthermore, the first data set includes charging voltage data, charging current data and charging temperature data of the lithium battery under each charge and discharge cycle; and the second data set includes historical data of battery capacity.

[0015] Furthermore, the first input data is expressed as follows:

[0016]

[0017] in, Represents the grayscale converted value of the i-th battery degradation data under the j-th cycle; Represents the battery degradation data from the jth cycle to the j+4th cycle; and Respectively represent the minimum and maximum values ​​of battery degradation data under the jth cycle; represents the degradation data of the i-th battery under the j-th cycle; represents the degradation data of the ith battery in the j+4th cycle; round(·) means rounding the result to the nearest integer; t represents the newly created time point; t i and t i+1 represents the two nearest neighboring points of the newly created time point in the original sequence; D(t) represents the interpolation value corresponding to the newly created time point; D(t i ) and D(t i+1 ) represent the voltage, current or temperature value of the corresponding nearest neighbor point respectively.

[0018] Furthermore, the second input data is expressed as follows:

[0019]

[0020] Among them, X represents the LSTM input layer data; Y represents the label data; C L Indicates the battery capacity after the Lth cycle.

[0021] Furthermore, the prediction model is obtained by connecting a convolutional neural network and a long short-term memory network optimized based on a sand cat colony in parallel, wherein:

[0022] The convolutional neural network comprises a convolutional layer, a pooling layer and a fully connected layer;

[0023] The long short-term memory network based on sand cat swarm optimization includes a sand cat swarm optimization module and a long short-term memory network;

[0024] The long short-term memory network includes a forget gate, an input gate and an output gate.

[0025] Further, the sand cat group optimization module is used to optimize the hyperparameters of the long short-term memory network, including the following steps:

[0026] Randomly generate a certain number of sand cat individuals and initialize the sand cat population;

[0027] Construct a fitness function and calculate the fitness value of the current position of the sand cat individual;

[0028] The position of the individual sand cat is updated based on the updating strategy in the searching prey phase and the updating strategy in the attacking prey phase;

[0029] The fitness value calculation and the position update steps of the individual sand cats are cyclically performed until a maximum number of iterations is met, and the optimal sand cat position is determined based on the fitness value as a hyperparameter of the long short-term memory network.

[0030] Furthermore, the step of training the prediction model based on the first input data and the second input data to obtain a trained prediction model specifically includes:

[0031] Inputting the first input data and the second input data into the prediction model to obtain a first prediction result;

[0032] Calculate the value of the model loss function based on the first prediction result and the actual observation data;

[0033] The parameters of the prediction model are adjusted based on the value of the model loss function and the fitness value to obtain a trained prediction model.

[0034] The beneficial effects of the method of the present invention are as follows: the present invention converts the one-dimensional degradation data of the battery into a grayscale image and inputs it into a convolutional neural network, which not only increases the diversity of the data, but also effectively utilizes the charging and discharging data of the battery of the satellite power system, so that the model can extract the deep fault characteristics of the battery, thereby improving the prediction accuracy; by introducing the sand cat swarm optimization algorithm to optimize the hyperparameters of the long short-term memory network, the convergence process is accelerated, and the prediction performance and generalization ability of the network are improved; and finally, it realizes the comprehensive analysis of different types of data of the power system, thereby more accurately predicting the potential failure mode and remaining life. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the steps of a satellite power system fault prediction method based on multi-model combination of the present invention;

[0036] Figure 2 It is a flowchart of a satellite power system fault prediction method based on multi-model combination of the present invention;

[0037] Figure 3 It is a convolutional neural network structure block diagram of a satellite power system fault prediction method based on multi-model combination of the present invention;

[0038] Figure 4 It is a long short-term memory network structure block diagram of a satellite power system fault prediction method based on multi-model combination of the present invention;

[0039] Figure 5 It is a schematic flow chart of a sand cat group optimization module of a satellite power system fault prediction method based on a multi-model combination of the present invention;

[0040] Figure 6 It is a schematic diagram of a battery degradation grayscale image generation process of a satellite power system fault prediction method based on a multi-model combination of the present invention;

[0041] Figure 7 It is a comparison chart of prediction results of a satellite power system fault prediction method based on a multi-model combination of the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0043] Reference Figure 1 and Figure 2 The present invention provides a satellite power system fault prediction method based on multi-model combination, the method comprising the following steps:

[0044] S1, collecting satellite power system battery data to obtain a first data set and a second data set;

[0045] Specifically, 90% of the spacecraft power systems in the world adopt the solar array-battery group structure. As a key component, the health status of the battery group has a direct impact on the working efficiency of the space power system. In recent years, the technology of lithium-ion batteries has developed rapidly, gradually replacing traditional silver-zinc batteries and nickel-metal hydride batteries, and has begun to be widely used in satellite power systems. As the most vulnerable component in the solar array-battery group space power system, the battery group collects the charging voltage data, charging current data and charging temperature data of the lithium battery under each charge and discharge cycle of the lithium battery as the first data set; collects the historical data of the battery capacity as the second data set; the first data set is used as the input data of the convolutional neural network; the second data set is used as the input data of the long short-term memory network based on the sand cat group optimization.

[0046] S2. Perform grayscale image degradation on the first data set to obtain first input data;

[0047] Specifically, refer to Figure 6 , since the length of historical data in each cycle is different, in order to facilitate subsequent transformation, the data needs to be resampled first so that the data in different cycles can maintain the waveform while maintaining the same length. Assume that the length of the original sequence is N and the resampling length is M, that is, M time points are created evenly between the start and end time of the original time series, where M≠N. For each new time point, find its two nearest neighboring points in the original sequence, and then perform linear interpolation between these two points. The corresponding interpolation at the newly created time point is expressed as follows:

[0048]

[0049] Where t represents the newly created time point; t i and t i+1 represents the two nearest neighboring points of the newly created time point in the original sequence; D(t) represents the interpolation value corresponding to the newly created time point; D(t i ) and D(t i+1 ) represent the voltage, current or temperature value of the corresponding nearest neighbor point respectively.

[0050] According to experience, in a specific embodiment of the present invention, the first and last cycle data of five consecutive cycle data are subtracted to obtain battery degradation data based on charging voltage, current and temperature within five cycles, and the expression is as follows:

[0051]

[0052] in, Represents the battery degradation data from the jth cycle to the j+4th cycle; represents the degradation data of the i-th battery under the j-th cycle; Represents the degradation data of the i-th battery under the j+4-th cycle.

[0053] Then, the battery degradation data of charging voltage, current and temperature are converted into grayscale, and the corresponding grayscale image is obtained as the first input data and input into the convolutional neural network; the shape of the grayscale image is Since the pixel value range of the grayscale image is from 0 to 255, the grayscale value of the battery degradation data after grayscale conversion is expressed as follows:

[0054]

[0055] in, Represents the grayscale converted value of the i-th battery degradation data under the j-th cycle; and represent the minimum and maximum values ​​of the battery degradation data under the j-th cycle respectively; round(·) means rounding the result to the nearest integer.

[0056] S3, performing data conversion on the second data set based on a rolling prediction method to obtain second input data;

[0057] Specifically, before the battery capacity history data is input into the long short-term memory network, it needs to be converted into input layer and label data. The present invention uses the rolling prediction method to complete the conversion of the battery capacity history data. The main idea is to divide the data into many windows of specified size, and then roll the window forward at each time step and remodel it to add a new time step, so as to better reflect the current state of the system. Similarly, if the rolling window size is set to 5 and the prediction step size is set to 1, L-5 capacity sequences can be generated, and the expression is as follows:

[0058]

[0059] Among them, X represents the LSTM input layer data; Y represents the label data; C L Indicates the battery capacity after the Lth cycle.

[0060] S4, construct prediction models based on convolutional neural networks and long short-term memory networks optimized by sand cat colonies;

[0061] Specifically, the prediction model is obtained by connecting a convolutional neural network and a long short-term memory network based on sand cat swarm optimization in parallel. The first part of the convolutional neural network can automatically capture the spatial hierarchical structure characteristics of the input data, reduce the number of network parameters by sharing weights and local connections, and improve the computational efficiency and generalization ability of the model. Figure 3 ,The convolutional neural network contains convolutional layers, pooling layers and fully connected layers.

[0062] The convolution layer is the core of the convolutional neural network. The convolution layer contains multiple convolution kernels. Each convolution kernel slides on the input data and generates a set of output nodes. Each output node extracts the features of the input node through the convolution operation. All output nodes generated by the same convolution kernel form a feature map, and the number of feature maps is determined by the number of convolution kernels. The expression of the convolution operation is as follows:

[0063] F k =W k ·E+b k

[0064] Among them, F k represents the feature map with index k; W k represents the convolution kernel with index k; E represents the input image; b k represents the bias term with index k; k represents the index of the feature map.

[0065] In order to increase the nonlinear expression ability of the model, a nonlinear activation function is usually connected after the convolution layer, such as the ReLU (Rectified Linear Unit) function; this function can set all negative values ​​to 0 and keep the positive values ​​unchanged, which helps the model learn more complex features. The expression of the ReLU function is as follows:

[0066]

[0067] In the pooling layer, in order to reduce the amount of calculation and prevent overfitting, a fixed-size sliding window is used to select the maximum value in the region, thereby reducing the spatial dimension of the feature map. For an input feature map of size n, a sliding window of size ω is used with a stride of s, and the size of the output feature map becomes:

[0068]

[0069] The fully connected layer is located at the end of the network, where each neuron is connected to all neurons in the previous layer, integrating the previously extracted local features into global features. Specifically, in the fully connected layer, the feature map extracted by the pooling layer is linearly transformed through a weight matrix and a bias term is added to serve as the data before fusion with the output of another part of the model.

[0070] The second part is a long short-term memory network based on sand cat swarm optimization, which includes a sand cat swarm optimization module and a long short-term memory network.

[0071] Reference Figure 4, the LSTM network achieves the function of long-term memory by introducing new internal states and gating mechanisms, selectively adding new information and forgetting previously accumulated information. The LSTM network contains a forget gate f t , input gate i t and output gate o t , which is calculated as follows:

[0072]

[0073] Among them, f t represents the forget gate; i t represents the input gate; o t represents the output gate; σ(·) represents the activation function, Its output is between 0 and 1; x t Indicates the input of the current unit; h t-1 represents the hidden state of the previous unit; W * and b * (*∈{f,o,i}) represents the weight matrix and bias corresponding to the forget gate, output gate and input gate respectively.

[0074] The forget gate controls how much information of the previous unit needs to be forgotten, and the input gate is used to decide how much information of the current unit to keep. Therefore, the calculation method for updating the internal state of the current unit is:

[0075] c t =f t ⊙c t-1 +i t ⊙tanh(W c [x t ,h t-1 ]+b c )

[0076] Among them, ⊙ represents the XOR operation; W c and b c Represent the weight matrix and bias respectively; activation function Its output interval is (-1,1); c t-1 Indicates the current internal state of the unit before updating; c t Represents the updated internal state of the current unit.

[0077] The output gate controls how much information of the internal state needs to be output. The final output of the current unit of the long short-term memory network is:

[0078] h t =o t ⊙tanh(c t )

[0079] The long short-term memory network can better capture longer-distance temporal dependencies, and the hidden state of the last unit of the long short-term memory network can represent the model's overall understanding and memory of the input sequence.

[0080] However, the LSTM network has many hyperparameters, such as the number of hidden layer neurons, learning rate, regularization coefficient, and number of training cycles, and the selection of these parameters has a significant impact on the model performance. The Sand Cat Swarm Optimization Algorithm is a new meta-heuristic algorithm that simulates the search and hunting behavior of sand cats. It can automatically search for the optimal or near-optimal parameter combination by adaptively adjusting parameters such as auditory sensitivity, randomly initialized search space, and random angles. The present invention uses the Sand Cat Swarm Optimization Algorithm to optimize the hyperparameters of the LSTM network, which can effectively improve the performance of the LSTM network and enable it to capture relevant feature information more comprehensively.

[0081] Reference Figure 5 ,The main steps of the sand cat swarm optimization algorithm are as follows:

[0082] Step 1: randomly generate a certain number of sand cat individuals and initialize the sand cat population;

[0083] Specifically, in the sand cat swarm optimization algorithm, a certain number of sand cat individuals are randomly generated, and each individual represents a potential solution containing long short-term memory network hyperparameters.

[0084] Step 2: Construct a fitness function and calculate the fitness value of the current position of the sand cat individual;

[0085] Specifically, the prediction error function of the entire model is used as the fitness function of the sand cat group. After substituting the position of each sand cat, different fitness values ​​can be obtained. The optimal position can be selected according to the optimal fitness value, and other sand cat individuals move to this position.

[0086] Step 3: updating the position of the individual sand cat based on the updating strategy in the prey search phase and the updating strategy in the prey attack phase;

[0087] Specifically, when the sand cat is in the prey search stage, the sand cat individual with a random initial position begins to search for the best prey position in each iteration; the search mechanism of the sand cat mainly relies on low-frequency perception, and its sensitivity range to low frequencies is r G , as the sand cat approaches its prey, r G The value of decreases linearly from 2 to 0, and its expression is as follows:

[0088]

[0089] Among them, S M represents the auditory characteristic parameter of the sand cat, and its value is set to 2; D c Indicates the current iteration number; Dd Indicates the maximum number of iterations.

[0090] During the search process, in order to find other possible best prey locations, each sand cat individual needs to constantly update its own position. The update mechanism of its position is as follows:

[0091] P(t+1)=r[P bc (t)-rand(0,1)P c (t)]

[0092] Where P(t+1) represents the position of each sand cat after the t+1th iteration; r = r G rand(0,1) represents different sensitivity ranges; rand(0,1) represents a random number between 0 and 1, which is used to improve randomness and avoid the optimization process from falling into local extreme values; P bc (t) represents the optimal individual position after the tth iteration; P c (t) represents the current individual position after the tth iteration.

[0093] When the sand cat is in the attacking stage, assuming that the sensitive range of the sand cat is a circle, each sand cat randomly obtains an angle to approach the prey through the roulette wheel selection algorithm. The position of each sand cat after iteration t+1 can be calculated as:

[0094]

[0095] Among them, P rnd It represents the optimal position P of each sand cat after t iterations b (t) and the current position P c (t); P(t+1) represents the position after iteration t+1 times; θ represents a random angle between 0 and 360°.

[0096] In order to ensure that sand cat individuals can seamlessly switch positions between the two stages of searching and attacking prey and accelerate the convergence speed, the parameter R = 2r is introduced. G rand(0,1)-r G , and finally the strategy of updating the individual position of the sand cat can be expressed as:

[0097]

[0098] When |R|≤1, the sand cat enters the prey attack stage at the next moment; when |R|>1, it enters the global prey search stage, and the position at the next moment can be any position between the current position and the hunting position.

[0099] Step 4, looping steps 2 and 3 until the maximum number of iterations is met, and determining the optimal sand cat position as a hyperparameter of the long short-term memory network based on the fitness value, which can effectively improve the performance and accuracy of the long short-term memory network.

[0100] S5. Training the prediction model based on the first input data and the second input data, allowing the model to learn and grasp the changing law of the battery state through the training data, and obtaining a trained prediction model;

[0101] S5.1. Inputting the first input data and the second input data into the prediction model to obtain a first prediction result;

[0102] S5.2. Calculate the value of the model loss function based on the first prediction result and the actual observation data;

[0103] Specifically, the value of the model loss function is expressed as follows:

[0104]

[0105] Where MSE stands for mean square error; N stands for the number of samples; Represents the output value of the i-th sample of the prediction model; Y i Represents the true value of the i-th sample.

[0106] S5.3. Adjust the parameters of the prediction model based on the value of the model loss function and the fitness value to obtain a trained prediction model.

[0107] S6. Predict the battery status in the future based on the trained prediction model to obtain a prediction result.

[0108] Specifically, the prediction result is a key indicator of the state of health (SOH). In order to evaluate the accuracy of the prediction, the present invention compares the prediction result with the actual observation data and calculates the model evaluation index to reflect the accuracy of the model prediction.

[0109] The model evaluation indicators include mean absolute error, mean absolute percentage error and determination coefficient, and their expressions are as follows:

[0110]

[0111]

[0112] Among them, MAE stands for mean absolute error; MAPE stands for mean absolute percentage error R 2 represents the coefficient of determination; N represents the sample size; Represents the output value of the i-th sample of the prediction model; Y iRepresents the true value of the i-th sample; Represents the mean of N samples.

[0113] In the specific embodiment of the present invention, the smaller the mean absolute error, mean absolute percentage error, mean square error and fitness value, the higher the accuracy of the prediction model; the determination coefficient range is (0,1), and the closer it is to 1, the better the fitting effect.

[0114] In order to verify the effectiveness of the method of the present invention, the specific embodiment of the present invention uses the charging voltage, current, temperature and capacity data of lithium batteries numbered B0005 and B0018 in the NASA public data set for experimental analysis. Among them, all the data of B0005 and the first 50 charge and discharge cycle data of B0018 are used as training sets, and the remaining data of B0018 are used as test sets. In terms of data preprocessing, taking the charging voltage, current and temperature data of B0005 based on the 1st to 5th cycles as an example, the generation process of the battery degradation grayscale map is as follows: Figure 6 As shown in Figure 2. In terms of model building, after obtaining the optimal hyperparameters of the long short-term memory network through the sand cat swarm optimization algorithm, the long short-term memory neural network was constructed, the number of hidden layers was set to 2, each hidden layer contained 64 neurons, the number of training iterations was 300, and the sliding window was 5. After training the prediction model, the test data was input into the trained prediction model, and the prediction results were as follows: Figure 7 The prediction performance evaluation of the prediction model is shown in Table 1. Combined with the chart, it can be seen that the prediction method proposed by the present invention not only has a good prediction effect, but also can measure the process of battery capacity regeneration.

[0115] Table 1 Prediction model prediction performance evaluation table

[0116] MAE MAPE <![CDATA[R 2 ]]> <![CDATA[1.36×10 -2 ]]> 0.9351 0.9579

[0117] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A satellite power system fault prediction method based on multi-model combination, It is characterized in that The following steps are involved: Collecting satellite power system battery data to obtain a first data set and a second data set; Performing grayscale image degradation on the first data set to obtain first input data; Performing data conversion on the second data set based on a rolling prediction method to obtain second input data; Constructing prediction models based on convolutional neural networks and long short-term memory networks optimized by sand cat colonies; Training the prediction model based on the first input data and the second input data to obtain a trained prediction model; Based on the trained prediction model, the battery status in the future is predicted to obtain the prediction result.

2. According to the satellite power system fault prediction method based on multi-model combination as described in claim 1, It is characterized in that The first data set includes charging voltage data, charging current data and charging temperature data of the lithium battery in each charge and discharge cycle; the second data set includes historical data of battery capacity.

3. According to the satellite power system fault prediction method based on multi-model combination as described in claim 1, It is characterized in that The first input data is expressed as follows: in, Represents the grayscale converted value of the i-th battery degradation data under the j-th cycle; Represents the battery degradation data from the jth cycle to the j+4th cycle; and Respectively represent the minimum and maximum values ​​of battery degradation data under the jth cycle; represents the degradation data of the i-th battery under the j-th cycle; represents the degradation data of the ith battery in the j+4th cycle; round(·) means rounding the result to the nearest integer; t represents the newly created time point; t i and t i+1 represents the two nearest neighboring points of the newly created time point in the original sequence; D(t) represents the interpolation value corresponding to the newly created time point; D(t i ) and D(t i+1 ) represent the voltage, current or temperature value of the corresponding nearest neighbor point respectively.

4. According to the satellite power system fault prediction method based on multi-model combination as described in claim 1, It is characterized in that The second input data is expressed as follows: Among them, X represents the LSTM input layer data; Y represents the label data; C L Indicates the battery capacity after the Lth cycle.

5. According to the satellite power system fault prediction method based on multi-model combination as described in claim 1, It is characterized in that The prediction model is obtained by connecting a convolutional neural network and a long short-term memory network based on sand cat swarm optimization in parallel, where: The convolutional neural network comprises a convolutional layer, a pooling layer and a fully connected layer; The long short-term memory network based on sand cat swarm optimization includes a sand cat swarm optimization module and a long short-term memory network; The long short-term memory network includes a forget gate, an input gate and an output gate.

6. According to the satellite power system fault prediction method based on multi-model combination as described in claim 1, It is characterized in that The sand cat swarm optimization module is used to optimize the hyperparameters of the long short-term memory network, including the following steps: Randomly generate a certain number of sand cat individuals and initialize the sand cat population; Construct a fitness function and calculate the fitness value of the current position of the sand cat individual; The position of the individual sand cat is updated based on the updating strategy in the searching prey phase and the updating strategy in the attacking prey phase; The fitness value calculation and position update steps of the individual sand cats are repeated until a maximum number of iterations is met, and the optimal sand cat position is determined based on the fitness value as a hyperparameter of the long short-term memory network.

7. A satellite power system fault prediction method based on multi-model combination according to claim 1, It is characterized in that The step of training the prediction model based on the first input data and the second input data to obtain a trained prediction model specifically includes: Inputting the first input data and the second input data into the prediction model to obtain a first prediction result; Calculate the value of the model loss function based on the first prediction result and the actual observation data; The parameters of the prediction model are adjusted based on the value of the model loss function and the fitness value to obtain a trained prediction model.

Citation Information

Patent Citations

  • Flotation process fault early warning method for predicting deviation degree based on BiLSTM

    CN116597350A

  • Lithium battery life prediction method based on multi-scale integrated deep learning model

    CN117473864A

  • Lithium battery remaining service life prediction method and system based on optimized neural network

    CN117665627A

  • Intelligent fault diagnosis method based on multi-task feature sharing neural network

    US20220269925A1

Cited By

  • Aero-engine infrared radiation curve prediction method and device and electronic equipment

    CN121328313A