Satellite battery life prediction method based on solar environment characteristics

Abstract

The invention relates to the technical field of satellite power supply management and life prediction, and discloses a solar environment characteristic-based satellite battery life prediction method, which comprises the following steps of: generating and preprocessing satellite battery data for training and solar panel data for training, the satellite battery data for training comprises satellite battery state data and satellite battery actual life data; building a CNN-Transform model, and training the CNN-Transform model by taking the satellite battery state data and the solar panel data for training as the input of the CNN-Transform model, and taking the satellite battery actual life data as the output of the CNN-Transform model; and inputting actual satellite battery state data and actual solar panel data, which are acquired in real time, into the trained CNN-Transformer model so as to output predicted satellite battery remaining life data. According to the invention, a more accurate and reliable result is provided for the residual life prediction of the satellite battery.

Description

Technical Field

[0001] The invention relates to the technical field of satellite power management and life prediction, and in particular to a satellite battery life prediction method based on solar energy environment characteristics. Background Art

[0002] With the continuous development of modern technology, satellites play an increasingly important role in our lives. According to the 2023 Global Satellite Industry Report, the number of satellites in orbit worldwide will total 7,316 in 2022, a year-on-year increase of 51%. In addition, according to a report by market research firm Allied Market Research, the global satellite market size is expected to grow from US$260 billion in 2020 to US$350 billion in 2027, with an average annual growth rate of 4.5%.

[0003] As the core energy source of satellites, the life and reliability of satellite batteries directly affect the normal operation and mission execution of satellites. The performance of satellite batteries is not only related to the stability and reliability of satellite operation in orbit, but also directly determines the success rate and service life of satellite missions. For example, a study by the National Aeronautics and Space Administration (NASA) of the United States showed that the decline in battery performance is one of the key factors in shortening the life of satellites. About 50% of satellites suffer performance degradation or failure due to battery problems during their missions. Therefore, accurate prediction of satellite battery life has important practical significance and application value. On the one hand, the prediction of battery life can provide a scientific basis for the maintenance and management of satellites, ensuring that corresponding maintenance measures are taken before the battery performance declines, and avoiding the interruption of satellite missions due to battery failure. On the other hand, accurate battery life prediction can improve the use efficiency of satellites, extend their on-orbit working time, and maximize their functions and benefits.

[0004] As satellite missions continue to increase in complexity and diversity, reliable battery life prediction can also help optimize satellite mission planning and resource allocation. For example, by accurately predicting battery life before a satellite enters a critical mission phase, necessary maintenance can be performed in advance to ensure the smooth progress of the mission. In addition, in a multi-satellite system, battery life prediction can help coordinate the task division and energy management of each satellite and improve the operating efficiency of the entire system.

[0005] However, existing satellite battery life prediction methods have some significant shortcomings and limitations. These methods mainly rely on the characteristic data of the battery itself, such as voltage, current, temperature, etc., and usually adopt data-driven deep learning models, such as LSTM (Long Short-Term Memory Network). Although these models perform well in some scenarios, there are still many deficiencies in practical applications, including poor model interpretability, and prediction accuracy and efficiency that are not ideal. In addition, existing methods often ignore the complex relationship between solar panels and battery systems. As an important part of the satellite power system, the working state and environmental characteristics of solar panels have a direct impact on the battery charging and discharging process. For example, when the solar panel is operating on the sun-facing side and the sun-facing side, the received light intensity will have a significant difference. This periodic light change directly affects the battery charging efficiency and life. In addition, fluctuations in ambient temperature will also affect the power generation efficiency of the solar panel, which in turn affects the battery charging and discharging process. Long-term exposure to high or low temperature environments may significantly change the discharge rate and health of the battery. These environmental factors jointly determine the battery life under different operating conditions. Therefore, developing an efficient prediction model that can integrate multi-source data and accurately capture the complex relationships within the system is of great significance for ensuring the reliability of satellite batteries and extending their service life. Summary of the invention

[0006] The present invention provides a satellite battery life prediction method based on solar energy environment characteristics. The method can break the strong dependence of satellite battery life prediction on single battery data, provide key technologies, and effectively alleviate the problem of low prediction accuracy caused by insufficient data or data noise in the battery life prediction process, thereby providing more accurate and reliable results for satellite battery life prediction.

[0007] The present invention is achieved through the following technical solutions:

[0008] A satellite battery life prediction method based on solar energy environment characteristics, comprising:

[0009] Generate and pre-process training satellite battery data and training solar panel data, wherein the training satellite battery data includes satellite battery status data and actual remaining life data of the satellite battery;

[0010] Constructing a CNN-Transformer model, and using the satellite battery status data and the training solar panel data as inputs of the CNN-Transformer model, and the satellite battery remaining life data as outputs of the CNN-Transformer model to train the CNN-Transformer model;

[0011] The actual satellite battery status data and actual solar panel data collected in real time are input into the trained CNN-Transformer model to output the predicted satellite battery remaining life data.

[0012] As an optimization, the training satellite battery data includes normal training satellite battery data and faulty training satellite battery data, and the training solar panel data includes normal training solar panel data and faulty training solar panel data.

[0013] As an optimization, the satellite battery status data includes battery voltage, battery current, battery temperature, single cell voltage, battery internal resistance and battery rated capacity;

[0014] The training solar panel data includes solar panel angle, light intensity, solar panel temperature, solar panel current and solar panel voltage.

[0015] As an optimization, the specific process of preprocessing the training satellite battery data and the training solar panel data includes:

[0016] Cleaning and normalizing the generated training satellite battery data and training solar panel data to remove noise and outliers and ensure data quality and consistency;

[0017] Using an interpolation algorithm and synchronization technology combined with linear interpolation or spline interpolation, the training satellite battery data and the training solar panel data are aligned in time series to ensure that each of the training satellite battery data and the corresponding training solar panel data have the same timestamp;

[0018] The training satellite battery data and the training solar panel data are grouped according to the phase, wherein the training satellite battery data and the training solar panel data within the same phase range are grouped together, and then the kernel density estimation method is used to expand the number of the training satellite battery data and the training solar panel data in each group, thereby obtaining a plurality of expanded groups of data sets, wherein each group of data sets includes the expanded training satellite battery data and the training solar panel data.

[0019] As an optimization, from the data flow transmission direction, the CNN-Transformer model includes a feature extraction module, a feature interaction module and a sequence modeling layer in sequence, wherein:

[0020] The feature extraction module is used to extract the first feature of the satellite battery status data and the second feature of the training solar panel data respectively;

[0021] The feature interaction module is used to fuse the first feature and the second feature corresponding to the same timestamp to obtain a third feature corresponding to the timestamp;

[0022] The sequence modeling module outputs satellite battery predicted life data through the third feature.

[0023] As an optimization, the feature extraction module includes two CNN modules, wherein one CNN module is used to extract a first feature of the satellite battery status data, and the other CNN module is used to extract a second feature of the training solar panel data.

[0024] As an optimization, each of the CNN modules includes two convolutional layers and an average pooling layer arranged after each convolutional layer.

[0025] As an optimization, the feature interaction module is the first multi-head attention mechanism model.

[0026] As an optimization, the sequence modeling module includes a Transformer model and a fully connected layer arranged at the output end of the Transformer model.

[0027] As an optimization, the Transformer model includes a position encoding layer and N encoders arranged after the position encoding layer. Along the transmission direction of the data stream, each of the encoders sequentially includes a second multi-head attention mechanism model, a first residual connection & normalization layer, a feedforward neural network layer and a second residual connection & normalization layer, wherein two feedforward neural network layers are provided, the output ends of the first residual connection & normalization layer are respectively connected to the input ends of the two feedforward neural network layers, and the output ends of the two feedforward neural network layers are respectively connected to the input ends of the second residual connection & normalization layer.

[0028] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0029] The method of the present invention can break the strong dependence of satellite battery life prediction on single battery data, provide key technologies, and effectively alleviate the problem of low prediction accuracy caused by insufficient data or data noise in the battery life prediction process, thereby providing more accurate and reliable results for satellite battery life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0031] Figure 1 A flowchart of a satellite battery life prediction method based on solar energy environment characteristics according to the present invention;

[0032] Figure 2 It is a structural schematic diagram of the CNN-Transformer model described in the present invention. Figure 3 It is a schematic diagram of the structure of the encoder in the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0034] In response to the problems raised by the background technology, the present invention predicts the life of satellite batteries through the CNN-Transformer model. By combining the attention mechanism of CNN and Transformer, the model can identify which input features have the greatest impact on the prediction results. In addition, the attention weights are visualized to intuitively display the decision-making basis of the model. This visualized attention weight can provide a clearer explanation, allowing researchers and engineers to understand the decision-making process of the model, and the transformer architecture usually adopts parallel computing, making the CNN-transformer model more efficient in processing satellite battery data, which can not only improve the performance and reliability of a single satellite, but also promote the technological progress and market development of the entire satellite industry. Next, the process of this method is specifically introduced.

[0035] This embodiment 1 provides a satellite battery life prediction method based on solar energy environment characteristics, such as Figure 1 As shown, including:

[0036] S1. Generate and pre-process training satellite battery data and training solar panel data, wherein the training satellite battery data includes satellite battery status data and satellite battery actual life data; generate satellite battery and solar panel data, and pre-process them to ensure data quality and consistency.

[0037] In some embodiments, the specific process of S1 is:

[0038] S1.1. Build a data twin model, including fault type definition module, physical model module, data generation module, parameter adjustment module and data output module.

[0039] First, it is necessary to generate training satellite battery data and training solar panel data. In some embodiments, the satellite battery status data includes battery voltage, battery current, battery temperature, single cell voltage, battery internal resistance and battery rated capacity; the training solar panel data includes solar panel angle, light intensity, solar panel temperature, solar panel current and solar panel voltage.

[0040] Specifically, the training satellite battery data and the number of training solar panels are generated by a data twin model, and the data twin model includes a fault type definition module, a physical model module, a data generation module, a parameter adjustment module, a data output module and a battery remaining life management module, wherein:

[0041] The fault type definition module is used to simulate and define data under different faults;

[0042] The physical model module is used to simulate the physical characteristics of the satellite in a normal environment based on physical principles, and generate satellite battery data for normal training and solar panel data for normal training;

[0043] The data generation module is used to generate satellite battery data for fault training and solar panel data for fault training under the conditions simulated by the fault type definition module and the physical model module;

[0044] There are two types of data generated here for training: 1. Normal data 2. Fault data (such as overheating, solar panel shading) or data under extreme conditions (such as solar magnetic storms). Normal data can be generated in the physical model module, and fault data must be generated in the data generation module.

[0045] The parameter adjustment module is used to adjust the parameters of the data twin model when the satellite operation state changes, so that the data twin model can better adapt to the new state;

[0046] The parameters here may refer to dynamic environmental parameters. For example, at extremely high or low temperatures, the battery charging and discharging efficiency will change. The parameter adjustment module can dynamically adjust the temperature adjustment coefficient to simulate this effect.

[0047] The data output module is used to output training satellite battery data and training solar panel data.

[0048] Different solar panel environmental characteristics are injected into the data twin model to generate solar panel data and battery data under different environmental conditions, as well as data under normal conditions. Specifically, solar panel data should include parameters such as solar panel angle, light intensity, temperature, current and voltage, while battery data should include parameters such as battery voltage, current, temperature, single cell voltage, battery internal resistance and rated capacity.

[0049] Solar panel environmental characteristics (solar panel data) refer to light intensity, temperature, angle of solar panels, shadows (satellite attitude, Earth or Moon shading), and possible astronomical phenomena (such as solar magnetic storms).

[0050] Some of these environmental characteristics will be input into the physical model module, such as solar panel angle, lighting, temperature, shadow, etc., and some will be input into the data generation module to generate characteristic data under different space environments, such as solar magnetic storms.

[0051] The battery remaining life management module is used to calculate the remaining life of the battery based on the current environment and battery status.

[0052] Alternatively, the training satellite battery data and training solar panel data can be obtained through long-term data collection. For the data collection of a satellite battery, the status data of the satellite battery and the solar panel data will be sampled regularly from a certain moment, so that normal data and fault data can be collected.

[0053] Next, the training satellite battery data and the training solar panel data are pre-processed.

[0054] S1.2. Cleaning and normalizing the generated training satellite battery data and training solar panel data to remove noise and outliers and ensure data quality and consistency;

[0055] The generated data is cleaned and normalized to remove noise and outliers to ensure data quality and consistency. Data cleaning includes deleting missing values ​​and smoothing abnormal spike data. Normalization ensures that different types of data are processed on the same scale by scaling the data to a standard range.

[0056] S1.3, using an interpolation algorithm, synchronization technology combined with linear interpolation or spline interpolation, aligning the training satellite battery data and the training solar panel data in time series to ensure that each of the training satellite battery data and the corresponding training solar panel data have the same timestamp;

[0057] Align data from different sources in time series to ensure that each data has the same timestamp. Use interpolation algorithms and synchronization technology to solve the time deviation problem during data collection. Use linear interpolation or spline interpolation to align data to a unified time base to reduce the impact of time deviation on CNN-Transformer model training.

[0058] It should be noted that time alignment does not mean that the data is aligned when it is collected or generated. It is sufficient to perform time alignment on the collected or generated data in the preprocessing step.

[0059] S1.4. Grouping the training satellite battery data and the training solar panel data according to the phase, wherein the training satellite battery data and the training solar panel data within the same phase range are grouped together, and then using the kernel density estimation method to expand the number of each group of the training satellite battery data and the training solar panel data, thereby obtaining several groups of expanded data sets, wherein each group of data sets includes the expanded training satellite battery data and the training solar panel data.

[0060] The data are grouped according to the phase, and the kernel density estimation method is used to expand the data of each group. In this way, the model's ability to handle periodic data can be enhanced and the accuracy of prediction can be improved.

[0061] Specifically, the steps of the kernel density estimation method are: calculate the probability density function of each point, and randomly sample from the probability density function to generate new data points. These new data points conform to the distribution characteristics of the original data and can be used to enhance the data set. The specific formula of the kernel density estimation method is as follows:

[0062]

[0063] in, represents the value corresponding to the new data point, n is the number of data points, x i is the ith data point, h is the bandwidth, and K represents the kernel function in the kernel density estimation method, which is used to calculate the density distribution of data points, usually using a Gaussian kernel.

[0064] S2. Build a CNN-Transformer model, and use the satellite battery status data and the training solar panel data as the input of the CNN-Transformer model, and the actual life data of the satellite battery as the output of the CNN-Transformer model to train the CNN-Transformer model; build and train a CNN-Transformer model that can fuse multi-source data to improve the accuracy and reliability of battery life prediction.

[0065] In some embodiments, the specific process of training the CNN-Transformer model by using the satellite battery status data and the training solar panel data as the input of the CNN-Transformer model and the actual life data of the satellite battery as the output of the CNN-Transformer model is as follows:

[0066] The satellite battery status data and the training solar panel data are used as inputs of the CNN-Transformer model, so that the satellite battery predicted life data is output through the CNN-Transformer model, the satellite battery predicted life data and the satellite battery actual life data are calculated by loss function, and the CNN-Transformer model is reversely optimized through the loss function, so as to achieve the training of the CNN-Transformer model. The loss function can be a conventional loss function.

[0067] In some embodiments, from the data stream transmission direction, the CNN-Transformer model includes a feature extraction module, a feature interaction module and a sequence modeling layer in sequence, wherein:

[0068] The feature extraction module is used to extract the first feature of the satellite battery status data and the second feature of the training solar panel data respectively;

[0069] The feature interaction module is used to fuse the first feature and the second feature corresponding to the same timestamp to obtain a third feature corresponding to the timestamp;

[0070] The sequence modeling module outputs satellite battery predicted life data through the third feature.

[0071] Specifically, the feature extraction module includes two CNN modules, one of which is used to extract a first feature of the satellite battery status data, and the other CNN module is used to extract a second feature of the training solar panel data.

[0072] Use convolutional neural network (CNN) to extract features from solar panel data and battery data.

[0073] In the feature extraction of battery data, CNN can capture the changing trend of battery voltage and current to detect the battery charging and discharging process, and can also capture abnormal changes in pressure, etc., that is, the first feature is the characteristic data of battery voltage, battery current, battery temperature, single cell voltage, battery internal resistance and battery rated capacity. In the feature extraction of solar panel data, CNN can capture the light intensity change pattern and periodic fluctuations of voltage and current, etc., that is, the second feature is the characteristic data of solar panel angle, light intensity, solar panel temperature, solar panel current and solar panel voltage.

[0074] In some embodiments, each of the CNN modules includes 2 convolutional layers and an average pooling layer arranged after each convolutional layer.

[0075] For any CNN module (also called convolutional neural network), the convolutional neural network (CNN) uses filters to perform convolution operations on the input data to extract local features.

[0076]

[0077] Where I is the input data, K is the convolution kernel, and f(x,y) is the output of the convolution layer operation. m and n are the height and width of the convolution kernel, and x and y represent the coordinates of the convolution output data.

[0078] The output of the convolutional layer operation is transformed using the activation function ReLU.

[0079] ReLU(f(x,y))=ma(0,f(x,y))

[0080] Finally, f(x,y) is downsampled to reduce the data dimension. The pooling layer usually uses maximum pooling or average pooling. Since the task is to predict the life of satellite batteries, average pooling is used.

[0081]

[0082] Among them, m*n is the size of the pooling window.

[0083] Two convolutional layers and two average pooling layers are used. First, the first convolutional layer inputs the satellite battery status data or the training solar panel data, and then gets an output through the first average pooling layer. Then the second convolutional layer inputs the output of the first average pooling layer, and then gets the output through the second average pooling layer. The output of the second average pooling layer is the first feature or the second feature.

[0084] In some embodiments, the feature interaction module is a first multi-head attention mechanism model. The outputs of the two CNN models are jointly input into the first multi-head attention mechanism model, and the first multi-head attention mechanism model performs feature interaction fusion on the first feature and the second feature.

[0085] Specifically, the high-level features of solar panel data and battery data extracted by the two CNN models are used as input in the form of multi-dimensional feature vectors. The self-attention mechanism captures the relationship between data features through the similarity between query, key, and value. The input feature vector is linearly transformed to generate Q, K, and V matrices, and then the dot product of Q and K is calculated and divided by the square root of the key dimension, and the attention score is obtained through the softmax function.

[0086]

[0087] Among them, Q, K, and V are query, key, and value matrices respectively, and d k The dimension of the key.

[0088] Each attention head calculates the score independently, and the calculation for each head is as follows:

[0089] head i =Attention(QW i Q ,KW i K ,VWV i V )

[0090] Finally, the outputs of these heads are concatenated and linearly transformed to obtain the fused feature representation.

[0091] MultiHead(Q,K,V)=Concat(head1,head2,...head n )W O

[0092] Among them, W i Q , W i K , W i V , W O is the parameter matrix that the model needs to learn.

[0093] The output of the first multi-head attention mechanism model effectively integrates the first feature of the battery data and the second feature of the solar panel data to form a high-quality feature vector (third feature) containing multi-source information (satellite battery data and solar panel data).

[0094] In some embodiments, the sequence modeling module includes a Transformer model and a fully connected layer arranged at the output end of the Transformer model.

[0095] In some embodiments, the Transformer model includes a position encoding layer and N encoders. Along the transmission direction of the data stream, the Transformer model includes a position encoding layer and N encoders arranged after the position encoding layer. Along the transmission direction of the data stream, each encoder includes a second multi-head attention mechanism model, a first residual connection & normalization layer, a feedforward neural network layer and a second residual connection & normalization layer in sequence, wherein two feedforward neural network layers are provided, and the output ends of the first residual connection & normalization layer are respectively connected to the input ends of the two feedforward neural network layers, and the output ends of the two feedforward neural network layers are respectively connected to the input ends of the second residual connection & normalization layer. N is a positive integer. In this embodiment, N is 6.

[0096] Use the Transformer model to capture the temporal dependency and long-distance dependency of data. The specific steps are as follows:

[0097] The fused feature vector (the third feature) is encoded, and periodic position encoding (PositionalEncoding) is added to obtain the encoding sequence corresponding to the third feature to retain the time series information of the satellite battery status data and the training solar panel data.

[0098]

[0099] Among them, pos is the position, P is the period, i is the dimension index, d model is the characteristic dimension.

[0100] Here, sine is used in even dimensions and cosine is used in odd dimensions in order to generate codes with different phases, so that the code at each position in a long sequence will not be repeated, and this can capture multi-scale temporal information and sequential relationships.

[0101] The encoded sequence obtained after encoding the third feature is input into the encoder, which contains multiple multi-head self-attention mechanisms and feedforward neural networks. The encoder is responsible for processing the input data and extracting high-dimensional feature representations. The calculation formula of the multi-head attention mechanism in the second multi-head attention mechanism model is the same as that in the first multi-head attention mechanism model.

[0102] Each head of the multi-head self-attention mechanism can independently learn and capture different pattern relationships in the input data. This mechanism enables the model to analyze the interaction between the solar panel and the battery from multiple angles and scales, thereby improving the expressiveness and predictive ability of the Transformer model. Each attention head has independent parameters, which allows each head to focus on different aspects or patterns in the input data. For example, an attention head can focus on the relationship between the pattern of light intensity changes over time and the battery charge state. An attention head can capture the long-term impact of the solar light cycle (such as the day-night cycle) on the battery health state. Batteries may decay faster when exposed to high-intensity light for a long time, while batteries rely mainly on the discharge process when they are on the dark side for a long time. This multi-head mechanism allows the Transformer model to learn and represent features in different subspaces in parallel, thereby enhancing the overall expressiveness of the Transformer model.

[0103] Since there is a strong linear relationship between the prediction of satellite life and battery capacity, the encoder pays special attention to the impact of the working status of the solar panels on the battery capacity when processing the data, thereby improving the accuracy of the life prediction.

[0104] The feature representation generated by the Transformer model is input into the fully connected layer, which is responsible for mapping the high-dimensional features to specific prediction values, namely the remaining useful life (RUL) of the satellite battery.

[0105] When training the CNN-Transformer model, 10,000 samples of satellite battery data and solar panel data were used for model training. The sample data came from the data twin system, covering the operating status of the satellite in the sun-facing and sun-facing environments, including multi-dimensional features such as battery voltage, current, temperature, and internal resistance. The data set is divided into a training set and a validation set in a ratio of 8:2. The CNN part contains 2 convolutional layers, the first layer contains 32 filters, the second layer contains 64 filters, the convolution kernel size is 3x3, the pooling method uses average pooling, and the window size is 2x2. The Transformer model uses a 6-layer encoder, each encoder contains 6 attention heads, the hidden layer dimension is 512, and the sequence length is set to 100. The period of the position encoding is set to 24, reflecting the daily cycle characteristics of the solar panel. The model is trained using the Adam optimizer, with an initial learning rate of 0.0001, a batch size of 64, and 50 rounds of training. The learning rate decay mechanism is used, the learning rate is reduced by 10% every 10 rounds, and a dropout of 0.1 is introduced in the Transformer part to prevent overfitting. The loss function is MSE, and the mean absolute error MAE on the validation set is monitored as the evaluation indicator.

[0106] S3, input the actual satellite battery status data and actual solar panel data collected in real time into the trained CNN-Transformer model, so as to output the predicted satellite battery remaining life data. That is, the trained CNN-Transformer model is used to predict the real-time data and output the remaining life of the satellite battery.

[0107] Before inputting the actual satellite battery status data and the actual solar panel data into the CNN-Transformer model, the actual satellite battery status data and the actual solar panel data are also preprocessed. The preprocessing process is the same as the preprocessing process of the training satellite battery data and the training solar panel data.

[0108] Specifically, the S3 process is:

[0109] S31: Collect the latest data from satellite batteries and solar panels in real time and pre-process it to ensure data quality and consistency.

[0110] S32: Input the preprocessed real-time data into the trained CNN-transformer model. The CNN-transformer model generates a prediction result of the remaining battery life through feature extraction, feature interaction and sequence modeling layers.

[0111] S33: Output the battery RUL result (the prediction result of the remaining battery life) and feed it back to the satellite control center as a basis for satellite maintenance and management.

[0112] The prediction results can help optimize satellite mission planning and resource allocation, and improve satellite operating efficiency and mission success rate.

[0113] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A satellite battery life prediction method based on solar energy environment characteristics, characterized in that: include: Generate and pre-process training satellite battery data and training solar panel data, wherein the training satellite battery data includes satellite battery status data and satellite battery actual life remaining data; Constructing a CNN-Transformer model, and using the satellite battery status data and the training solar panel data as inputs of the CNN-Transformer model, and the satellite battery actual life data as outputs of the CNN-Transformer model to train the CNN-Transformer model; The actual satellite battery status data and actual solar panel data collected in real time are input into the trained CNN-Transformer model to output the predicted satellite battery remaining life data.

2. A satellite battery life prediction method based on solar energy environment characteristics according to claim 1, characterized in that: The training satellite battery data includes normal training satellite battery data and fault training satellite battery data, and the training solar panel data includes normal training solar panel data and fault training solar panel data.

3. The satellite battery life prediction method based on solar energy environment characteristics according to claim 1 is characterized in that: The satellite battery status data includes battery voltage, battery current, battery temperature, single cell voltage, battery internal resistance and battery rated capacity; The training solar panel data includes solar panel angle, light intensity, solar panel temperature, solar panel current and solar panel voltage.

4. The satellite battery life prediction method based on solar energy environment characteristics according to claim 1 is characterized in that: The specific process of preprocessing the training satellite battery data and the training solar panel data includes: Cleaning and normalizing the generated training satellite battery data and training solar panel data to remove noise and outliers and ensure data quality and consistency; Using an interpolation algorithm and synchronization technology combined with linear interpolation or spline interpolation, the training satellite battery data and the training solar panel data are aligned in time series to ensure that each of the training satellite battery data and the corresponding training solar panel data have the same timestamp; The training satellite battery data and the training solar panel data are grouped according to the phase, wherein the training satellite battery data and the training solar panel data within the same phase range are grouped together, and then the kernel density estimation method is used to expand the number of the training satellite battery data and the training solar panel data in each group, thereby obtaining a plurality of expanded groups of data sets, wherein each group of data sets includes the expanded training satellite battery data and the training solar panel data.

5. The satellite battery life prediction method based on solar energy environment characteristics according to claim 1 is characterized in that: From the data flow transmission direction, the CNN-Transformer model includes a feature extraction module, a feature interaction module and a sequence modeling layer in sequence, wherein: The feature extraction module is used to extract the first feature of the satellite battery status data and the second feature of the training solar panel data respectively; The feature interaction module is used to fuse the first feature and the second feature corresponding to the same timestamp to obtain a third feature corresponding to the timestamp; The sequence modeling module outputs satellite battery predicted life data through the third feature.

6. A satellite battery life prediction method based on solar energy environment characteristics according to claim 5, characterized in that: The feature extraction module includes two CNN modules, one of which is used to extract a first feature of the satellite battery status data, and the other CNN module is used to extract a second feature of the training solar panel data.

7. A satellite battery life prediction method based on solar energy environment characteristics according to claim 6, characterized in that: Each of the CNN modules includes two convolutional layers and an average pooling layer arranged after each convolutional layer.

8. The satellite battery life prediction method based on solar energy environment characteristics according to claim 5 is characterized in that: The feature interaction module is the first multi-head attention mechanism model.

9. The satellite battery life prediction method based on solar energy environment characteristics according to claim 5, characterized in that: The sequence modeling module includes a Transformer model and a fully connected layer arranged at the output end of the Transformer model.

10. A satellite battery life prediction method based on solar energy environment characteristics according to claim 9, characterized in that: The Transformer model includes a position encoding layer and N encoders arranged after the position encoding layer. Along the transmission direction of the data stream, each of the encoders sequentially includes a second multi-head attention mechanism model, a first residual connection & normalization layer, a feedforward neural network layer and a second residual connection & normalization layer, wherein there are two feedforward neural network layers, the output ends of the first residual connection & normalization layer are respectively connected to the input ends of the two feedforward neural network layers, and the output ends of the two feedforward neural network layers are respectively connected to the input ends of the second residual connection & normalization layer.

Images