Method and system for predicting residual life of battery based on hybrid deep learning

Abstract

The invention discloses a hybrid deep learning-based battery residual life prediction method and system, and the method comprises the steps: extracting an effective charging interval from the multi-dimensional time series data of a battery in a charging and discharging process; battery health state prediction is carried out based on the battery health state sample extracted from the effective charging interval and the charging curve characteristics; predicting the remaining life of the battery in real time according to the prediction result of the battery health state; meanwhile, a battery health state prediction model combining a convolutional neural network and a long-short-term memory network is constructed, and the battery health state prediction model deeply extracts a feature relationship from multi-dimensional original time sequence data through the convolutional neural network; and then the extracted feature sequence is used as the input of the long-short-term memory network to learn the descending trend of the battery health state, and the recession features in the data can be better captured, so that the prediction result of the battery health state and the remaining service life becomes more accurate.

Description

Technical Field

[0001] This invention belongs to the field of battery health management technology, specifically relating to a method and system for predicting battery remaining life based on hybrid deep learning. Background Technology

[0002] In recent years, with the booming development of the new energy industry, electric vehicles and other means of transportation have become widely used, especially in high-frequency usage scenarios such as food delivery and express delivery. As the core power source, the performance, lifespan, and safety of batteries have become crucial. To improve work efficiency and user experience, battery swapping using charging cabinets has become a new battery swapping model. Accurately assessing and predicting the health status and remaining lifespan of each battery in the charging cabinet has become the core of achieving refined asset management.

[0003] Battery remaining life (RUL) prediction is a key technology in battery health management (BHM). It can provide early warning of potential safety risks to batteries, providing crucial decision-making basis for battery maintenance, replacement and retirement, thereby significantly reducing operating costs and improving safety.

[0004] Traditional battery life prediction methods mainly rely on simplified physical models, empirical models, or simple threshold alarms. For example, they may make rough estimates by recording the number of battery cycles or usage time, or trigger an alarm when a single indicator such as voltage or temperature exceeds a preset threshold. These methods have significant limitations: First, they are insufficient in describing the complex and nonlinear internal electrochemical degradation process of batteries, making it difficult to adapt to changing real-world operating environments (such as different charge / discharge rates and ambient temperatures); second, threshold alarms are a passive response mechanism, only alerting when a fault is about to occur or has already occurred, failing to achieve true "prediction" and "early warning"; third, these methods typically ignore the rich state information contained in the complete time-series curves of voltage, current, and temperature during battery charging and discharging.

[0005] Therefore, how to automatically extract deep features of battery aging from the massive time-series data provided by the battery management system (BMS) and use advanced machine learning algorithms to accurately predict the future lifespan of the battery has become a technical challenge and research hotspot that urgently needs to be solved in the field of battery management. Summary of the Invention

[0006] To address the technical problems mentioned in the background section, such as low accuracy, inability to effectively utilize massive time-series data, and poor early warning capabilities of traditional battery life prediction methods, the present invention aims to provide a battery remaining life prediction system and method based on hybrid deep learning to accurately predict battery health status and remaining life.

[0007] In a first aspect, this invention proposes a battery remaining life prediction method based on hybrid deep learning, the method comprising:

[0008] Acquire multidimensional time-series data of the battery; identify different effective charging intervals from the multidimensional time-series data of the battery and mark the number of cycles corresponding to each effective charging interval; obtain the battery health status value corresponding to each effective charging interval; construct a battery health status sample based on the battery health status value and the number of cycles, and extract charging curve features from the multidimensional time-series data based on the battery health status sample;

[0009] A battery health state prediction model is constructed. The battery health state prediction model includes a feature extraction module, a trend analysis module, and a fully connected layer connected in sequence. The feature extraction module is used to extract local features from the charging curve features. The trend analysis module is used to analyze the evolution of local features over time. The fully connected layer is used to map the output of the trend analysis module to the final battery health state value.

[0010] The battery health status prediction model is trained using battery health status samples and charging curve features. The trained battery health status prediction model is then used to predict the battery health status value corresponding to the latest cycle count of the tested battery. The battery health status values ​​corresponding to the historical cycle count and the latest cycle count are fitted, and the life end cycle count is obtained using the fitted curve. The difference between the life end cycle count and the latest cycle count is used as the remaining lifespan to complete the prediction of the remaining battery lifespan.

[0011] As a preferred method, the battery health status sample is cleaned in multiple stages before extracting charging curve features. The specific process is as follows:

[0012] Remove abnormal battery health status values ​​from the battery health status samples; perform first-order polynomial fitting on the remaining battery health status samples of different batteries respectively, and obtain the slope of the fitted line. If the number of remaining battery health status samples is less than the sample threshold or the slope is not less than zero, then remove all battery health status samples corresponding to that battery.

[0013] Preferably, the effective charging range satisfies the following conditions:

[0014] The starting point of the effective charging interval is greater than the current threshold, and the ending point of the effective charging interval is less than the current threshold; the number of data points within the effective charging interval is not less than the data point threshold; and the remaining capacity percentage corresponding to the ending point of the effective charging interval is 100%.

[0015] Preferably, the battery health status value corresponding to the effective charging interval is the battery health status value at the stable point of the effective charging interval; the stable point is obtained by adding a set index interval to the end index of the effective charging interval to obtain the stable point index.

[0016] Preferably, the fitting curve is obtained by fitting the battery health state value using a second-order polynomial function.

[0017] Preferably, the training process of the battery health status prediction model is as follows: the charging curve features corresponding to the historical cycle number are used as the input of the battery health status prediction model, the battery health status value corresponding to the current cycle number is used as the label to train the battery health status prediction model, and the root mean square error is used as the loss function to update the model parameters.

[0018] Preferably, the multidimensional timing data includes total voltage, current, motherboard temperature, remaining capacity percentage, remaining capacity, and nominal capacity; the charging curve features are composed of total voltage, current, and motherboard temperature.

[0019] Preferably, the battery health status value is the ratio of remaining capacity to nominal capacity.

[0020] Preferably, the feature extraction module employs a convolutional neural network, which includes multiple convolutional blocks connected in sequence and a flattening layer; the trend analysis module employs a long short-term memory network.

[0021] Secondly, this invention provides a battery remaining life prediction system based on hybrid deep learning, which is used to execute the aforementioned battery remaining life prediction method. The battery remaining life prediction system includes a data acquisition module, a preprocessing module, a feature extraction module, a battery health status prediction module, and a remaining life prediction module. The data acquisition module is used to acquire multi-dimensional time-series data of the battery during the charging and discharging process. The feature extraction module is used to extract effective charging intervals from the multi-dimensional time-series data and obtain battery health status samples and charging curve features corresponding to the effective charging intervals. The preprocessing module is used to transform, clean, and normalize the multi-dimensional time-series data. The battery health status prediction module is used to predict the battery health status based on the charging curve features. The remaining life prediction module is used to fit the battery health status and obtain the remaining life based on the fitted curve.

[0022] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor performs the above-described battery remaining life prediction method.

[0023] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for predicting remaining battery life.

[0024] The beneficial effects of this invention are:

[0025] 1. This invention predicts the battery health status during the effective charging period using a battery health status prediction model and obtains the remaining battery life based on the prediction results. This method can extract rich state information contained in the complete time-series curve. Compared with the prior art, which can only issue warnings when a fault is about to occur or has already occurred, this invention can achieve real-time prediction of battery health status. At the same time, this invention can effectively remove noise and improve data quality by cleaning and filtering the samples multiple times, thereby improving the predictive ability of the model.

[0026] 2. This invention constructs a battery health status prediction model using convolutional neural networks and long short-term memory networks. Combining the powerful local feature extraction capabilities of convolutional neural networks with the time trend analysis capabilities of long short-term memory networks, it can better capture degradation characteristics in the data, thus making the prediction results of battery health status and remaining lifespan more accurate. This method can effectively improve the early warning accuracy of battery health status and can accurately predict the remaining lifespan of batteries in most cases. The model has strong robustness and also has high application value in battery asset management fields such as battery swapping cabinets and electric vehicle fleets. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention.

[0028] Figure 2 This is a diagram of the battery health status prediction model in this invention.

[0029] Figure 3 This is a comparison chart showing the predicted and actual values ​​of battery health status using a long short-term memory network.

[0030] Figure 4 This is a comparison chart showing the results of battery health status prediction using the battery health status prediction model in this invention and the results of battery health status prediction using a convolutional neural network.

[0031] Figure 5 This is a schematic diagram illustrating the predictive performance of the battery health status prediction model in this invention.

[0032] Figure 6 This is a comparison of the R² values ​​of the battery health status prediction model and the long short-term memory network in this invention under multiple samples.

[0033] Figure 7This is a comparison chart showing the battery health state prediction performance of the battery health state prediction model in this invention with that of convolutional neural networks and long short-term memory networks.

[0034] Figure 8 This is an R² comparison chart of the battery health state prediction model in this invention and existing battery health state prediction models.

[0035] Figure 9 This is an R² comparison chart of battery health status prediction using data before and after preprocessing. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings.

[0037] A battery remaining life prediction method based on hybrid deep learning is proposed. The battery remaining life prediction system includes a data acquisition module, a preprocessing module, a feature extraction module, a battery health status prediction module, and a remaining life prediction module. The data acquisition module acquires multi-dimensional time-series data of the battery during charging and discharging. The feature extraction module extracts effective charging intervals from the multi-dimensional time-series data and obtains battery health status samples and charging curve features corresponding to these intervals. The preprocessing module transforms, cleans, and normalizes the multi-dimensional time-series data. The battery health status prediction module predicts the battery health status based on the charging curve features. The remaining life prediction module fits the battery health status and obtains the remaining life based on the fitted curve.

[0038] like Figure 1 As shown, the battery remaining life prediction method includes the following steps:

[0039] Step 1: Data Collection

[0040] A dataset is constructed using multidimensional time-series data of the battery. The multidimensional time-series data of the battery includes timestamp, device ID, total voltage (V), charging current (A), discharging current (A), motherboard temperature (°C), remaining capacity percentage (SOC), remaining capacity (Ah), nominal capacity (Ah), and abnormal alarms.

[0041] In this embodiment, the data in the dataset comes from the Big Data System Intelligent Cloud BMS.

[0042] Step 2: Data Preprocessing

[0043] First, the format of the multidimensional time-series data in the dataset is standardized by mapping Chinese field names to English and translating Chinese descriptions in abnormal alarm messages into standardized English codes to facilitate program processing. Then, missing or outlier values ​​in the multidimensional time-series data are filled or removed. Finally, charging and discharging currents are merged into a unified "current" feature, with charging being positive and discharging negative, and all voltage units are standardized to volts (V). The dataset is then divided into independent subsets based on the battery's device ID for subsequent personalized analysis.

[0044] Step 3: Obtain battery health status tags and charging curve characteristics

[0045] 3-1. Identify the effective charging range

[0046] Different effective charging intervals are identified from the multidimensional time-series data of the battery. The effective charging intervals meet the following conditions:

[0047] (1) The charging current corresponding to the start of the effective charging range is greater than the preset current threshold for the first time. Furthermore, the charging current at the end of the effective charging range falls back to the current threshold. the following.

[0048] In this embodiment, the current threshold Set it to 0.1A.

[0049] (2) The number of data points within the effective charging range is not less than the preset data point threshold. This is to filter out brief charging noise.

[0050] In this embodiment, the data point threshold is set to 5.

[0051] (3) The remaining capacity percentage at the end of the effective charging zone is 100%.

[0052] 3-2. Constructing a Battery Health Status Sample

[0053] The stable point corresponding to each effective charging zone is obtained using the following method:

[0054]

[0055] in, Index for stable points; This is the endpoint index of the effective charging area; The index interval between the set stable point index and the endpoint index.

[0056] In this embodiment, the index interval Set to 2.

[0057] The battery health status value is obtained based on the remaining capacity at the stable point and the battery's nominal capacity. Its expression is:

[0058]

[0059] in, The remaining capacity at the stable point; This refers to the nominal capacity.

[0060] Battery health status value corresponding to the stable point and number of loops As a sample of battery health status ( ,SOH) is added to the dataset.

[0061] 3-3. Multi-stage cleaning of battery health status samples

[0062] Due to variations in the measurement environment, the battery health status calculated in a single instance can fluctuate. To obtain a smooth and reliable degradation trend, multi-level cleaning is required for all battery health status samples. The specific process is as follows:

[0063] (1) Group all battery health status samples by device ID and cycle number. Within each group, outlier battery health status values ​​are removed using the interquartile range (IQR) method. The average battery health status value of the remaining battery health status samples in the group is used as the standard battery health status value for that cycle number.

[0064] (2) Collect the standard battery health status values ​​of different devices at different cycles, and use the IQR method again to remove global outliers and remove those points that significantly deviate from the overall data distribution.

[0065] (3) For the remaining battery health status samples of each device, a first-order polynomial is used for fitting to calculate the degradation slope of the battery health status. If the number of remaining battery health status samples is less than the set sample threshold or the degradation slope... If the value is not less than zero, then remove the data from that device.

[0066] In this embodiment, the sample threshold is set to 5.

[0067] 3-4. Extracting charging curve features

[0068] For each battery health status sample, three sets of time-series curves are obtained: total voltage, current, and motherboard temperature within the effective charging range. To ensure that all curves have consistent lengths to meet the neural network input requirements, linear interpolation is used to resample each curve to a fixed length. (Set to 200 points), forming a dimension of The charging curve feature matrix is ​​then added to the dataset.

[0069] Step 4: Data Normalization Process

[0070] The battery health state samples and charging curve feature matrices in the dataset are subjected to min-max normalization, scaling them to the [0, 1] interval to accelerate model convergence. The formula is as follows:

[0071]

[0072] in, and These are the data before and after normalization, respectively. and These are the minimum and maximum values ​​of the data, respectively.

[0073] Step 5: Construct a battery health status prediction model

[0074] like Figure 2 As shown, the battery health state prediction model comprises a feature extraction module, a trend analysis module, and a fully connected layer connected in sequence. The feature extraction module automatically learns local features related to battery aging from the original feature matrix of a single charge. The trend analysis module analyzes the evolution of the local features extracted by the feature extraction module over time. The fully connected layer maps the output of the trend analysis module to the final battery health state value.

[0075] The feature extraction module employs a convolutional neural network (CNN), which consists of two concatenated convolutional blocks and a flattening layer. The two convolutional blocks have identical structures, each including a cascaded convolutional layer (Conv1D), a ReLU activation function, and a max-pooling layer (MaxPool1D). The mathematical expression for the convolutional layer is:

[0076]

[0077] in, This is the k-th element of the output feature map of the convolutional layer; For activation functions; For convolution kernel; Input data; is the bias; m is the kernel size.

[0078] The mathematical expression for the ReLU activation function is:

[0079]

[0080] in, is the output of the ReLU activation function; x is the input of the ReLU activation function.

[0081] The trend analysis module employs a Long Short-Term Memory (LSTM) network. LSTM networks control the flow of information through their internal input, forget, and output gates, effectively capturing long-term dependencies. Its core update formula is as follows:

[0082] (1) Gate of Oblivion:

[0083]

[0084] (2) Input gate:

[0085]

[0086] (3) Candidate cell status:

[0087]

[0088] (4) Cell state renewal:

[0089]

[0090] (5) Output gate:

[0091]

[0092] (6) Hidden state update:

[0093]

[0094] in, It is the Sigmoid activation function. and These are the weight matrix and the bias vector, respectively.

[0095] Step Six: Model Training

[0096] The battery health state prediction model was trained using a dataset. During training, a sliding window method was employed, using the charging curve features corresponding to historical cycle counts as input to the model and the battery health state value corresponding to the current cycle count as the label. The root mean square error (RMSE) was used as the loss function. The Adam optimizer was used to minimize the loss function through backpropagation to update the model parameters. The loss function L of the battery health state prediction model is expressed as:

[0097]

[0098] in, This represents the actual battery health status value. is the battery health state value predicted by the model; n is the number of samples.

[0099] In this embodiment, the charging curve features corresponding to the past 5 cycles are used to predict the battery health status corresponding to the 6th cycle.

[0100] Step 7: Predict remaining useful life

[0101] Predict the latest cycle number using a battery health status prediction model. Corresponding battery health status value Because the degradation process of battery health status typically exhibits a non-linear characteristic, with slower degradation in the early stages and accelerated degradation in later stages, a second-order polynomial function is used to accurately describe this pattern. Fitted battery health status values and historical loop count ( The corresponding historical battery health status value is in the following form:

[0102]

[0103] in, The coefficients are to be fitted.

[0104] Choosing a second-order polynomial is a common approach that strikes a good balance between model complexity and fitting accuracy.

[0105] The coefficients are determined using the least squares method. The optimal value minimizes the total squared error between the fitted curve and all historical battery health state values, i.e.:

[0106]

[0107] By solving the above optimization problem, the optimal decay curve function is finally obtained. ;in, These are the optimal coefficients.

[0108] Set the battery health status threshold when the battery reaches the end of its lifespan. Set the battery health status threshold. Substitute into the decay curve function to obtain the number of cycles at the end of the lifespan. Based on the number of cycles at the end of lifespan And the latest cycle count Get remaining service life This allows for the prediction of the remaining battery life.

[0109] In some embodiments, since the end of the lifespan of a lithium battery is defined as the degradation of its maximum usable capacity to 80% of its nominal capacity, the battery health state threshold is set to 80, i.e. .

[0110] Step 8: Model Evaluation

[0111] Using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination respectively The predicted battery health status value in this invention is evaluated.

[0112] Mean Absolute Error The expression is:

[0113]

[0114] in, The actual value; is the predicted value; n is the number of samples.

[0115] Root mean square error The expression is:

[0116]

[0117] Coefficient of determination The expression is:

[0118]

[0119] A hybrid deep learning model is employed to analyze the data, thereby obtaining battery health status prediction evaluation metrics for different datasets and individual model scenarios. This invention designs comparative experiments to evaluate the accuracy of the proposed CNN-LSTM hybrid model based on charging curve features and time series data. The evaluation metrics are mainly compared with other commonly used time series prediction models in the field, including pure Long Short-Term Memory (LSTM) models, Transformer models, Gated Recurrent Unit (GRU) models, and traditional machine learning models based on feature engineering. For the pure LSTM model, two hidden layers with 50 units each are used; for the Transformer model, two encoding layers and four attention heads are used, with a model dimension of 64; for the GRU model, two hidden layers with 50 units each are used; for feature engineering combined with XGBoost, key statistical features of the charging curve are manually extracted, and the XGBoost tree size is set to 200 trees, a maximum depth of 5, and a learning rate of 0.1. All models are evaluated using the same test dataset, and the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are calculated.

[0120] The prediction results of the battery health status value using the LSTM model are as follows: Figure 3 As shown. From Figure 3As can be seen, the overall trend of the predicted battery health status values ​​is basically consistent with the trend of the actual values, indicating that the model has learned the changing trend of battery health status quite well. The predicted curve is also relatively smooth, and can accurately reflect the overall changing trend of battery health status values. However, there are some deviations from the actual values ​​at individual abrupt change points, showing some slight delayed response or underestimation. Overall, the model has high prediction accuracy and good fitting effect in most test samples, indicating that the LSTM model has strong stability and robustness in battery health status prediction, and can achieve a high-accuracy estimate of battery health status.

[0121] from Figure 4 As can be seen, there are significant differences in the performance of the battery health state prediction model (CNN-LSTM model) and the CNN model in predicting battery health state. The CNN-LSTM model has a more concentrated distribution of prediction points, with most closely following the ideal fitting line, indicating that its prediction results are highly consistent with the actual battery health state values, demonstrating high accuracy and stability. In contrast, the CNN model's prediction points are more scattered, with some samples deviating significantly from the ideal line, indicating limitations in learning complex features and temporal dependencies. Overall, the CNN-LSTM model exhibits superior fitting ability and generalization performance compared to the traditional CNN model, more accurately capturing the changing patterns of battery health state, achieving higher prediction accuracy, and demonstrating strong robustness and reliability.

[0122] from Figure 5 As can be seen, the CNN-LSTM model demonstrates high accuracy and stability in the battery health state prediction task. The model's mean absolute error (MAE) is 0.110%, and its root mean square error (RMSE) is 0.135%, indicating a small deviation between the predicted and actual values, and the model exhibits good fit. Furthermore, its coefficient of determination (R²) reaches 0.895, indicating good correlation and generalization ability. Overall, the CNN-LSTM model shows significant advantages in the joint learning of multidimensional and temporal features, validating its effectiveness and reliability in the battery health state prediction task.

[0123] from Figure 6As can be seen, there are significant differences in the R² scores of the LSTM and CNN-LSTM models across different devices. Overall, the CNN-LSTM model consistently achieves higher R² values ​​than the LSTM model across all devices, indicating that by combining convolutional feature extraction with temporal modeling, CNN-LSTM can more effectively capture the spatiotemporal features of the data, thereby improving the prediction accuracy of battery health status. The performance of the LSTM model fluctuates slightly across different devices, with overall R² values ​​concentrated between 0.75 and 0.79; while the R² values ​​of CNN-LSTM consistently remain above 0.82, reaching a maximum of nearly 0.85, demonstrating more stable generalization ability. In summary, CNN-LSTM exhibits superior prediction performance across various devices, demonstrating its significant advantages and robustness in handling this type of time series task.

[0124] from Figure 7 As can be seen, the three different models (LSTM, CNN, and CNN-LSTM) show some differences in their performance in predicting battery health. Specifically, the CNN-LSTM model performs best in terms of R² score, reaching 0.837, significantly better than LSTM's 0.747 and CNN's 0.802, demonstrating its stronger ability to fit data and capture time-series relationships. Furthermore, the CNN-LSTM model also performs well in MAE and RMSE metrics. Although the LSTM model's RMSE is slightly lower than CNN's, overall, CNN-LSTM leads in all metrics. In summary, CNN-LSTM demonstrates the best predictive ability in the battery health prediction task.

[0125] from Figure 8As can be seen, the R² scores of the four battery health prediction models (CNN-LSTM, Transformer, GRU, and Feature Eng. + ML) also show slight differences across different devices. The CNN-LSTM model exhibits the best prediction accuracy on most devices, with its R² value approaching or exceeding 0.85 on multiple devices, reaching as high as close to 0.90. In contrast, the Transformer and GRU models perform similarly, with R² values ​​generally remaining around 0.80, showing relatively stable performance, but overall lagging behind CNN-LSTM. Furthermore, the Feature Eng. + ML model performs relatively weakly; although its R² value approaches 0.80 on some devices, it is generally more scattered and cannot consistently outperform other models. In summary, the CNN-LSTM model has an advantage on most devices, and its stability and high prediction accuracy make it the most effective model for this task. While the Transformer and GRU also perform well on some devices, overall, the advantage of CNN-LSTM is more pronounced.

[0126] from Figure 9 As can be seen, models based on preprocessed data show a significant improvement in R² score compared to models based on raw data. Models with preprocessed data exhibit higher R² values ​​on most devices, especially on s0, s5, and s8 devices, where R² scores approach 0.90, and in some cases even exceed 0.90, indicating that preprocessed data can significantly improve prediction accuracy. In contrast, models based on raw data generally have lower R² scores, hovering around 0.80, performing worse than models based on preprocessed data. This demonstrates that data preprocessing can effectively remove noise and improve data quality, thereby enhancing the predictive power of the model.

[0127] As can be seen, this invention is practiced and described through some examples. Those skilled in the art will recognize that appropriate changes or equivalent substitutions can be made to these embodiments and features without departing from the scope and spirit of this invention. Furthermore, under the teachings of this invention, these features and embodiments can be appropriately modified to adapt to specific situations and corresponding materials, so as not to depart from the scope and spirit of this invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A battery remaining life prediction method based on hybrid deep learning, characterized in that: The method includes: Acquire multidimensional time-series data of the battery; identify different effective charging intervals from the multidimensional time-series data of the battery and mark the number of cycles corresponding to each effective charging interval; obtain the battery health status value corresponding to each effective charging interval; construct a battery health status sample based on the battery health status value and the number of cycles, and extract charging curve features from the multidimensional time-series data based on the battery health status sample; A battery health state prediction model is constructed. The battery health state prediction model includes a feature extraction module, a trend analysis module, and a fully connected layer connected in sequence. The feature extraction module is used to extract local features from the charging curve features. The trend analysis module is used to analyze the evolution of local features over time. The fully connected layer is used to map the output of the trend analysis module to the final battery health state value. The battery health status prediction model is trained using battery health status samples and charging curve features. The trained battery health status prediction model is then used to predict the battery health status value corresponding to the latest cycle count of the tested battery. The battery health status values ​​corresponding to the historical cycle count and the latest cycle count are fitted, and the life end cycle count is obtained using the fitted curve. The difference between the life end cycle count and the latest cycle count is used as the remaining lifespan to complete the prediction of the remaining battery lifespan.

2. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: Before extracting charging curve features, the battery health status samples undergo multi-level cleaning, the specific process of which is as follows: Remove abnormal battery health status values ​​from the battery health status samples; perform first-order polynomial fitting on the remaining battery health status samples of different batteries respectively, and obtain the slope of the fitted line. If the number of remaining battery health status samples is less than the sample threshold or the slope is not less than zero, then remove all battery health status samples corresponding to that battery.

3. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: The effective charging range meets the following conditions: The starting point of the effective charging interval is greater than the current threshold, and the ending point of the effective charging interval is less than the current threshold; the number of data points within the effective charging interval is not less than the data point threshold; and the remaining capacity percentage corresponding to the ending point of the effective charging interval is 100%.

4. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: The battery health status value corresponding to the effective charging interval is the battery health status value at the stable point of the effective charging interval; the stable point is obtained by adding a set index interval to the end index of the effective charging interval to obtain the stable point index.

5. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: The fitted curve is obtained by fitting the battery health status value using a second-order polynomial function.

6. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: The training process of the battery health status prediction model is as follows: the charging curve features corresponding to the historical cycle number are used as the input of the battery health status prediction model, the battery health status value corresponding to the current cycle number is used as the label to train the battery health status prediction model, and the root mean square error is used as the loss function to update the model parameters.

7. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: The multidimensional time-series data includes total voltage, current, motherboard temperature, remaining capacity percentage, remaining capacity, and nominal capacity; the charging curve features are composed of total voltage, current, and motherboard temperature.

8. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: The battery health status value is the ratio of remaining capacity to nominal capacity.

9. The battery remaining life prediction method based on hybrid deep learning according to claim 1, characterized in that: The feature extraction module employs a convolutional neural network, which includes multiple convolutional blocks connected in sequence and a flattening layer; the trend analysis module employs a long short-term memory network.

10. A battery remaining life prediction system based on hybrid deep learning, characterized in that: This system is used to execute a battery remaining life prediction method based on hybrid deep learning as described in claim 1. The battery remaining life prediction system includes a data acquisition module, a feature extraction module, a battery health status prediction module, and a remaining life prediction module. The data acquisition module acquires multi-dimensional time-series data of the battery during charging and discharging. The feature extraction module extracts effective charging intervals from the multi-dimensional time-series data and acquires battery health status samples and charging curve features corresponding to the effective charging intervals. The battery health status prediction module predicts the battery health status based on the charging curve features. The remaining life prediction module fits the battery health status and obtains the remaining life based on the fitted curve.

Images