Early prediction method for battery life based on pre-classification and LSTM-CNN

Abstract

A battery life early prediction method based on pre-classification and LSTM-CNN comprises the following steps: extracting features by using capacity difference data in a discharge process, determining a classification interval of a sample through an unsupervised learning clustering method, dividing the sample into a long life group and a short life group according to the actual life distribution condition of the battery, according to the method, a brand-new superposition model of LSTM and CNN networks suitable for battery sample classification and life prediction is constructed and trained, new samples are classified through an LSTM-CNN classification model, different life prediction models are matched, and then the life of the battery is predicted through the LSTM-CNN prediction model, so that early prediction of the life of the battery by using a small amount of data is realized, and the prediction efficiency of the life of the battery is improved. The model structure is simple, and the prediction accuracy and efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery life prediction, and in particular to an early prediction method for battery life based on pre-classification and LSTM-CNN. Background Art

[0002] As an energy storage device, batteries have extensive and important applications in new energy vehicles, smart grids, portable electronic devices and other fields. However, during the life cycle of the battery, the electrical performance will gradually decline due to material aging, interface defects, irreversible internal reactions and other reasons. Typical performance degradation is manifested as capacity decay and power decline, that is, a reduction in battery life. Accurately predicting the remaining life of the battery not only helps users to rationally plan equipment replacement cycles and reduce operating costs, but also can provide early warning of potential safety risks and improve the overall stability of the system. Traditional battery life prediction methods mostly rely on theoretical models and empirical formulas. The assumptions are usually very strict and it is difficult to accurately reflect the battery degradation mechanism under actual working conditions.

[0003] In recent years, with the development of big data and artificial intelligence technology, more and more research has turned to using real-time monitoring data during battery operation to predict battery life. Despite this, there are still many challenges in predicting battery life through data-driven methods. For example, when using the LSTM model, although the model performs well in capturing long-term dependencies in time series data, it is not obvious in extracting local features. This is because the design focus of LSTM is to use its internal memory units and gating mechanisms to maintain long-term memory of previous information, which makes it better at identifying and predicting important events or patterns with long time intervals in sequence data. The autoregressive model (AR model) has been applied to the estimation of battery remaining useful life (RUL) by using linear combinations of time series data for prediction. Its advantages are simple model, small amount of calculation, and few samples required, but it also has limited generalization ability and difficulty in capturing nonlinearity and uncertainty in the battery aging process. The application of convolutional neural network (CNN) in battery aging modeling is mainly reflected in its ability to automatically extract useful features from high-dimensional data of batteries. Although it has limitations in capturing the time series characteristics of battery aging, it can effectively improve the accuracy of battery remaining useful life (RUL) prediction by combining it with models such as LSTM. At present, the model combining convolutional neural network (CNN) and long short-term memory network (LSTM) has shown good accuracy in battery life prediction, but there is still a lot of room for optimization in terms of model structure and feature extraction methods to reduce model complexity and improve model prediction efficiency and practicality. The patent with publication number CN116879753B discloses a battery life prediction method based on big data, and proposes a multi-dimensional feature battery life prediction method combining Bi-LSTM and XGboost models, which makes up for the shortcomings of a single model and improves the accuracy and robustness of the prediction results. However, due to the high time complexity and space complexity of the XGboost model, the overall computing power of the prediction model is high and the implementation is difficult. "Tang Y, Yang K, Zheng H, et al. Early prediction of lithium-ion battery lifetime via a hybrid deep learning model [J]. Measurement, 2022, 199: 111530." combines LSTM and CNN networks and proposes a PCLN model, including a data processing module, an LSTM extractor, a CNN extractor and a prediction layer. The discharge data in the early stage of battery degradation is used as the model input, and good prediction results are obtained. However, when the sample life is relatively dispersed, the prediction results for long-life samples are generally low. Therefore, it is of great significance to reduce the complexity of the fusion model, improve the accuracy of early prediction, and realize personalized prediction of samples. Summary of the invention

[0004] In order to overcome the difficulty of traditional battery life prediction methods to achieve accurate prediction of early life, in view of the limitations of existing technologies such as high data volume requirements, insufficient prediction accuracy and complex model structure, the purpose of the present invention is to propose an early prediction method for battery life based on pre-classification and LSTM-CNN, which utilizes capacity difference data in the discharge process to extract features, determines the classification interval of samples through an unsupervised learning clustering method, and divides them into long-life group and short-life group according to the actual battery life distribution. A new superposition model of LSTM and CNN networks suitable for battery sample classification and life prediction is constructed and trained. First, new samples are classified through the LSTM-CNN classification model and matched with different life prediction models, and then the battery life is predicted through the LSTM-CNN prediction model, which realizes early prediction of battery life using a small amount of data, streamlines the model structure, and significantly improves the prediction accuracy and efficiency.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A battery life early prediction method based on pre-classification and LSTM-CNN includes the following steps:

[0007] S1: First, extract the QV data in each cycle discharge process from the collected battery charge and discharge cycle data, and perform preprocessing operations on the extracted data;

[0008] S2: Based on data preprocessing, feature extraction and feature screening are performed. The selected features will be used as input for subsequent battery grade classification and life prediction models;

[0009] S3: Using the Mean Shift unsupervised learning clustering method, the battery data set is divided into two distinct life span intervals, namely the long life span group and the short life span group, based on the actual distribution of the life span of the battery data set. The clustering results will serve as the basis for sample grouping to guide subsequent battery classification and life span prediction;

[0010] S4: Build an LSTM-CNN classification model framework, use the features extracted from the training set data as input data, and the grouping results shown in step S3 as label data to train the battery life classification model;

[0011] S5: Building an LSTM-CNN regression model framework, and using the characteristic data of the long-life group and the short-life group batteries in the training set to train the corresponding battery life prediction model according to the clustering results described in step S3;

[0012] S6: Use the test set data to verify and evaluate the model prediction effect.

[0013] In the process of preprocessing the data in step S1, redundant invalid data outside the discharge process is first identified and removed, and only valid data reflecting the changes in the intrinsic characteristics of the battery during the discharge process is retained. On this basis, the first-order difference method is used to detect and identify possible abnormal discharge cycles and abnormal data points in individual cycles, and then these abnormal values ​​are reasonably repaired with the help of a linear interpolation method. Finally, the voltage data is standardized to unify the voltage sampling points and sampling intervals, so that the data between different batteries are comparable.

[0014] In the feature extraction and screening process of step S2, for each round of charge and discharge cycle, first calculate the Q-Q1 data in each cycle, and extract seven basic characteristic parameters, including variance, minimum value, maximum value, average value, kurtosis, skewness and capacity difference corresponding to the voltage of 2V; then, by calculating three correlation coefficients between the seven characteristic parameters and the total capacity of the battery, the three correlation coefficients of the total capacity of the battery are: Pearson correlation coefficient, Spearman rank correlation coefficient and Kendall rank correlation coefficient, quantitatively analyze the correlation between the seven characteristic parameters and the battery capacity attenuation, and then screen out four characteristic indicators closely related to the battery life for subsequent model training and prediction input, and the screening principle is: select characteristic items with correlation coefficients greater than 0.9.

[0015] In the life interval division process of step S3, the Mean shift clustering algorithm combined with kernel density estimation is used to perform unsupervised learning clustering on the original battery life data. The specific method is: first, a Gaussian kernel function is used to estimate the probability density of the sample life distribution in the data set, and the Mean shift algorithm is used to calculate the drift vector of the sample point, and each sample point is moved in the direction of increasing probability density. The probability density calculation and sample point movement are repeated until the maximum number of iterations is reached. Finally, the battery samples will converge to two points with the largest local density, i.e., the clustering centers. According to the clustering of the long life group and the short life group, the classification intervals, i.e., the long life group and the short life group, are obtained as the label basis for the subsequent classification model.

[0016] The step S4 is specifically as follows:

[0017] A composite deep learning classification model structure integrating a long short-term memory network (LSTM) and a convolutional neural network (CNN) is constructed, including two LSTM layers, a Dropout layer, five one-dimensional convolutional layers, a maximum pooling layer, a flattening layer, a fully connected layer, and an output layer. The activation function of the output layer is specified as a Softmax function to convert the output into a probability distribution form. In the model training process, the training data is first labeled using the clustering result of step S3, and the feature data of the discharge data of the first 10%-15% cycles of the training set is used as the model input to train a classification model, namely, an LSTM-CNN classification model. The model classifies the battery into a corresponding life span according to its current health status, so that different life prediction models can be used to predict the remaining life in the future.

[0018] The step S5 is specifically as follows:

[0019] First, build an LSTM-CNN regression model framework with the same hierarchical structure as the LSTM-CNN classification model described in step S4, which consists of two LSTM layers, five one-dimensional convolutional layers, a fully connected layer and an output layer. The output layer uses a Linear activation function, and the output result is the life value predicted by the model. During the LSTM-CNN regression model training process, according to the clustering results, the long-life group and the short-life group samples are trained separately, and the feature data of the first 10%-15% cycles of the long-life group and the short-life group samples are input respectively, and different model parameters are assigned to capture the battery degradation law in different life intervals, and accurate life prediction of different batteries is performed. After the training is completed, the long-life LSTM-CNN prediction model and the short-life LSTM-CNN prediction model are obtained.

[0020] The step S6 is specifically:

[0021] First, the feature data of the battery samples in the test set are input into the LSTM-CNN classification model for pre-classification judgment to obtain the classification results of the battery samples. Then, based on the pre-classification results, the corresponding LSTM-CNN life prediction model is called for the long-life group / short-life group samples, and the feature data is input to predict the remaining battery life. The predicted value uses the root mean square error and mean absolute percentage error as evaluation indicators to compare the remaining life predicted by the model with the actual battery life.

[0022] The present invention also includes a battery life early prediction system based on pre-classification and LSTM-CNN, including a processing instruction module for executing each step in the above-mentioned battery life early prediction method based on pre-classification and LSTM-CNN.

[0023] A battery life early prediction device based on pre-classification and LSTM-CNN, comprising:

[0024] Memory: a computer-readable device storing the computer program of the battery life early prediction method based on pre-classification and LSTM-CNN;

[0025] Processor: used to implement the battery life early prediction method based on pre-classification and LSTM-CNN when executing the computer program.

[0026] A computer-readable storage medium stores a computer program, which, when executed by a processor, can implement a battery life early prediction method based on pre-classification and LSTM-CNN.

[0027] Compared with the prior art, the advantages of the present invention are:

[0028] (1) Since step S2 adopts a feature extraction method based on the loss mechanism of negative active materials in lithium-ion batteries, Q-Q1 is calculated and the multi-dimensional strongly correlated features extracted on the basis of calculating the capacity difference in the discharge process are used as model input through statistical feature extraction and feature screening methods based on correlation coefficients, the early aging characteristics of the battery are identified, the feature space is optimized, the risk of dimensional disaster is reduced, the data demand is reduced to achieve early life prediction, and the prediction accuracy of the data-driven model is improved.

[0029] (2) Since step S3 adopts the unsupervised learning method of mean shift clustering, the classification boundary is defined according to the distribution of the data set itself, which provides a more flexible and reasonable classification rule for the subsequent step S4 to perform sample pre-classification. It has the advantage that the classification boundary is consistent with the actual engineering situation and can improve the accuracy of the prediction results.

[0030] (3) Since steps S4 and S5 adopt the method of constructing an LSTM-CNN fusion model framework, they can effectively capture the time series characteristics and complex patterns of battery cycle data, and combine the learning ability of LSTM units for long-term dependencies with the high efficiency of CNN in feature extraction, which significantly improves the accuracy and robustness of classification models and prediction models.

[0031] (4) Since step S6 adopts a personalized prediction method combined with pre-classification, the long-life LSTM-CNN prediction model or the short-life LSTM-CNN prediction model is matched to the long-life and short-life samples, which has the advantages of refined, personalized and high-precision battery life prediction.

[0032] In summary, the present invention adopts a feature extraction method based on the loss mechanism of negative electrode active materials of lithium-ion batteries, combines unsupervised learning clustering and pre-classification strategies, constructs an LSTM-CNN classification model and an LSTM-CNN prediction model, and realizes early personalized prediction of battery life. It has the advantages of streamlined model structure, small data requirement, strong generalization performance, and high prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the overall framework of the life prediction method of the present invention.

[0034] Figure 2 This is the Mean-shift clustering effect diagram.

[0035] Figure 3 Schematic diagram of the LSTM-CNN fusion model.

[0036] Figure 4 is the confusion matrix of the classification model.

[0037] Figure 5 It is a scatter plot of the predicted results and the actual test values. DETAILED DESCRIPTION

[0038] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0039] The present invention proposes a battery life prediction method based on pre-classification and deep learning. On the basis of constructing a new LSTM-CNN classification model and LSTM-CNN regression model suitable for battery life prediction, the battery life is accurately predicted by combining data preprocessing, feature extraction and screening, and Mean Shift clustering. Figure 1 As shown in the figure, early, accurate, efficient and reliable prediction can be achieved when the amount of data is not large. The following is a specific implementation method.

[0040] A battery life early prediction method based on pre-classification and LSTM-CNN includes the following steps:

[0041] S1: First, extract the QV data in each cycle discharge process from the collected battery charge and discharge cycle data, and perform preprocessing operations on the extracted data; the specific steps are as follows:

[0042] S11: Collect capacity-voltage (QV) data of a set of batteries during all charge and discharge cycles in the complete life cycle, ensuring that the data contains enough information to fully characterize the battery performance;

[0043] The complete life cycle refers to all charge and discharge cycles before the battery capacity degrades to 80% of the rated capacity;

[0044] S12: Cleaning the collected data to remove redundant data in the non-discharging stage, such as the charging stage and data under other atypical working conditions before and after the discharging process, to ensure that only valid data reflecting the battery discharge characteristics is retained;

[0045] S13: Standardize the voltage data, unify the voltage sampling points and sampling intervals, and ensure that the data of different batteries under the same sampling conditions are comparable;

[0046] S14: Apply the first-order difference method to analyze the effective (QV) data to find and identify possible abnormal discharge cycles (e.g., abnormally fast capacity drop or recovery) and abnormal data points within a single cycle;

[0047] Furthermore, the first-order difference calculation method is ΔX i =X i+1 -X i , the trend component in the data can be eliminated through the first-order difference. If outlier data appears in the calculation result, it means that the original sequence has a mutation or abnormality, and the cycle with abnormal data is marked accordingly.

[0048] S14: For abnormal data points found, linear interpolation is performed to reasonably repair the capacity data of the same voltage in the previous and next cycles to ensure the consistency and continuity of the data sequence

[0049] S15: Finally, the voltage data is standardized to unify the voltage sampling points and sampling intervals so that the data between different batteries are comparable.

[0050] S2: Based on data preprocessing, feature extraction and feature screening are performed. The selected features will be used as input for subsequent battery grade classification and life prediction models. The specific steps are as follows:

[0051] S21: For each battery cycle, calculate the Q-Q1 data (i.e., the capacity difference between each sampling point in the current cycle and the first cycle) to reflect the change in battery performance between cycles. First, calculate the seven basic characteristic parameters of each cycle: variance, minimum, maximum, average, kurtosis, skewness, and capacity difference when the voltage is 2V;

[0052] S22: calculating three correlation coefficients between the above seven characteristic parameters and the total battery capacity, the three correlation coefficients of the total battery capacity are: Pearson correlation coefficient, Spearman rank correlation coefficient and Kendall rank correlation coefficient, so as to quantify the correlation between the characteristic parameters and the battery capacity attenuation;

[0053] Furthermore, the Pearson correlation coefficient is used to measure the linear correlation between the characteristic parameter and the capacity, and the Spearman rank correlation coefficient and the Kendall rank correlation coefficient are used to measure the rank correlation between the characteristic parameter and the capacity, regardless of the actual numerical value of the variable, and the calculation methods are:

[0054]

[0055]

[0056]

[0057] Where n is the number of samples, d is the rank difference of the variable, C and D are the number of consistent and inconsistent pairs, respectively;

[0058] S23: Quantitatively analyze the correlation between the seven characteristic parameters and the battery capacity decay, and then select four characteristic indicators closely related to the battery life for input of subsequent model training and prediction. The screening principle is: select characteristic items with correlation coefficients greater than 0.9.

[0059] S3: In order to refine the battery life prediction problem, the Mean Shift unsupervised learning clustering method is used to divide the battery data set into two distinct life intervals, namely the long life group and the short life group, according to the actual distribution of the battery life. The clustering results will serve as the basis for sample grouping to guide subsequent battery classification and life prediction; the specific method is as follows:

[0060] S31: In order to refine the battery life prediction problem, the battery life data is estimated by kernel density, and the original battery life data is clustered by unsupervised learning using the Mean shift clustering algorithm combined with kernel density estimation. First, the Gaussian kernel function is selected to depict the overall shape of the data distribution by converting the local density around the data point into a probability density function. The calculation method of the kernel function and the probability density estimate of each data point is as follows:

[0061]

[0062]

[0063] Where n is the total number of data points, h is the bandwidth, and d is the data dimension;

[0064] S32: Based on the kernel density estimation graph, the Mean Shift clustering algorithm is used for clustering. First, each data point is initialized as a candidate cluster center. For each data point, its gradient ascending direction under the kernel density function is calculated, that is, it moves in the direction of higher density. After the data point moves, the position of the data point is updated. This process is repeated until the position of the data point converges. Each convergence point represents the center of a cluster, such as Figure 2 As shown, all batteries are automatically divided into two life spans with obvious distinction: long life group and short life group;

[0065] S33: Combined with the clustering results, a category label is assigned to each battery sample, indicating whether it belongs to the long life or short life interval, which serves as the label basis for the subsequent classification model. This pre-classification strategy can effectively guide the subsequent model training and prediction process.

[0066] S4: Build an LSTM-CNN classification model framework, use the features extracted from the training set data as input data, and the grouping results shown in step S3 as label data to train the battery life classification model; the specific method is as follows:

[0067] S41: Use the keras framework to build a composite deep learning classification model structure that integrates long short-term memory network LSTM and convolutional neural network CNN, such as Figure 3 As shown, including:

[0068] Two LSTM layers are used to capture long-distance dependencies in the training data and improve the model's ability to handle time-order features. The data input dimension is (number of loops, number of features) and the number of hidden units is 64.

[0069] A Dropout layer: discard a certain proportion of neurons to prevent overfitting. In this example, the discard ratio is 0.3;

[0070] Five one-dimensional convolutional layers (Conv1D): the first layer has 32 convolution kernels, a convolution kernel width of 16, and uses the Relu activation function. The last four layers have 64 convolution kernels, a convolution kernel width of 4, and still use the Relu activation function.

[0071] A maximum pooling layer: the pooling window size is 8, which effectively reduces data complexity and retains local features;

[0072] A flattening layer: flattens the multi-dimensional features into one-dimensional data to facilitate processing by the fully connected layer;

[0073] A fully connected layer: uses 100,000 neurons to extract more complex and abstract features;

[0074] An output layer: Using the Softmax activation function, the model outputs the probability distribution of each category (long life or short life), realizing the classification of battery life range;

[0075] S42: The preprocessed and feature-extracted data is used as the model input, the clustering results are used as the label data, and the labels are one-hot encoded. The data is then randomly divided into a training set and a test set, and the proportion of the test set is set to 20%;

[0076] S43: Compile and train the model. Set the loss function to the cross entropy loss function, use the Adam optimizer to automatically adjust the learning rate, update the model weights, set the model training batch size to 16, the model iteration number to 500, and input the training set data to train the model;

[0077] S44: Input the test set data into the trained model to obtain the predicted classification results, such as Figure 4 As shown, the calculated prediction accuracy is 88.89%, and finally the prediction model is output and saved.

[0078] S5: Build an LSTM-CNN regression model framework, and use the characteristic data of the long-life group and the short-life group batteries in the training set to train the corresponding battery life prediction model according to the clustering results described in step S3; the specific method is as follows:

[0079] S51: Similarly, two LSTM-CNN regression models with similar structures to the classification model described in step S41 are constructed. The structures of the layers before the output layer are the same, but the activation function of the output layer is changed to a linear activation function to directly output the remaining life prediction value. The two models are used for life prediction of long-life samples and short-life samples respectively.

[0080] S52: Perform model compilation and model training. Set the loss function to the root mean square error loss function, use the Adam optimizer to automatically adjust the learning rate, set the model training batch size to 4, and the number of model iterations to 1000. According to the life span range shown in the clustering results, the training set data is divided into long-life and short-life parts. The characteristic data of long-life and short-life batteries are used to train two independent life prediction models, and the model parameters are optimized according to the characteristics of each sample to more accurately capture the battery degradation law in different life spans;

[0081] S6: For new samples, life prediction is performed according to steps S1-S5 to obtain the prediction result of battery life. The prediction effect of the model is verified and evaluated using the test set data, that is, the overall model is verified and the prediction effect is evaluated. The specific method is as follows:

[0082] S61: First, the trained classification model described in step S4 is called, and the characteristic data of the test set samples are used as model input to obtain its pre-classification results (i.e., prediction of the life span), and the test set samples are divided into a long life group and a short life group;

[0083] S62: Based on the classification result, the life prediction model described in step S5 is called, and the feature data is used as input again. The long life prediction model and the short life prediction model are used to predict the remaining life of the test sample, and the predicted value of the sample battery life is output, such as Figure 5 As shown;

[0084] S63: In order to quantify the prediction performance of the model, the root mean square error and mean absolute percentage error are used as evaluation indicators to compare the remaining life predicted by the model with the actual battery life to verify the prediction performance and accuracy of the model.

[0085] The calculation method of the evaluation index is as follows:

[0086]

[0087]

[0088] According to the prediction results and the actual data of battery life, the root mean square error is calculated to be 150.94 and the average percentage error is 12.5%. In particular, the prediction effect for the short-life group is better, with a root mean square error of 75.79 and an average percentage error of 10.61%, indicating that this method can effectively use a small amount of data to achieve early prediction of battery life and can obtain a higher prediction accuracy, especially for short-life battery samples. This method can provide more accurate prediction results. In general, this method constructs a simpler LSTM-CNN composite model, combined with a pre-classification strategy, which can perform early personalized predictions for long-life and short-life batteries with a small amount of data, improving the prediction accuracy and robustness.

[0089] The present invention also includes a battery life early prediction system based on pre-classification and LSTM-CNN, including a processing instruction module for executing each step in the above-mentioned battery life early prediction method based on pre-classification and LSTM-CNN.

[0090] A battery life early prediction device based on pre-classification and LSTM-CNN, comprising:

[0091] Memory: a computer-readable device storing the computer program of the battery life early prediction method based on pre-classification and LSTM-CNN;

[0092] Processor: used to implement the battery life early prediction method based on pre-classification and LSTM-CNN when executing the computer program.

[0093] A computer-readable storage medium stores a computer program, which, when executed by a processor, can implement a battery life early prediction method based on pre-classification and LSTM-CNN.

[0094] In summary, the present invention aims at the problems that the traditional battery life prediction method has low early prediction accuracy, complex model structure, large data usage, etc., and proposes a battery life early prediction method based on pre-classification and LSTM-CNN (long short-term memory network-convolutional neural network) deep learning technology. First, the effective QV data reflecting the intrinsic characteristics of the battery are extracted from the battery charge and discharge cycle data. Secondly, seven features highly correlated with the battery life are calculated, including variance, minimum value, maximum value, average value, kurtosis, skewness, and the capacity difference corresponding to the voltage of 2V for feature extraction, and the correlation coefficient is calculated for screening. On this basis, the Mean shift clustering algorithm combined with kernel density estimation is used to perform unsupervised learning clustering on the battery life data, and the battery samples are divided into two life intervals with significant distinction, namely the long life group and the short life group. Then, a composite deep learning model integrating LSTM and CNN is designed and developed, which is used for the classification of battery samples and the prediction of the specific life of the battery. Among them, the LSTM-CNN classification model is used to identify the health status of the battery and classify it into the corresponding life interval; the LSTM-CNN regression model uses independently trained life prediction models to make final life predictions for long-life and short-life batteries based on the pre-classification results. Finally, the root mean square error (RMSE) and mean absolute percentage error (MAPE) indicators are used to evaluate the prediction performance of the model. The method of the present invention can achieve early and accurate prediction of battery life by using only a small amount of limited cycle discharge data, and makes more refined predictions for long-life and short-life batteries respectively through independently trained life prediction models. In addition, the model structure is relatively simple, easy to implement and maintain, and significantly improves the efficiency, accuracy, reliability and economy of battery prediction.

Claims

1. A battery life early prediction method based on pre-classification and LSTM-CNN, characterized in that: The following steps are involved: S1: First, extract the QV data in each cycle discharge process from the collected battery charge and discharge cycle data, and perform preprocessing operations on the extracted data; S2: Based on data preprocessing, feature extraction and feature screening are performed. The selected features will be used as input for subsequent battery grade classification and life prediction models; S3: Using the Mean Shift unsupervised learning clustering method, the battery data set is divided into two distinct life span intervals, namely the long life span group and the short life span group, based on the actual distribution of the life span of the battery data set. The clustering results will serve as the basis for sample grouping to guide subsequent battery classification and life span prediction. S4: Build an LSTM-CNN classification model framework, use the features extracted from the training set data as input data, and the grouping results shown in step S3 as label data to train the battery life classification model; S5: Building an LSTM-CNN regression model framework, and using the characteristic data of the long-life group and the short-life group batteries in the training set to train the corresponding battery life prediction model according to the clustering results described in step S3; S6: Use the test set data to verify and evaluate the model prediction effect.

2. According to claim 1, a method for early prediction of battery life based on pre-classification and LSTM-CNN is characterized in that: In the process of preprocessing the data in step S1, redundant invalid data outside the discharge process is first identified and removed, and only valid data reflecting the changes in the intrinsic characteristics of the battery during the discharge process is retained. On this basis, the first-order difference method is used to detect and identify possible abnormal discharge cycles and abnormal data points in individual cycles, and then these abnormal values ​​are reasonably repaired with the help of a linear interpolation method. Finally, the voltage data is standardized to unify the voltage sampling points and sampling intervals, so that the data between different batteries are comparable.

3. The method for early prediction of battery life based on pre-classification and LSTM-CNN according to claim 1, characterized in that: In the feature extraction and screening process of step S2, for each round of charge and discharge cycle, first calculate the Q-Q1 data in each cycle, and extract seven basic characteristic parameters, including variance, minimum value, maximum value, average value, kurtosis, skewness and capacity difference corresponding to the voltage of 2V; then, by calculating three correlation coefficients between the seven characteristic parameters and the total capacity of the battery, the three correlation coefficients of the total capacity of the battery are: Pearson correlation coefficient, Spearman rank correlation coefficient and Kendall rank correlation coefficient, quantitatively analyze the correlation between the seven characteristic parameters and the battery capacity attenuation, and then screen out four characteristic indicators closely related to the battery life for subsequent model training and prediction input, and the screening principle is: select characteristic items with correlation coefficients greater than 0.

9.

4. The method for early prediction of battery life based on pre-classification and LSTM-CNN according to claim 1, characterized in that: In the life interval division process of step S3, the Mean shift clustering algorithm combined with kernel density estimation is used to perform unsupervised learning clustering on the original battery life data. The specific method is: first, a Gaussian kernel function is used to estimate the probability density of the sample life distribution in the data set, and the Mean shift algorithm is used to calculate the drift vector of the sample point, and each sample point is moved in the direction of increasing probability density. The probability density calculation and sample point movement are repeated until the maximum number of iterations is reached. Finally, the battery samples will converge to two points with the largest local density, i.e., the clustering centers. According to the clustering of the long life group and the short life group, the classification intervals, i.e., the long life group and the short life group, are obtained as the label basis for the subsequent classification model.

5. The method for early prediction of battery life based on pre-classification and LSTM-CNN according to claim 1, characterized in that: The step S4 is specifically as follows: A composite deep learning classification model structure integrating a long short-term memory network (LSTM) and a convolutional neural network (CNN) is constructed, including two LSTM layers, a Dropout layer, five one-dimensional convolutional layers, a maximum pooling layer, a flattening layer, a fully connected layer, and an output layer. The activation function of the output layer is specified as a Softmax function to convert the output into a probability distribution form. In the model training process, the training data is first labeled using the clustering result of step S3, and the feature data of the discharge data of the first 10%-15% cycles of the training set is used as the model input to train a classification model, namely, an LSTM-CNN classification model. The model classifies the battery into a corresponding life span according to its current health status, so that different life prediction models can be used to predict the remaining life in the future.

6. The method for early prediction of battery life based on pre-classification and LSTM-CNN according to claim 1, characterized in that: The step S5 is specifically as follows: First, build an LSTM-CNN regression model framework with the same hierarchical structure as the LSTM-CNN classification model described in step S4, which consists of two LSTM layers, five one-dimensional convolutional layers, a fully connected layer and an output layer. The output layer uses a Linear activation function, and the output result is the life value predicted by the model. During the LSTM-CNN regression model training process, according to the clustering results, the long-life group and the short-life group samples are trained separately, and the feature data of the first 10%-15% cycles of the long-life group and the short-life group samples are input respectively, and different model parameters are assigned to capture the battery degradation law in different life intervals, and accurate life prediction of different batteries is performed. After the training is completed, the long-life LSTM-CNN prediction model and the short-life LSTM-CNN prediction model are obtained.

7. The method for early prediction of battery life based on pre-classification and LSTM-CNN according to claim 1, characterized in that: The step S6 is specifically: First, the feature data of the battery samples in the test set are input into the LSTM-CNN classification model for pre-classification judgment to obtain the classification results of the battery samples. Then, based on the pre-classification results, the corresponding LSTM-CNN life prediction model is called for the long-life group / short-life group samples, and the feature data is input to predict the remaining battery life. The predicted value uses the root mean square error and mean absolute percentage error as evaluation indicators to compare the remaining life predicted by the model with the actual battery life.

8. A battery life early prediction system based on pre-classification and LSTM-CNN, characterized in that: It includes a processing instruction module for executing each step of a battery life early prediction method based on pre-classification and LSTM-CNN as described in any one of the claims.

9. A battery life early prediction device based on pre-classification and LSTM-CNN, characterized in that: include: Memory: a computer-readable device storing a computer program of a method for early prediction of battery life based on pre-classification and LSTM-CNN as described in any one of the above claims; Processor: used to implement a battery life early prediction method based on pre-classification and LSTM-CNN as described in any one of the claims when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement a battery life early prediction method based on pre-classification and LSTM-CNN as described in any one of the claims.

Images