Battery remaining life prediction optimization method based on interpretable deep learning algorithm

Abstract

The invention provides a battery remaining life prediction optimization method based on an interpretable deep learning algorithm. The method comprises the steps of constructing a data set and performing preprocessing; constructing model input data based on the preprocessed data set; constructing a battery residual life prediction model based on deep learning; interpreting decision behaviors of the model in the prediction process through an interpretable deep learning algorithm, and quantifying the contribution degree of input data in the prediction process; and optimizing aging data and a prediction process according to a model explanation result. According to the invention, the marginal contribution of the input features to the predicted value of the remaining life of the battery output by the neural network is calculated through the interpretable deep learning algorithm, and the contribution degrees of data of different working conditions, different feature parameters and different aging stages in the life prediction process are quantified, so that important aging features are identified and useless data are removed. The prediction process is optimized from the aspect of a data source, and the prediction speed is accelerated while the prediction accuracy is ensured, so that the consumption of computing resources is reduced.

Description

Technical Field

[0001] The present invention belongs to lithium-ion battery technology and relates to a battery remaining life prediction optimization method based on an explainable deep learning algorithm. Background Art

[0002] As a green and efficient energy storage device, lithium-ion batteries have broad application prospects in promoting energy transformation, promoting economic development, and improving environmental quality. However, the capacity of lithium-ion batteries decreases with the increase of charge and discharge cycles, and the performance gradually degrades. Battery life failure may lead to safety accidents, so battery life prediction is particularly important.

[0003] There are three typical methods for predicting the remaining life of lithium-ion batteries: the first is an experience-based method, the second is a model-based method, and the third is a data-driven method. The experience-based method is to complete the battery life prediction through the rules of historical data. The experience-based method is simple and fast, but it is only a rough estimate of RUL and does not meet the current accuracy requirements. The model-based method is to complete the life prediction by analyzing the operating environment of the battery, internal material properties, etc., but it takes a lot of time to establish an accurate model and the operation is cumbersome. The data-driven method is based on statistical analysis and realizes life prediction by mining the intrinsic relationship between input and response output. The data-driven method does not require specific knowledge of material properties, structure or failure mechanism. Although it is more convenient when modeling, the specific aging mechanism of the battery cannot be clearly expressed. Summary of the invention

[0004] In view of the problems and defects described in the above-mentioned prior art, a battery remaining life prediction optimization method based on an interpretable deep learning algorithm is proposed, and the method includes:

[0005] Build and preprocess the dataset;

[0006] Build model input data based on preprocessed dataset;

[0007] Build a battery remaining life prediction model based on deep learning;

[0008] Explain the decision-making behavior of the model during the prediction process and quantify the contribution of input data to the prediction process through interpretable deep learning algorithms;

[0009] Optimize aging data and prediction processes based on model interpretation results.

[0010] Furthermore, the data set is constructed and preprocessed, specifically including obtaining cycle aging data of the entire life cycle of the battery, using a filtering method to remove noise in the data and remove abnormal data, constructing evolution data based on target requirements based on the original data, and normalizing the preprocessed cycle aging data set.

[0011] Furthermore, the model input data is constructed based on the preprocessed data set, specifically including: taking the corresponding values ​​of the basic parameters and the voltage as the first dimension, taking the corresponding values ​​of the basic parameters in each cycle as the second dimension, and constructing matrix data based on the original cycle data according to the set cycle period window. For parameters that are not continuously sampled during the charging and discharging process, sequence data is constructed, which together with the matrix data serve as input data for the battery remaining life prediction model.

[0012] Furthermore, the deep learning-based construction of a battery remaining life prediction model specifically includes establishing a neural network suitable for processing large matrix data, automatically capturing important local features and spatial information in matrix parameters and screening them; establishing a neural network suitable for processing the relationship between time series data and time information in sequence data, combining the screened features with the sequence data for input, learning the global dependencies related to the remaining life, and realizing the battery remaining life prediction.

[0013] Furthermore, the method uses an interpretable deep learning algorithm to explain the decision-making behavior of the model in the prediction process and quantify the contribution of the input data in the prediction process. Specifically, it includes explaining the model through an interpretable algorithm, calculating the contribution value of each input data to the life prediction result, and clarifying the model's attention to each area of ​​the input data.

[0014] Furthermore, the aging data and prediction process are optimized according to the model interpretation results, specifically including: based on the interpretation of the model by an interpretable deep learning algorithm, statistically dividing the importance of each region, reflecting the areas where important aging characteristics are located, removing data irrelevant to aging characteristics, and simplifying the amount of battery aging data required in the prediction process.

[0015] Furthermore, the matrix data includes voltage, current, temperature, capacity, and energy that change with battery charging or discharging during each cycle under different cycle conditions, and the sequence data includes total charge and discharge capacity, charge and discharge time, and internal resistance during each cycle.

[0016] Furthermore, the cyclic aging data is filtered using a Savitzky-Golay filtering algorithm; the data normalization uses Min-Max normalization or Z-score standardization method to normalize the data; the neural network for large matrix data uses a convolutional neural network CNN or Transformer; the neural network for processing the relationship between time series data and time information in sequence data uses a long short-term memory network LSTM or a gated recurrent unit GRU; the interpretable deep learning algorithm uses a gradient weighted class activation map Grad-CAM or a hierarchical relevance propagation LRP, SHapley additional interpretation SHAP.

[0017] Furthermore, the contribution value of each input data to the life prediction result is calculated to clarify the model's attention to each area of ​​the input data, specifically including, according to the correlation between the basic parameters and the voltage in each cycle process, using the voltage as the basis for regional division, that is, dividing the basic parameters in the matrix data of the input model into several areas within the cut-off voltage range of charge and discharge, accumulating the contribution values ​​of all data in each area to obtain the total contribution of the data in the area, and using the total contribution as an indicator to measure the importance difference between areas.

[0018] Furthermore, the verification process of the method comprises the following steps:

[0019] The region with the largest contribution rate is used as input to calculate the evaluation index;

[0020] Gradually increase the input area range according to the contribution rate and calculate the evaluation index;

[0021] The evaluation indicators calculated each time are sorted. A larger evaluation value means that the accuracy of life prediction using partial aging data is higher than that using complete data, and the prediction time will be significantly shortened. Finally, a highest evaluation value is obtained, and its corresponding data range is used as the final data input range.

[0022] The present invention has the following beneficial technical effects:

[0023] (1) The present invention calculates the marginal contribution of input features to the battery remaining life prediction value output by the neural network through an interpretable deep learning algorithm, quantifies the contribution of data under different operating conditions, different characteristic parameters, and different aging stages in the life prediction process, thereby identifying important aging features and removing useless data, optimizing the prediction process from the data source, speeding up the prediction while ensuring the accuracy of the prediction, and reducing the consumption of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention has the following accompanying drawings:

[0025] Figure 1The voltage is used as a region division relationship diagram during the cycle of the present invention;

[0026] Figure 2 The present invention is based on contribution values ​​of different aging intervals. DETAILED DESCRIPTION

[0027] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0028] A battery remaining life prediction optimization method based on an interpretable deep learning algorithm, the specific optimization process includes the following steps:

[0029] S1: Construct a data set and perform preprocessing. Obtain the cycle aging data of the battery's entire life cycle, use appropriate filtering methods to remove the noise in the data caused by equipment sampling and remove the data that is relatively abnormal relative to the overall situation, build evolutionary data based on target requirements based on the original data, and finally select the commonly used normalization method to normalize the preprocessed aging data set.

[0030] S2: Construct model input data based on the preprocessed data set. Considering that the changes in each basic parameter during the battery charging and discharging process have a certain corresponding relationship with the battery voltage, the battery voltage is used as the horizontal axis, and other basic parameters in each cycle process are used as the vertical axis. According to the set cycle window, the matrix data based on the original cycle data is constructed. For the statistical parameters of each cycle that do not change with the charging and discharging process, the sequence data is constructed, and the matrix data is used together as the input data of the battery remaining life prediction model.

[0031] S3: Build a battery remaining life prediction model based on deep learning. The matrix data contains a lot of information about battery aging characteristics. By establishing a neural network suitable for processing large matrix data such as images, the important local features and spatial information in the matrix parameters are automatically captured and effectively screened. Then a neural network suitable for processing the relationship between time series data and time information in sequence data is established. The screened features are combined with the sequence data and input together to learn the global dependency related to the remaining life and realize the prediction of the remaining battery life.

[0032] S4: Use interpretable deep learning algorithms to explain the decision-making behavior of the model during the prediction process and quantify the contribution of input data in the prediction process. Neural networks have complex structures and poor interpretability. They cannot directly describe how the model makes decisions through clear rules or mathematical formulas like traditional algorithms. Therefore, the model can be explained through emerging interpretable algorithms, and the contribution of each input data to the life prediction results can be calculated to clarify the model's attention to each area of ​​the input data.

[0033] S5: Optimize aging data and prediction process based on model interpretation results. Based on the interpretation of the model by the interpretable deep learning algorithm, the importance of each region is statistically divided to reflect the areas where important aging characteristics are located, and data with fewer aging characteristics is removed, thereby simplifying the amount of battery aging data required in the prediction process, and achieving faster prediction speed while ensuring prediction accuracy.

[0034] Battery cycle aging data includes but is not limited to matrix data such as voltage, current, temperature, capacity, energy, etc. that change with battery charging or discharging during each cycle under different cycle conditions, as well as vector data such as total charge and discharge capacity, charge and discharge time, internal resistance, etc. during each cycle.

[0035] Battery aging data can be filtered using filtering algorithms such as Savitzky-Golay, and the data can be normalized using methods such as Min-Max normalization and Z-score standardization.

[0036] For neural networks suitable for processing large matrix data such as images, convolutional neural networks CNN, Transformer, etc. can be used. For neural networks suitable for processing the relationship between time series data and time information in sequence data, long short-term memory networks LSTM, gated recurrent units GRU, etc. can be used.

[0037] For interpretable deep learning algorithms used to explain models, gradient-weighted class activation mapping Grad-CAM, layer-wise relevance propagation LRP, SHapley additional interpretation SHAP, etc. can be used.

[0038] like Figure 1 As shown in the figure, after calculating the contribution value of each input data in the prediction process, according to the correlation between each basic parameter and the voltage in each cycle, the voltage is used as the basis for regional division, that is, each basic parameter in the matrix data of the input model is divided into several regions within the cut-off voltage range of charge and discharge, and the region number is set to 1, 2, 3, ..., n. The total contribution of the data in each region can be obtained by adding up the contribution values ​​of all data in the region, and it is used as an indicator to measure the importance difference between regions.

[0039] An interpretable deep learning algorithm is used to explain and analyze the learning process of the training set corresponding to the battery under different working conditions; among them, considering the different degradation paths of the experimental batteries under different working conditions, there are differences in the learning process of the aging data set, and these differences lead to different contributions of each data to the remaining life prediction results. Although the overall change trend of the data importance from region 1 to region n is similar under the same characteristic parameters of the battery under different working conditions, in order to improve the versatility of the method, the average value of the characteristic importance of the experimental batteries in each region under a variety of different working conditions is calculated, and the average value is used as the reference value for the final contribution evaluation.

[0040] The training set learning process corresponding to different battery characteristic parameters is explained and analyzed through an interpretable deep learning algorithm; in the prediction of the remaining battery life, there are a large number of characteristic parameters that can be used as the input of the neural network. Many parameters are coupled with each other and work together to obtain good prediction results. Therefore, it is impossible to determine the ranking of regional importance based on the contribution value distribution of only one parameter. In order to improve the reliability of the method, different characteristic parameter matrix data are used as the input of the prediction model, and the contribution value of each characteristic data is calculated using an interpretable deep learning algorithm to obtain the contribution distribution of different characteristic matrix data. In the contribution evaluation, the contribution of different features is weighted to obtain the contribution value ranking based on different characteristic parameters.

[0041] like Figure 2 As shown in the figure, the training set learning process corresponding to different battery aging stages is explained and analyzed by an interpretable deep learning algorithm; the battery remaining life prediction model is trained based on all battery aging data from unused to failure threshold, so it is still necessary to use the matrix data of different aging stages to analyze its contribution in predicting the remaining life, so as to comprehensively evaluate each performance area in the entire learning process. The contribution value of each characteristic parameter data matrix in different aging intervals is calculated using an interpretable neural network algorithm, and the contribution value based on different aging intervals is obtained by adding them up.

[0042] The present invention also proposes an evaluation index based on prediction error and prediction time to determine the optimal area to verify whether the optimization method adopted can accurately identify important features, thereby greatly saving computing resources while ensuring accuracy. The specific calculation includes the following steps:

[0043] (1) The region with the largest contribution rate is used as input to calculate the evaluation index;

[0044] (2) Gradually increase the input area range according to the contribution rate and calculate the evaluation index;

[0045] (3) The evaluation indicators calculated each time are ranked. A larger evaluation value means that the accuracy of life prediction using partial aging data is higher than that using complete data, and the prediction time will be significantly shortened. Finally, a highest evaluation value is obtained, and its corresponding data range is used as the optimal data input range.

[0046] The above embodiments describe the technical solutions of the present invention in detail. Obviously, the present invention is not limited to the described embodiments. Based on the embodiments of the present invention, people familiar with the technical field can also make various changes accordingly, but any changes that are equivalent or similar to the present invention belong to the scope of protection of the present invention.

[0047] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A battery remaining life prediction optimization method based on an interpretable deep learning algorithm, characterized in that: The method comprises: Build and preprocess the dataset; Build model input data based on preprocessed dataset; Build a battery remaining life prediction model based on deep learning; Explain the decision-making behavior of the model during the prediction process and quantify the contribution of input data to the prediction process through interpretable deep learning algorithms; Optimize aging data and prediction processes based on model interpretation results.

2. The method for optimizing the remaining battery life prediction according to claim 1, characterized in that: The data set is constructed and preprocessed, specifically including obtaining cycle aging data of the battery throughout its life cycle, removing noise in the data and abnormal data using a filtering method, constructing evolution data based on target requirements based on the original data, and normalizing the preprocessed cycle aging data set.

3. The method for optimizing the remaining battery life prediction according to claim 2, characterized in that: The model input data is constructed based on the preprocessed data set, specifically including: taking the corresponding values ​​of the basic parameters and the voltage as the first dimension, taking the corresponding values ​​of the basic parameters in each cycle as the second dimension, constructing matrix data based on the original cycle data according to the set cycle period window, and constructing sequence data for parameters that are not continuously sampled during the charging and discharging process, which are used together with the matrix data as input data for the battery remaining life prediction model.

4. The method for optimizing the remaining battery life prediction according to claim 3, characterized in that: The deep learning-based construction of a battery remaining life prediction model specifically includes: establishing a neural network suitable for processing large matrix data, automatically capturing important local features and spatial information in matrix parameters and screening them; establishing a neural network suitable for processing the relationship between time series data and time information in sequence data, combining the screened features with the sequence data for input, learning the global dependency relationship related to the remaining life, and realizing the battery remaining life prediction.

5. The method for optimizing the remaining battery life prediction according to claim 4, characterized in that: The method of explaining the decision-making behavior of the model in the prediction process and quantifying the contribution of the input data in the prediction process through an interpretable deep learning algorithm specifically includes explaining the model through an interpretable algorithm, calculating the contribution value of each input data to the life prediction result, and clarifying the model's attention to each area of ​​the input data.

6. The method for optimizing the remaining battery life prediction according to claim 5, characterized in that: The optimization of aging data and prediction process according to the model interpretation results specifically includes: based on the interpretation of the model by an interpretable deep learning algorithm, statistically dividing the importance of each region, reflecting the areas where important aging characteristics are located, removing data irrelevant to aging characteristics, and simplifying the amount of battery aging data required in the prediction process.

7. The method for optimizing the remaining battery life prediction according to claim 6, characterized in that: The matrix data includes voltage, current, temperature, capacity, and energy that change with battery charging or discharging during each cycle under different cycle conditions, and the sequence data includes total charge and discharge capacity, charge and discharge time, and internal resistance during each cycle.

8. The method for optimizing the remaining battery life prediction according to claim 7, characterized in that: The cyclic aging data is filtered using the Savitzky-Golay filtering algorithm; the data normalization uses the Min-Max normalization or Z-score standardization method to normalize the data; the neural network for large matrix data uses a convolutional neural network CNN or Transformer; the neural network for processing the relationship between time series data and time information in sequence data uses a long short-term memory network LSTM or a gated recurrent unit GRU; the interpretable deep learning algorithm adopts gradient weighted class activation mapping Grad-CAM or layer-wise relevance propagation LRP, SHapley additional interpretation SHAP.

9. The method for optimizing the remaining battery life prediction according to claim 8, characterized in that: The calculation of the contribution value of each input data to the life prediction result clarifies the degree of attention the model pays to each area of ​​the input data, specifically including, according to the correlation between the basic parameters and the voltage in each cycle process, taking the voltage as the basis for regional division, that is, dividing the basic parameters in the matrix data of the input model into several areas within the cut-off voltage range of charge and discharge, accumulating the contribution values ​​of all data in each area to obtain the total contribution of the data in the area, and using the total contribution as an indicator to measure the importance difference between areas.

10. The method for optimizing the remaining battery life prediction according to claim 9, characterized in that: The verification process of the method comprises the following steps: The region with the largest contribution rate is used as input to calculate the evaluation index; Gradually increase the input area range according to the contribution rate and calculate the evaluation index; The evaluation indicators calculated each time are sorted. A large evaluation value means that the accuracy of life prediction using partial aging data is higher than that using complete data, and the prediction time will be significantly shortened. Finally, a highest evaluation value is obtained, and its corresponding data range is used as the final data input range.

Images