Lithium ion battery remaining service life prediction method based on deep learning fusion model

Abstract

The invention discloses a lithium ion battery remaining service life prediction method based on a deep learning fusion model, and aims to solve the problems that feature extraction depends on complex engineering design, prediction precision is insufficient and adaptability is poor in an existing method. According to the method, charging and discharging times, charging and discharging voltage, charging and discharging time and battery capacity data in the charging and discharging process of a battery are collected, and a CNN-BiLSTM-attention mechanism (CNN-BiLSTM-attention) fusion model is constructed and used for predicting the RUL of the battery. The method comprises the specific steps of data acquisition and preprocessing, model construction, model training and testing, and prediction of the RUL of the battery by using historical data. The convolutional neural network is used for extracting local features, the bidirectional long-short-term memory network captures time dependence in a time sequence, and the attention mechanism endows features with different weights. According to the method, the accuracy and adaptability of the RUL prediction of the lithium ion battery are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage battery life prediction. Specifically, it is a method for predicting the remaining useful life of lithium-ion batteries. This method is a fusion deep learning model based on a Convolutional Neural Network (CNN), a Bi-directional Long Short-Term Memory (BiLSTM), and an Attention mechanism, aiming to achieve high-precision prediction of the remaining useful life of lithium-ion batteries. Background Art

[0002] With the increasing depletion of traditional fossil energy and the exacerbation of environmental pollution problems, the development of clean energy has become a key area of global concern. Clean energies such as wind energy and solar energy are highly favored due to their renewable nature. However, due to the volatility and instability of their energy supply, energy storage devices have become a key enabling technology to improve the utilization efficiency of clean energy.

[0003] Lithium-ion batteries are widely used in the energy storage field due to their high energy density, long cycle life, and other advantages. However, the battery will gradually age and fail during long-term use, resulting in a decline in battery performance. Accurately predicting the remaining useful life of the battery is of great significance for the maintenance and management of the energy storage battery system.

[0004] Currently, the methods for predicting the remaining useful life (RUL) of lithium-ion batteries are mainly divided into the following three categories: model-based prediction, data-driven prediction, and fusion method prediction. The core of model-based prediction is to establish a physical model of the battery through mathematical equations to describe its dynamic characteristics and degradation trends. For example, electrochemical models and equivalent circuit models. Among them, electrochemical models describe the physical and chemical changes occurring in the battery through detailed mathematical equations and can more accurately reflect the internal electrochemical mechanism of the battery. However, since the battery aging process involves multiple complex and interrelated degradation mechanisms, existing electrochemical models are difficult to comprehensively cover all aging factors. In addition, this method is highly dependent on experimental parameters, usually requiring high-precision parameter measurements, and is very sensitive to noise, resulting in insufficient robustness of model parameters. Similarly, equivalent circuit models describe the charge and discharge process of the battery through electronic components such as resistors, capacitors, and inductors. Although its structure is simple and the calculation efficiency is high, its applicability is poor, it is difficult to accurately reflect the complex dynamic characteristics of the battery, and the identification accuracy of model parameters is low. Generally speaking, model-based methods are highly dependent on model selection, with complex and time-consuming experimental parameters. Especially under specific or extreme working conditions, the applicability and prediction ability of the model are often limited. In addition, some models require high computing resources and show certain limitations in practical applications. Although this method can theoretically provide high prediction accuracy, its complexity and sensitivity to specific conditions make it difficult to be popularized in practical applications. Therefore, it is urgent to explore more flexible and efficient life prediction methods to overcome the above deficiencies.

[0005] Data-driven methods extract useful information from historical data through data processing and machine learning methods without relying on complex modeling of the internal aging mechanism of the battery. However, due to the limitations of the algorithm itself, traditional single data-driven methods have low prediction accuracy and are difficult to meet the actual application requirements. To overcome this problem, fusion methods have gradually emerged, mainly including the combination of physical models and data-driven methods and the fusion of multiple data-driven methods. Fusion methods can integrate the advantages of various methods, make up for the deficiencies of single methods, and significantly improve the accuracy and applicability of prediction.

[0006] In recent years, with the continuous improvement of hardware performance, data-driven fusion deep learning methods have shown great potential in the field of battery life prediction. Among them, CNN can extract local features from input data such as battery voltage, current, and temperature, BiLSTM can consider both the forward and backward information of the data to capture the dynamic laws during the battery aging process, and Attention can assign different weights to the features at different time steps to highlight the information that is more critical to the prediction result. The organic combination of these three models is expected to achieve accurate prediction of the RUL of lithium-ion batteries while significantly improving the prediction accuracy and robustness of the model. Summary of the Invention

[0007] The present invention aims to provide a deep learning fusion method based on a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism (Attention) for realizing high-precision prediction of the remaining useful life (RUL) of lithium-ion batteries (especially lithium-rich manganese-based energy storage batteries). To achieve the above object, the present invention adopts the following technical solution steps:

[0008] 1) Collect key data during the charge and discharge cycles of lithium-ion batteries, including voltage, time, capacity, and number of cycles for each charge and discharge cycle.

[0009] 2) Clean and normalize the data of the number of battery charge and discharge cycles, voltage of each charge and discharge cycle, time of each charge and discharge cycle, and battery capacity, and divide them into a training set and a test set.

[0010] 3) Establish a CNN-BiLSTM-Attention fusion deep learning model according to the normalized data sequence.

[0011] 4) Input the processed data into the model for training, adjust the model parameters through an optimization algorithm, and verify the model using the test set.

[0012] 5) Use the historical cycle data of the battery to be predicted as input, obtain the predicted value of the battery capacity in the later stage of the battery using the prediction model, and then calculate the remaining useful life value of the battery at the current moment through the corresponding relationship between the battery capacity prediction value and the number of battery cycles.

[0013] 6) Use the test data to evaluate the model performance indicators, calculate the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE) of the model to verify the prediction accuracy and robustness of the model.

[0014] The specific operation of step 2) is as follows:

[0015] 21) Clean the data to remove outliers and noise.

[0016] 22) Convert the charge and discharge voltage, charge and discharge time, charge and discharge times, and battery capacity into data matrices, and label them as x Vct , x Vdt , x Tdt , x Tct , x t , Q dt .

[0017] 23) Normalize the battery capacity data to facilitate model training and prediction. The normalization formula is:

[0018]

[0019] 24) Divide the normalized data into a training set and a test set according to a certain ratio, where the proportion of the training set is more than 50%.

[0020] The specific operation of step 3) is as follows:

[0021] 31) Input the preprocessed data into the CNN module, including charge and discharge voltage, battery capacity, number of cycles, and charge and discharge time, and input it as a time series with a sliding window size of N. The expression is:

[0022] z = [A, A t2 , …, A t N (2)

[0023] A t = [Q dt , x Vct , x Vdt , x Tdt , x Tct , x t N (3)

[0024] Among them, Z is the data sequence input with N as the sliding window, A t is the input data sequence, Q dt is the capacity data of the lithium-ion battery at the t-th cycle, A t is the input time series, x Vct is the average value of the charging voltage of the lithium-ion battery at the t-th cycle, x Vdt is the average value of the discharge voltage of the lithium-ion battery at the t-th cycle. x Tct is the average value of the charging time of the lithium-ion battery at the t-th cycle, x Tdt is the average value of the charging time of the lithium-ion battery at the t-th cycle, x t is the number of cycles.

[0025] 32) Extract the local features of the data through the convolutional layer:

[0026] Z j = f(∑A i * W i ) + b i (4)

[0027] And reduce the dimension through the pooling layer:

[0028] x i = max(C 1 , C 2 , … C​​m ) (5)

[0029] The ReLu function speeds up the training process and accelerates the convergence rate of gradient descent:

[0030] ReLu(x) = max(0, x) (6)

[0031] Where Z j is the local feature of the data extracted by the convolutional layer, f is the activation function, A i is the input sequence, * is the convolutional operation, W i is the weight matrix, b i is the bias matrix, C m is the data sequence of the input pooling layer, x i is the feature extracted by the convolutional layer, m is the width of the pooling layer, and x is the feature vector.

[0032] 33) Input the features extracted by the CNN into the BiLSTM module. The BiLSTM simultaneously models the sequence data through two LSTM networks, the forward and backward ones, to capture the bidirectional temporal dependencies in the battery aging process. In the t-th cycle, the output state of the forward LSTM network is H t,i , and the output state of the backward LSTM network is The output of the BiLSTM layer at the cycle number t is: H t = [H t,1 , H t,2 , … H t,i … H t,N N .

[0033] By introducing forget gates, output gates, and input gates, the LSTM realizes the selective retention and update of important information in the time series, thus solving the deficiencies of traditional time series models in long-term dependency modeling. The forget gate is used to determine the information to be forgotten in the current cell state; the input gate controls the writing of new information; the output gate is responsible for generating the output state at the current time step. This process is executed using the following mathematical formulas:

[0034] i t = σ(W i × [h t-1 , X t + b i ) (7)

[0035] f t = σ(W f × [h t-1 , X t + b f ) (8)

[0036] O t ​= σ(W o × [h t-1 , X t + b o ) (9)

[0037]

[0038] h t = o t × tanh(C t ) (12)

[0039] where W and b are LSTM parameters, is the candidate value of the new cell C t state. i t , f t , O t represent the forget gate, input gate, and output gate respectively. The activities of these gating mechanisms depend on the current input X t and the previous output h t-1 . The forget gate f t is responsible for deciding whether to inherit data from the previous state h t-1 , and updates the state h t-1 to the state C t through formula (10). Meanwhile, formula (11) generates the potential state value C t , and formula (12) calculates the current output h t of each LSTM cell. Subsequently, C t and h t are sequentially sent to the next process, and this loop is repeated through the time series. Through the above mechanism, BiLSTM can effectively capture the changing trends of long-term and short-term information in time series data, thus providing in-depth support for the prediction of the battery capacity attenuation trend.

[0040] 34) Apply the attention mechanism to the output of BiLSTM, assign different weights to the features at different time steps, and highlight the features that contribute more to the prediction of the remaining useful life of the battery. The weight assignment process of the attention mechanism is as follows:

[0041] Calculate the score of the attention weight:

[0042]

[0043] where, represents the context vector of the previous time step, and H t,i is the hidden state output of BiLSTM at the t-th time step.

[0044] Normalize the attention weight:

[0045]

[0046] Normalized α t,i is the weight assigned to the t-th time step, ensuring that the sum of all weights is 1.

[0047] 35) The attention mechanism receives its input from the hidden state H generated by the BiLSTM layer. At each time step, the output from the attention layer S t is calculated as follows: t

[0048] S t = α t,i H t,i (15)

[0049] 36) By introducing an attention mechanism for differential weight allocation in the hidden layer of the CNN-BiLSTM model, the feature vectors highly relevant to the prediction of the remaining battery service life are effectively highlighted, and all relevant information in the time series is comprehensively aggregated. Subsequently, Sigmoid is selected as the activation function of the output layer, and the aggregated vector S t is transformed into the predicted value Q t at the t-th time step. The calculation formula is:

[0050] Q t = sigmoid(wS t + b) (16)

[0051] where w is the weight of the output layer, b is the bias of the output layer, and Q t is the predicted battery capacity at time t.

[0052] The specific operations of step 4) are as follows:

[0053] 41) Use historical data to train the CNN-BiLSTM-Attention model, and adjust the model parameters through an optimization algorithm (Adam algorithm), with the root mean square error (RMSE) as the loss function.

[0054] 42) After training is completed, input the test set into the trained model to obtain the predicted value of the remaining service life of the lithium-ion battery.

[0055] 43) Denormalize the predicted value. Perform denormalization on the predicted value using the formula opposite to that of normalization to obtain the actual predicted result of the remaining service life.

[0056] The specific operations of step 6) are as follows:

[0057] 61) Calculate the coefficient of determination:

[0058]

[0059] ​62) Calculate the root mean square error:

[0060]

[0061] 63) Calculate the mean absolute error:

[0062]

[0063] Among them, is the predicted value, Q i is the actual value, and Q is the average value of the actual values.

[0064] Compared with the prior art, the present invention has the following advantages:

[0065] High precision: The CNN-BiLSTM-Attention model can extract deep features from the time-series data of the battery, and fully consider the bidirectional information of the time-series data and the weights of key features, significantly improving the accuracy of lithium-ion battery RUL prediction.

[0066] Strong adaptability: The model has good adaptability to the aging process of the battery under different working conditions, and can accurately predict the RUL of the battery under complex working conditions, providing a reliable reference for the actual application of the battery.

[0067] Good real-time performance: The training and prediction processes of the model are efficient, and can quickly respond to changes in the battery state, providing technical support for the real-time monitoring and maintenance of the battery

[0068] Using the charge and discharge time, charge and discharge voltage, charge and discharge times, charge and discharge capacity and other data of the test battery as training data, and a small amount of cycle data of the target battery as historical data, using the CNN-BilSTM-Attention fusion deep learning method to learn the deep high-order features between the data, obtaining the basis vectors that can express the essential common features between the data, so as to discover the deep relationship between different battery data and then realize life prediction. This method has good prediction accuracy and provides a good economic idea. The present invention has good application value and engineering prospects. Description of the Drawings

[0069] Figure 1 is the life prediction flow chart based on the deep learning fusion model

[0070] Figure 2 is the structure diagram of the CNN-BiLSTM-Attention model

[0071] Figure 3 is the dataset of the cycle times and capacity of the lithium-rich manganese-based energy storage battery collected in the laboratory

[0072] Figure 4The prediction results of the CNN-BiLSTM-Attention model for the lithium-rich manganese-based energy storage battery numbered LLO-1 tested in the laboratory

[0073] Figure 5 The prediction results of the CNN-BiLSTM-Attention model for the lithium-rich manganese-based energy storage battery numbered LLO-2 tested in the laboratory

[0074] Figure 6 The prediction results of the CNN-BiLSTM-Attention model for the data of the lithium cobalt oxide battery numbered CS2-38 tested by the University of Maryland

[0075] Figure 7 The prediction results of the CNN-BiLSTM-Attention model for the data of the lithium-ion battery numbered B0005 tested by NASA Laboratory Specific implementation manners

[0076] Specific implementation manner one combines with the attached Figure 1 And 2 To further introduce the present invention: The specific steps are as follows:

[0077] 1) Collect key data during the charge and discharge cycles of the lithium-rich manganese-based soft-pack lithium-ion battery, including voltage, time, capacity, and number of cycles for each charge and discharge cycle. As Figure 3 shown, the cycle dataset of the lithium-rich manganese-based soft-pack battery tested in the laboratory of the present invention.

[0078] 2) Clean, normalize, and divide the data such as the number of charge and discharge cycles of the battery, voltage of each charge and discharge cycle, time of each charge and discharge cycle, and battery capacity into a training set and a test set.

[0079] The specific operation of step 2 is as follows:

[0080] 21) Clean the data to remove outliers and noise.

[0081] 22) Convert the charge and discharge voltage, charge and discharge time, number of charge and discharge times, and battery capacity into data matrices respectively, and mark them as x

[0082] matrix, and sequentially mark them as x Vct ,x Vdt ,x Tdt ,x Tct ,x t ,Q dt 。

[0083] 23) Normalize the battery capacity data for easy model training and prediction.

[0084] 24) Divide the normalized data into a training set and a test set according to a certain ratio, where the proportion of the training set is 70%.

[0085] 3) Based on the normalized data sequence, establish a CNN-BiLSTM-Attention fusion deep learning model. The structure is as Figure 2 shown.

[0086] The specific operation is as follows:

[0087] 31) Input the preprocessed data into the CNN module, including charge and discharge voltage, battery capacity, number of cycles, and charge and discharge time, and input it as a time series with a sliding window size of N. Among them, the input feature dimension of the CNN layer is 5, the data is flattened into a 1D form, the convolutional kernel is [3,1], the sliding window N is 32, and the number of channels is 32.

[0088] 32) Extract the local features of the data through the convolutional layer and reduce the dimension through the pooling layer. The parameters of the pooling layer are 32×3×1. Use the ReLu function to accelerate the training speed.

[0089] 33) Input the features extracted by the CNN into the BiLSTM module. The BiLSTM simultaneously models the sequence data through two LSTM networks, the forward and the backward, to capture the bidirectional time dependence in the battery aging process. Among them, the BiLSTM layer is 2 layers. The output state of the forward LSTM network is H t,i , and the output state of the backward LSTM network is The output of the BiLSTM layer at the loop count t is: H t = [H t,1 , H t,2 , … H t,i … H t,N N .

[0090] 34) Apply the attention mechanism to the output of the BiLSTM, assign different weights to the features at different time steps, and highlight the features that contribute more to the prediction of the remaining useful life of the battery. And normalize the attention weights.

[0091] 4) Input the processed data into the model for training, adjust the model parameters through the Adam optimization algorithm, and use the test set to verify the model.

[0092] The specific operation is as follows:

[0093] 41) The CNN-BiLSTM-Attention model is trained using historical data, and the model parameters are adjusted through an optimization algorithm. The root mean square error (RMSE) is used as the loss function (RMSE < 0.1). The maximum number of iterations is 1000, the initial learning rate is 0.01, and the learning rate decay factor is 0.1.

[0094] 42) After training is completed, the test set is input into the trained model to obtain the predicted value of the remaining useful life of the lithium-ion battery.

[0095] 43) Denormalization is performed on the predicted value. The formula opposite to the normalization formula is used to obtain the actual predicted result of the remaining useful life.

[0096] 5) The historical cycle data of the battery to be predicted in the early stage is used as the input, and the predicted model is used to obtain the predicted value of the battery capacity of the battery in the later stage. Then, the remaining useful life value of the battery at the current moment is calculated through the corresponding relationship between the battery capacity predicted value and the battery cycle number.

[0097] 6) The performance metrics of the model are evaluated using the test data, and the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute error (MAE) of the model are calculated.

[0098] Specific Embodiment 2: Combining the attached Figure 4 , the RUL prediction results of the lithium-rich manganese-based energy storage battery under 25°C cycling are introduced. The specific implementation steps are as shown in Embodiment 1. Among them, for the prediction of a single battery, the lithium-rich manganese-based soft-pack battery numbered LLO-1 is selected. Its data is obtained through constant current charge and discharge tests at 25°C, 2.5 - 4.3V, and 1C current density. The prediction results are as Figure 4 shown. It can be noted that the prediction results are very close to the true values. Among them, R 2 is as high as 0.952, and the RMSE and MAE are only 0.147 and 0.118 respectively.

[0099] Specific Embodiment 3: Combining the attached Figure 5 introduces the cycle life prediction results of the lithium-rich manganese-based energy storage battery under 45°C cycling. The specific implementation steps are as shown in Embodiment 1. Among them, for the prediction of a single battery, the lithium-rich manganese-based soft-pack battery numbered LLO-2 is selected. Its data is obtained through constant current charge and discharge at 45°C, 2.5 - 4.3V, and 1C current density. The prediction results are as Figure 5 shown. It can be noted that the prediction results are very close to the true values. Its R 2 is as high as 0.982, and the RMSE and MAE are only 0.083 and 0.062 respectively. It proves that for the lithium-rich manganese-based energy storage battery operating under complex working conditions, this method can still effectively predict its remaining useful life.

[0100] Embodiment 4. To verify the generalization ability of the model, combined with the attached Figure 6 introduce the prediction results of the original data from the tests of the Center for Advanced Life Cycle Engineering (CALCE) at the University of Maryland. The specific implementation steps are as shown in Embodiment 1. The cathode material of the CALCE battery is lithium cobalt oxide (LiCoO 2 ), with a nominal capacity of 1.1 Ah. The serial number is CS2_38. This battery is charged at a constant current rate of 0.5C at the same room temperature condition (25°C) until the voltage reaches 4.2V, and then continues to be charged in a constant voltage mode until the charging current drops to 50 mA. The discharge test is carried out in a constant current mode at a rate of 1C until the voltage drops to 2.7V. The prediction results are as Figure 6 shown. It can be noted that the prediction results are very close to the true values, and its R 2 is as high as 0.987, and the RMSE and MAE are only 0.009 and 0.033 respectively.

[0101] Embodiment 5. To verify the generalization ability of the model, combined with the attached Figure 7 introduce the prediction results of the original data from the tests of the NASA laboratory. The specific implementation steps are as shown in Embodiment 1. The battery is a commercial 18650 battery with a nominal capacity of 2 Ah. The serial number is B0005. This battery is charged at a constant current of 1.5A at the same room temperature condition (25°C) until the voltage reaches 4.2V, and the discharge test is carried out in a constant current mode at a current of 1A until the voltage drops to 2.5V. The prediction results are as Figure 7 shown. It can be noted that the prediction results are very close to the true values, and its R 2 is as high as 0.989, and the RMSE and MAE are only 0.019 and 0.010 respectively.

[0102] Embodiment 6. To further verify the generalization ability of the model, introduce the results of other battery data combined with Table 1 in the appendix. The specific implementation steps are as shown in Embodiment 1. Here, the remaining useful life of different batteries is predicted simultaneously. In addition to the lithium-rich manganese-based soft-pack batteries measured in the laboratory (serial numbers LLO-1 to 5), it also includes the data of lithium cobalt oxide soft-pack batteries from CACLE (serial numbers CS2_37, CS2_38), and also includes the 18650-type batteries from NASA (serial numbers B0005, B0006). It can be found from Table 1 that basically all the prediction results R 2 are above 0.9, and both the RMSE and MAE are within 0.2.

[0103] Embodiment 7: Test the influence of different proportions of training data on the prediction results. It will be introduced in combination with Appendix 2. The specific implementation steps are as shown in Embodiment 1. The difference here is that three training sets with proportions of 60%, 70%, and 80% are set. When predicting the two battery data sets of LLO-2 and LLO-5 measured in the laboratory, it can be found that the more training set data there is, the 2 higher the R of the prediction result is, and the smaller the error is, which proves that the prediction effect is better.

[0104] Through experimental verification, the present invention has high prediction accuracy, can predict the RUL of the battery in the early stage, can greatly shorten the time required for battery life testing, and can realize the prediction of the remaining service life of the lithium-rich manganese-based energy storage battery with less data volume and simple feature engineering. Moreover, it not only has a certain application in lithium-rich manganese-based energy storage batteries, but also has a certain prediction accuracy in other batteries, which proves that the model has a certain generalization ability.

[0105] In summary, the method for predicting the RUL of lithium-ion batteries based on the CNN-BiLSTM-Attention model of the present invention has high precision, strong adaptability, good real-time performance and generalization ability, can provide reliable prediction results for the actual application of batteries, and has broad application prospects and important economic value.

[0106] Finally, it should be noted that the above embodiments are used to illustrate the superiority of the technical solution of the present invention, but do not limit it.

[0107] Although the above embodiments have described the present invention in detail, those skilled in the art should understand that they can still modify the solutions of the above embodiments, or perform equivalent replacements on some of the technical features, but these replacements do not cause the corresponding technical solutions to deviate from the scope defined by the claims of the present invention in essence.

[0108] Table 1 shows the error results of predictions for different data sets.

[0109] Table 2 shows the prediction error results at different prediction starting points.

[0110] Table 1

[0111] Battery number <![CDATA[R 2 > RMSE MAE LLO-1 0.952 0.147 0.118 LLO-2 0.982 0.083 0.062 LLO-3 0.991 0.080 0.017 LLO-4 0.965 0.106 0.101 LLO-5 0.972 0.091 0.081 CS2-37 0.966 0.104 0.092 CS2-38 0.987 0.009 0.033 B0005 0.989 0.019 0.010 B0006 0.992 0.016 0.009

[0112] Table 2

[0113]

Claims

1. A method for predicting the remaining service life of lithium-ion batteries based on a deep learning fusion model, characterized in that The specific steps include: 1) Collect multi-dimensional data generated by lithium-ion batteries during the charge and discharge cycle, including the number of charge and discharge times, the charge and discharge voltage of each charge and discharge cycle, the charge and discharge time, and the battery capacity; 2) Preprocess the collected multidimensional data, including cleaning outliers and normalizing features. And divide the training set and test set; 3) Based on the normalized multi-dimensional time series data, a fusion deep learning model of convolutional neural network (CNN)-bidirectional long short-term memory network (BiLSTM)-attention mechanism (Attention) is constructed to extract features and capture bidirectional dependencies in time series; 4) Input the processed data into the model, complete the model training, and use the test set to verify the accuracy of the model; 5) Input the early cycle data of the battery to be predicted into the trained model, predict the changing trend of the battery capacity with the number of cycles, and calculate the remaining service life of the battery based on the corresponding relationship between the capacity change trend and the number of cycles.

2. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that: The specific operations of step 2) are: 21) Clean the data to remove outliers and noise; 22) Convert the charge and discharge voltage, charge and discharge time, charge and discharge times and battery capacity into data matrix Array, and marked as x Vct , x Vdt , x Tdt , x Tct , x t , Q dt ; 23) Normalize the battery capacity data to facilitate model training and prediction. The normalization formula is: 24) The normalized data is divided into a training set and a test set, with the training set accounting for more than 50%.

3. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that: The specific operations of step 3) are: 31) Input the preprocessed data into the CNN module, including the charge and discharge voltage, battery capacity, number of cycles, and charge and discharge time, as a time series input with a sliding window size of N, and the expression is: Z=[A,A t2 ,…,A t ] N (2) A t =[Q dt ,x vct ,x Vdt ,x Tdt ,x Tct ,x t ] N (3) Among them, Z is the data sequence with N as the sliding window, A t is the input data sequence, Q dt is the capacity data of the lithium-ion battery at the tth cycle, x Vct is the average value of the charging voltage of the lithium-ion battery in the tth cycle, x Vdt is the average discharge voltage of the lithium-ion battery in the tth cycle; X Tct is the average charging time of the lithium-ion battery in the tth cycle, x Tdt is the average charging time of the lithium-ion battery in the tth cycle, x t is the number of cycles; 32) Extract local features of data through convolutional layers: Z j =f(∑A i *W i )+b i (4) And reduce the dimension through the pooling layer: x i =max(C1,C2,...C m ) (5) The ReLu function speeds up training and accelerates the convergence of gradient descent: ReLu(x)=max(0,x) (6) Where Z j is the local feature of the data extracted by the convolution layer, f is the activation function, A i is the input sequence, * is the convolution operation, W i is the weight matrix, b i is the bias matrix, C m is the input pooling layer data sequence, x i is the feature extracted by the convolutional layer, m is the width of the pooling layer, and x is the feature vector; 33) The features extracted by CNN are input into the BiLSTM model. BiLSTM models the sequence data simultaneously through two LSTM networks, forward and backward, to capture the bidirectional time dependency in the battery aging process. In the tth cycle, the output state of the forward LSTM network is H t，i , the output state of the backward LSTM network is The output of the BiLSTM layer at cycle number t is: H t =[H t，1 , H t，2 , …H t，i …H t，N ] N ; LSTM achieves the selective retention and update of important information in the time series by introducing forget gate, output gate and input gate, thus solving the shortcomings of traditional time series models in modeling long-term dependencies; the forget gate is used to determine the information that needs to be forgotten in the current unit state; the input gate controls the writing of new information; the output gate is responsible for generating the output state of the current time step; the process is executed using the following mathematical formula: i t =σ(W i ×[h t-1 ,X t ]+b i ) (7) f t =σ(W f ×[h t-1 ,X t ]+b f ) (8) The t =σ(W o ×[h t-1 ,x t ]+b o ) (9) h t =o t ×tanh(C t ) (12) Where W and b are LSTM parameters, X t is the current input sequence; It is a new unit C t Candidate value of state; i t , f t , o t Represent the forget gate, input gate, and output gate respectively; the activities of these gating mechanisms depend on the current input X t and the previously output h t-1 ; Forget gate f t Responsible for deciding whether to start from the previous state h t-1 Inherit data, and use formula (10) to convert the state h t-1 Updated to status C t ; At the same time, formula (11) generates the potential state value C t , and formula (12) calculates the current output h of each LSTM unit t ; Then, C t and h t are sent to the next process in turn, repeating the cycle through the time series; 34) Apply the attention mechanism to the output of BiLSTM, assign different weights to the features at different time steps, and highlight the features that contribute more to the prediction of the remaining battery life; the weight allocation process of the attention mechanism is as follows: Calculate the score of attention weight: Among them, e t,i is the attention weight score, represents the vector of the previous time step, H t,i is the hidden state output of BiLSTM at the tth time step; Normalized attention weights: Normalized α t,i is the weight assigned to the tth time step, ensuring that the sum of all weights is 1; 35) The attention mechanism generates the hidden state H from the BiLSTM layer t Receives its input, at each time step, from the attention layer S t The output calculation formula is as follows: S t =a t,i H t,i (15) 36) Select Sigmoid as the activation function of the output layer and convert the aggregated vector S t Transformed into the predicted value Q at the tth time step t ; The calculation formula is: Q t =sigmoid(wS t +b) (16) Among them, w is the output layer weight, b is the output layer bias, Q t is the predicted battery capacity at time t.

4. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that The specific operations of step 4) are: 41) Use historical data to train the CNN-BiLSTM-Attention fusion model and adjust the model parameters through the optimization algorithm; specifically, use the charge and discharge cycle data as input, perform forward propagation to calculate the prediction error; use the back propagation algorithm to optimize the model weights and bias parameters to minimize the loss function; repeat the above process until the training converges and reaches the set performance target, that is, RMSE < 0.1; 42) After the training is completed, the historical data of the battery to be predicted is input into the model, the change trend of the battery's future capacity is output, and its remaining service life is calculated based on the relationship between capacity and cycle number; 43) Denormalization: Denormalize the predicted value and use the opposite formula to obtain the actual remaining useful life prediction result.

5. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that: The method also includes step 6) using the predicted values ​​output by the model to calculate the model performance indicators, including the coefficient of determination (R-Square, R 2 ), Root Mean Square Error (RMSE) and Mean Absolute Error (MAE); The specific operations of step 6) are: 61) Calculate the coefficient of determination: 62) Calculate the root mean square error: 63) Calculate the mean absolute error: in, is the predicted value, Q i is the actual value, and Q is the average value of the actual value.

6. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that The specific structure of the CNN layer in step 3) is as follows: the input data is flattened into a one-dimensional sequence, the convolution layer adopts a two-layer convolution structure to extract multi-scale local features, the pooling layer is used for dimensionality reduction to reduce feature redundancy, and the convolution layer and the pooling layer are followed by a connection layer.

7. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 6, characterized in that: In step 3), the convolution layer is any one of 128×3×1, 64×3×1, and 32×3×1, the pooling layer is any one of 128×3×1, 64×3×1, and 32×3×1, and the flattening layer is any one of 128×1, 64×1, and 32×1.

8. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that: In step 4), the maximum number of iterations is 800-2000, the batch size is selected as any one of 32, 64, and 128, the initial learning rate is 0.01, and after 500-1500 trainings, the learning rate is 0.01×0.

1. The Adam optimization algorithm is used for network parameter optimization.

9. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that Step 6) R 2 The evaluation result must reach 0.9 or above.

10. The method for predicting the remaining service life of a lithium-ion battery based on a deep learning fusion model according to claim 1, characterized in that The prediction of remaining service life is based on the battery capacity change trend and is used to evaluate the relationship between capacity and cycle number.

Images