Early cycle life prediction method for lithium batteries based on improved CNN algorithm based on degradation classification

Abstract

The present invention relates to a method for predicting the early cycle life of lithium batteries based on an improved CNN algorithm for degradation classification. A clustering algorithm is used to classify batteries, and corresponding improved CNN networks are designed for batteries with different degradation modes. The prediction is performed by combining manual and self-extracted features, thereby ultimately improving the cycle life prediction accuracy. The method has the advantages of high accuracy and ease of use. The method uses domain adaptation and SE modules to design CNN networks respectively, and can use degradation classification results to optimize regression prediction results, improve the generalization and versatility of the network, and thus improve the cycle life prediction accuracy. Compared with the existing CNN prediction model, by designing appropriate CNN model parameters, unsupervised clustering is used to preliminarily divide batteries into short-life batteries and long-life batteries, and then domain adaptation and SE modules are introduced respectively, thereby improving the life prediction accuracy of short-life batteries and long-life batteries.

Description

Technical Field

[0001] The present invention belongs to the field of lithium battery life prediction, and relates to a lithium battery early cycle life prediction method based on a degradation classification improved CNN algorithm, especially using battery early degradation data and a deep learning algorithm to predict the cycle life of a lithium battery. Background Art

[0002] Lithium batteries, due to their high energy density, compact size, and low pollution, are widely used in applications such as autonomous driving, aerospace, and consumer electronics. The global lithium-ion battery market is projected to exceed $170 billion by 2030. However, due to the complex chemical mechanisms inherent in batteries, the capacity of lithium-ion batteries inevitably decreases during the charge and discharge process. When the battery capacity reaches 70%-80% of its initial capacity, it is considered to have reached the end of its cycle life. If a battery is still used after reaching the end of its cycle life, it will not only cause financial losses but also pose risks to human health. Therefore, accurately predicting the cycle life of lithium batteries is extremely important for ensuring their safe use. The earlier the cycle life is predicted, the more timely fault diagnosis can be performed to prevent potential risks. However, in the early stages of battery degradation, capacity degradation is not obvious, making accurate cycle life prediction extremely challenging.

[0003] Currently, the prediction of early-stage degradation cycle life is primarily based on data-driven methods, including those based on manual features and those using deep learning-based self-extracted features, such as elastic nets, gradient boosted regression trees (GBRT), random forests (RF), Gaussian process regression (GPR), stacked denoising autoencoders (SDAE), recurrent neural networks (RNN), and convolutional neural networks (CNN). CNNs outperform algorithms like RNNs and RFs in early-stage cycle life prediction. However, these methods still overlook the role of degradation classification in early-stage prediction and the complementarity of different features. If both manual and self-extracted features can be considered simultaneously, and a CNN more adapted to battery degradation patterns can be designed, the accuracy of cycle life prediction can be further improved.

[0004] The related papers of the prior art are:

[0005] 1. Severson KA, Attia PM, Jin N, Perkins N, Jiang B, Yang Z, et al. Data-driven prediction of battery cycle life before capacity degradation. NatEnergy 2019; 4(5): 383.

[0006] 2. Ma G, Wang Z, Liu W, et al. A two-stage integrated method for early prediction of remaining useful life of lithium-ion batteries [J]. Knowledge-Based Systems, 2023, 259: 110012.

[0007] 3. Xu Q, Wu M, Khoo E, et al. A hybrid ensemble deep learning approach forearly prediction of battery remaining useful life [J]. IEEE / CAA Journal of Automatica Sinica, 2023, 10(1): 177-187.

[0008] 4. Attia PM, Severson KA, Witmer J D. Statistical learning for accurate and interpretable battery lifetime prediction[J]. Journal of The Electrochemical Society, 2021, 168(9):090547. Summary of the Invention

[0009] Technical problems to be solved

[0010] To overcome the shortcomings of existing technologies, this paper proposes a method for predicting the early cycle life of lithium batteries based on a degradation classification-based improved CNN algorithm. This method uses a clustering algorithm to classify batteries, designs an improved CNN network tailored to the degradation patterns of batteries, and combines manual and self-extracted features for prediction, ultimately improving cycle life prediction accuracy. The method boasts high accuracy and ease of use.

[0011] Technical Solution

[0012] A method for predicting the early cycle life of lithium batteries based on a degradation classification improved CNN algorithm is characterized by the following steps:

[0013] Step 1: Divide the discharge data of the early degradation process into training sets and test sets, remove and correct the abnormal data in the training sets and test sets, extract statistical features from the cleaned data, and normalize the cleaned data;

[0014] Step 2: Use the K-means clustering algorithm to classify the battery data obtained in step 1 and divide the batteries into long-life batteries and short-life batteries;

[0015] Step 3: The discharge data of all batteries in the training set and the short-life batteries in the test set, as well as the statistical feature data from step 1, are input into the Domain Alignment CNN (AL-CNN) network. The AL-CNN network is trained to obtain the AL-CNN network for cycle life prediction of short-life batteries.

[0016] The discharge data of all batteries in the training set and the statistical feature data from step 1 are input into the Squeeze-and-Excitation CNN (SE-CNN) network. The SE-CNN network is trained to obtain the SE-CNN network for cycle life prediction of long-life batteries.

[0017] The AL-CNN network includes a CNN network composed of two-dimensional convolutional modules Conv2d and a fully connected network, wherein: each two-dimensional convolutional module Conv2d is connected to a maximum pooling layer, and the back end of the CNN network is connected to two fully connected networks for regression. The final prediction output is a 1×1 cycle life;

[0018] The SE-CNN network includes a CNN network consisting of a two-dimensional convolutional module Conv2d, an SE module inserted between two convolutional layers, and a fully connected network connected to the back end of the CNN network. The two-dimensional convolutional module Conv2d is connected to a maximum pooling layer, which is the same as the two-dimensional convolutional module Conv2d and maximum pooling layer in the AL-CNN network. The SE module compresses the data output by the first maximum pooling layer, converting each two-dimensional channel into a one-dimensional scalar. Then, an excitation operation is used to learn the weight of each channel. After learning the corresponding weight of each channel, the weight is reassigned to each channel, and a product operation is performed to output a new feature map. The new feature map is then passed through the convolution, pooling layers, and two fully connected networks, and the final predicted output is a 1×1 cycle life.

[0019] Step 4: After processing the battery measurement data using steps 1 and 2, the long-life battery discharge data and characteristics or the short-life battery discharge data and characteristic data are obtained, which are input into the SE-CNN network or the AL-CNN network respectively to obtain the cycle life prediction value of the long-life battery or the cycle life prediction value of the short-life battery.

[0020] The extracted statistical features are the variance Var, minimum Min, skewness Skew, and kurtosis Kurt of the difference between the discharge capacity-voltage curves of the 100th cycle and the 10th cycle, the discharge capacity of the second cycle, and the difference between the maximum discharge capacity of the first 100 cycles and the discharge capacity of the second cycle, expressed as:

[0021]

[0022] The K-means clustering algorithm is as follows: select the number of clusters n, select n cluster centers from the input data, calculate the sum of the Euclidean distances between all input data and their cluster centers, and find new cluster centers by minimizing the cost function; repeat the above steps until the cluster centers tend to be stable. The cost function of this algorithm is defined as follows:

[0023]

[0024] Among them, x represents the input data, which is usually manually extracted features, and θ i is the center of the i-th category, C i is the i-th category.

[0025] The training and testing steps of the AL-CNN network in step 3 are as follows:

[0026] Step (1) uses a convolutional network to extract features from the input data: no padding operation is performed in the convolution layer, and the convolution process is:

[0027]

[0028] Among them, E represents the input matrix, the shape of the convolution kernel W is (f1, f2), C represents the output feature map, C i,j is the convolution feature located at (i, j) in C;

[0029] The process of maximum pooling after the convolutional layer is:

[0030]

[0031] Among them, the size of the pooling window is k×k, the step size is s, X and Y represent the input and output tensors respectively, (i, j) are the spatial coordinates of the tensor, c represents the channel index, and max represents the maximum function; if the input shape is H×W×C, then the shape of Y is H′×W′×C, where

[0032] In step (2), after the convolutional network extracts features, that is, at the end of the CNN network, the flattened features and statistical features learned by the CNN network are fused using a connection operation to form a new fused feature vector; then, two fully connected layers are added for regression, and the final prediction output is a 1×1 cycle life;

[0033] The nonlinear activation function used after each convolutional layer and fully connected layer during training is the rectified linear unit ReLU;

[0034] The weighted summation formula of the fully connected layer is as follows:

[0035]

[0036] ReLU(x)=max(0,x)

[0037] where w i is the weight in the fully connected layer, b i is the bias, x i is the input neuron;

[0038] Step (3) parameter update:

[0039] First, the CORAL loss function is used for feature alignment. The CORAL loss function is defined as follows:

[0040]

[0041] Among them, C S and C T are the two-dimensional feature covariance matrices of the source domain and the target domain, respectively, || || F represents the F norm;

[0042] The correction of the loss function l in the AL-CNN network is as follows:

[0043] l = MSE CL +lbd*L coral

[0044]

[0045] Among them, y i is the actual cycle life, To predict the cycle life, n is the number of samples, lbd is the weight of the CORAL function, which is a hyperparameter that needs to be adjusted;

[0046] Then, the back-propagation algorithm is used to update the CNN weights:

[0047]

[0048] Where Δw i,jis the change in weight, η is the learning rate, is the partial derivative of the loss function l with respect to the weight;

[0049] Through back propagation, the weights of each layer can be updated according to the target loss function, so that the parameters of the model are continuously updated, thereby extracting features corresponding to the task and improving prediction accuracy;

[0050] Step (4) As the model parameters are updated, the network will continue to converge. The validation set and the "early stopping mechanism" are used to determine whether the model has converged. If it has converged, the trained network model is saved. Finally, the short-life battery discharge data and statistical features in the test set are input into the AL-CNN network to obtain the cycle life prediction results of the short-life battery.

[0051] The loss function of the SE-CNN network is different from that of the AL-CNN network. The loss function is the mean square error (MSE). The update of the model parameters is the same as that of the AL-CNN network.

[0052] The compression operation in the SE-CNN network compresses the shape of the tensor U from H×W×C to 1×1×C, where C is the number of channels. Through the squeezing operation, each two-dimensional channel becomes a global value. The squeezing operation is described as follows:

[0053]

[0054] Global spatial information is squeezed into the channel descriptor through global average pooling (GAP); the correlation between channels is fully utilized, and each two-dimensional channel is converted into a one-dimensional scalar, ignoring the internal redundant spatial distribution information.

[0055] The excitation operation uses the information obtained from the compression operation to learn the weight of each channel. The process of the excitation operation is as follows:

[0056] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0057]

[0058] In the formula, δ represents the ReLU function, σ refers to the sigmoid function, r is a hyperparameter used to reduce the number of channels and reduce the amount of computation.

[0059] After learning the corresponding weights of each channel, the product reassigns the weights to each channel and performs a product operation to output a new feature map. The process is described as follows:

[0060]

[0061] Among them, uc represents the cth channel in U, s c Represents the weight of the corresponding channel.

[0062] A readable storage medium, characterized in that a computer program is stored on the readable storage medium, and when the computer program is executed by a processor, the steps of the lithium battery early cycle life prediction method based on the degradation classification improved CNN algorithm are implemented.

[0063] A computer program product, characterized in that it includes computer executable instructions, which, when executed, are used to implement the lithium battery early cycle life prediction method based on the degradation classification improved CNN algorithm.

[0064] Beneficial effects

[0065] The present invention proposes a lithium battery early cycle life prediction method based on a degradation classification improved CNN algorithm. The method uses a clustering algorithm to classify batteries, designs an improved CNN network corresponding to batteries with different degradation modes, and combines manual and self-extracted features for prediction, ultimately improving the cycle life prediction accuracy. Therefore, the method has the advantages of high accuracy and ease of use.

[0066] This method combines two types of features, manually extracted and self-extracted, to design CNN networks for batteries with different degradation patterns using domain adaptation and SE modules, respectively. This method leverages degradation classification results to optimize regression predictions, improving the network's generalization and versatility, and thus enhancing cycle life prediction accuracy. Compared to existing CNN prediction models, this method improves the accuracy of cycle life prediction by designing appropriate CNN model parameters and employing unsupervised clustering to initially classify batteries into short-life and long-life types. Domain adaptation and SE modules are then introduced, respectively, to improve cycle life prediction accuracy for short-life and long-life batteries.

[0067] This method takes into account the optimization effect of classification on prediction and provides a new perspective to solve the problem of battery life prediction. Figure 4 It can be seen from the simulation results that the predicted cycle life is basically very close to the actual cycle life, which illustrates the effectiveness of the present invention. Figure 5As the ablation experiment results, by comparing with other CNN methods that do not perform classification, the present invention has the lowest mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). This can further illustrate the optimization effect of classification on prediction and reflects the innovativeness of the present invention. Finally, Table 1 shows a comparison between the present invention and the existing mainstream lithium battery cycle life prediction methods. It can be seen that the battery life prediction accuracy of the present invention is the highest, which can further illustrate the superiority of the present invention. Applying this method to the prediction of early battery cycle life can significantly improve the accuracy of battery cycle life prediction, and it is a simple and effective method. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 Flow chart of the method of the present invention

[0069] Figure 2 AL-CNN network structure diagram

[0070] Figure 3 SE-CNN network structure diagram

[0071] Figure 4 Cycle life prediction results of this method on the entire MIT dataset

[0072] Figure 5 Comparison of the cycle life prediction of this method with CNN, SE-CNN, AL-CNN, and SE-AL-CNN (combination of the two) on 82 batteries DETAILED DESCRIPTION

[0073] The present invention will now be further described with reference to the embodiments and accompanying drawings:

[0074] The present invention proposes a lithium battery early cycle life prediction method based on an improved CNN algorithm for degradation classification. A clustering algorithm is applied to classify batteries, and an improved CNN network corresponding to batteries with different degradation modes is designed. The prediction is performed by combining manual and self-extracted features, ultimately improving the cycle life prediction accuracy. Therefore, the method has the advantages of high accuracy and ease of use.

[0075] The present invention provides a method for predicting the early cycle life of lithium batteries based on a degradation classification-improved CNN algorithm. Taking the MIT battery dataset as an example, the charge and discharge experimental data of the early stage (the first 100 cycles) of the lithium battery are collected, and the cycle life of the lithium battery is predicted based on the degradation classification-improved CNN algorithm. The method comprises the following steps:

[0076] Step 1: Collect discharge data from the MIT dataset during the early degradation process of batteries, remove and correct abnormal data, manually extract statistical features from the cleaned data, and normalize the cleaned data.

[0077] The features extracted manually are as follows:

[0078] Var(ΔQ100-10(V))

[0079] Min(ΔQ100-10(V))

[0080] Skew(ΔQ100-10(V))

[0081] Kurt(ΔQ100-10(V))

[0082] Discharge_cap_2

[0083] Difference_max_cap_cycle_2

[0084] They are the variance, minimum, skewness, and kurtosis of the difference between the discharge capacity-voltage curves of the 100th cycle and the 10th cycle, the discharge capacity of the second cycle, and the difference between the maximum discharge capacity of the first 100 cycles and the discharge capacity of the second cycle.

[0085] Step 2: Use the K-means clustering algorithm to classify the batteries into long-life batteries and short-life batteries.

[0086] Specifically, we select the number of clusters n, select n cluster centers from the input data, calculate the sum of the Euclidean distances between all input data and their cluster centers, and find new cluster centers by minimizing the cost function. Repeat the above steps until the cluster centers become stable. The cost function of this algorithm is defined as follows:

[0087]

[0088] Among them, x represents the input data, which is usually manually extracted features, and θ i is the center of the i-th category, C i is the i-th category.

[0089] Step 3: Design the Domain Alignment CNN (AL-CNN) network to achieve cycle life prediction of short-life batteries. The network input is the discharge data and manual features of all batteries in the entire training set and the short-life batteries in the test set. Its purpose is to make the short-life battery data in the test set continuously adapt to the training set. The AL-CNN network mainly includes a convolutional network for feature extraction, a fully connected network and a parameter update strategy, such as Figure 2 The training and testing steps of AL-CNN are as follows:

[0090] 3.1. First, a convolutional network is used to extract features from the input data. The CNN convolutional network used for feature extraction in AL-CNN relies entirely on the two-dimensional convolution module Conv2d in the Pytorch framework. This convolutional network extracts features from the input sequence. This convolutional network consists of two convolutional layers, with 6 and 16 kernels, respectively. The kernel size is 3, the stride is 1, and no padding is performed in the convolutional layers. The convolution process can be written as:

[0091]

[0092] Among them, E represents the input matrix, the shape of the convolution kernel W is (f1, f2), C represents the output feature map, C i,j is the convolution feature located at (i, j) in C.

[0093] After each convolutional layer, there is a maximum pooling layer with a pooling kernel size of 2×1 and a pooling step size of 2. The second pooling layer performs a zero-filling operation. The maximum pooling process is:

[0094]

[0095] Where the size of the pooling window is k×k, the step size is s, X and Y represent the input and output tensors respectively, (i, j) are the spatial coordinates of the tensor, c represents the channel index, and max represents the maximum function. If the input shape is H×W×C, then the shape of Y is H′×W′×C, where It should be noted that the size of channel C does not change.

[0096] 3.2. After feature extraction by the convolutional network, at the end of the CNN network, the flattened features learned by the CNN network and the manually extracted features are fused using a concatenation operation to form a new fused feature vector. Subsequently, two fully connected layers are added for regression, and the final predicted output is a 1×1 cycle life.

[0097] In addition, the nonlinear activation function used after each convolutional layer and fully connected layer is the rectified linear unit ReLU. The weighted summation formula of the fully connected layer is as follows:

[0098]

[0099] ReLU(x)=max(0,x)

[0100] where w i is the weight in the fully connected layer, b i is the bias, x i is the input neuron.

[0101] 3.3 Parameter update

[0102] First, in order to shorten the difference between the source domain and the target domain, the CORAL loss function is used for feature alignment. The CORAL loss function is defined as follows:

[0103]

[0104] Among them, C S and C T are the two-dimensional feature covariance matrices of the source domain and the target domain, respectively, || || F represents the F-norm.

[0105] Then, the correction of the loss function l in the AL-CNN network is as follows:

[0106] l = MSE CL +lbd*L coral

[0107]

[0108] Among them, y i is the actual cycle life, To predict the cycle life, n is the number of samples, lbd is the weight of the CORAL function, and is a hyperparameter that needs to be adjusted.

[0109] Then, the back-propagation algorithm is used to update the CNN weights:

[0110]

[0111] Where Δw i,j is the change in weight, η is the learning rate, is the partial derivative of the loss function l with respect to the weights. Through backpropagation, the weights of each layer can be updated according to the target loss function, so that the parameters of the model can be continuously updated, thereby extracting features corresponding to the task and improving prediction accuracy.

[0112] 3.4. As model parameters are updated, the network continues to converge. The validation set and the "early stopping mechanism" are used to determine whether the model has converged. If converged, the trained network model is saved. Finally, the short-life battery discharge data from the test set is input into the AL-CNN network to obtain the cycle life prediction results of the short-life batteries.

[0113] Step 4: Design the Squeeze-and-Excitation CNN (SE-CNN) network, such as Figure 3As shown in the figure, the cycle life prediction of long-life batteries is achieved. By learning global information, useful channel features are selectively emphasized while less useful channel features are suppressed, thereby increasing the network representation power, thereby reducing the prediction error from a global perspective. The input data is the discharge data and manual features of all batteries in the entire training set. The SE-CNN training and testing steps are similar to those of AL-CNN, but the network structure is different. The SE module in SE-CNN mainly includes three steps: compression, excitation, and product, as follows:

[0114] 4.1. The compression operation compresses the shape of the tensor U from H×W×C to 1×1×C, where C is the number of channels. Through the compression operation, each two-dimensional channel becomes a global value. The compression operation can be described as follows:

[0115]

[0116] Global spatial information is squeezed into the channel descriptor through global average pooling (GAP). In order to fully utilize the correlation between channels, each two-dimensional channel is converted into a one-dimensional scalar, ignoring the internal redundant spatial distribution information.

[0117] 4.2. In order to utilize the information obtained by the compression operation, we use the excitation operation to learn the weight of each channel. The process of the excitation operation is as follows:

[0118] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0119]

[0120] In the formula, δ represents the ReLU function, σ refers to the sigmoid function, r is a hyperparameter used to reduce the number of channels and reduce the amount of computation.

[0121] 4.3. After learning the corresponding weights for each channel, the weights are reassigned to each channel, and the product operation is performed to finally output a new feature map. The process is described as follows:

[0122]

[0123] Among them, u c represents the cth channel in U, s c Represents the weight of the corresponding channel.

[0124] 4.4. Similar to the structure of AL-CNN, SE-CNN inserts the SE module between the first and second convolutional layers in AL-CNN, but the loss function remains unchanged, using the mean squared error (MSE). Model parameter updating is similar to AL-CNN. After training the model and saving the parameters, long-life batteries in the test set are tested to obtain cycle life prediction results.

[0125] Step 5: Summarize the cycle life prediction values ​​of long-life and short-life batteries to test the overall prediction effect of the test set.

[0126] Attachment Figure 4 The horizontal axis is the actual cycle life, and the vertical axis is the predicted cycle life. The closer the point is to the middle diagonal line, the closer the predicted life is to the actual life. The circle represents the trained battery. It can be seen that the circle is very close to the middle diagonal line, indicating that the model training effect is good; the triangle represents the tested battery, which is used to test the accuracy of the model prediction. It can be seen that the triangle as a whole is still very close to the diagonal line, which means that the predicted value is closer to the true value, that is, the prediction accuracy is relatively high.

[0127] Attachment Figure 5 The horizontal axis represents different CNN methods; the vertical axis on the left represents the prediction error metrics RMSE and MAE, measured in cycles, both plotted as bar graphs; the vertical axis on the right represents the prediction error metric MAPE, measured in %, plotted as a broken line. As can be seen from the figure, the method of the present invention (hybrid clustered CNN) achieves the lowest results for all three error metrics, indicating that the present invention has the highest battery life prediction accuracy.

[0128] Comparison of cycle life prediction simulation results: Using the MIT experimental data set, the CNN algorithm improved by degradation classification can accurately predict the battery cycle life, with an overall accuracy of 6.68%. The battery cycle life prediction results are as follows: Figure 4 As shown in FIG. 1 , it can be seen that the predicted cycle life is basically very close to the actual cycle life, which illustrates the effectiveness of the present invention. Figure 5As the ablation experiment results, by comparing with other CNN methods that do not perform classification, the present invention has the lowest mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). This can further illustrate the optimization effect of classification on prediction and reflects the innovation of the present invention. Finally, Table 1 shows a comparison between the present invention and existing mainstream lithium battery cycle life prediction methods. As can be seen from the table, the RMSE of this method in the main test set is 47.67 cycles and the MAPE is 5.05%; the RMSE of the secondary test set is 158.63 cycles and the MAPE is 8.4%. The prediction errors of this method on the two test sets are smaller than those of other methods, proving that this method has higher accuracy and better prediction effect for early cycle life prediction.

[0129] Table 1 Comparison of cycle life prediction results of this method and other methods

[0130]

Claims

1. A method for predicting the early cycle life of lithium batteries based on degradation classification and improved CNN algorithm, characterized in that Here are the steps: Step 1: Divide the discharge data of the early degradation process into training sets and test sets, remove and correct the abnormal data in the training sets and test sets, extract statistical features from the cleaned data, and normalize the cleaned data; Step 2: Use the K-means clustering algorithm to classify the battery data obtained in step 1 and divide the batteries into long-life batteries and short-life batteries; Step 3: The discharge data of all batteries in the training set and the short-life batteries in the test set, as well as the statistical feature data from step 1, are input into the Domain Alignment CNN (AL-CNN) network. The AL-CNN network is trained to obtain the AL-CNN network for cycle life prediction of short-life batteries. The discharge data of all batteries in the training set and the statistical feature data from step 1 are input into the Squeeze-and-Excitation CNN (SE-CNN) network. The SE-CNN network is trained to obtain the SE-CNN network for cycle life prediction of long-life batteries. The AL-CNN network includes a CNN network composed of two-dimensional convolution modules Conv2d and a fully connected network, wherein: each two-dimensional convolution module Conv2d is connected to a maximum pooling layer, and the back end of the CNN network is connected to two fully connected networks for regression. The final prediction output is a 1×1 cycle life; The AL-CNN network is trained using the CORAL loss function for feature alignment. C s and C T The training set is the source domain, and the test set is the target domain; The SE-CNN network includes a CNN network consisting of a two-dimensional convolutional module Conv2d, an SE module inserted between two convolutional layers, and a fully connected network connected to the back end of the CNN network. The two-dimensional convolutional module Conv2d is connected to a maximum pooling layer, which is the same as the two-dimensional convolutional module Conv2d and maximum pooling layer in the AL-CNN network. The SE module compresses the data output by the first maximum pooling layer, converting each two-dimensional channel into a one-dimensional scalar. Then, an excitation operation is used to learn the weight of each channel. After learning the corresponding weight of each channel, the weight is reassigned to each channel, and a product operation is performed to output a new feature map. The new feature map is then passed through the convolution, pooling layers, and two fully connected networks, and the final predicted output is a 1×1 cycle life. Step 4: After processing the battery measurement data using steps 1 and 2, the long-life battery discharge data and features or the short-life battery discharge data and features are obtained, which are input into the SE-CNN network or the AL-CNN network respectively to obtain the cycle life prediction value of the long-life battery or the cycle life prediction value of the short-life battery.

2. The method for predicting early cycle life of lithium batteries based on degradation classification and improved CNN algorithm according to claim 1, characterized in that: The extracted statistical features are the variance of the difference between the discharge capacity-voltage curves of the 100th cycle and the 10th cycle. Var , minimum value Min , skewness Skew , kurtosis Kurt , the discharge capacity of the second cycle, the difference between the maximum discharge capacity of the first 100 cycles and the discharge capacity of the second cycle, is expressed as: 。 3. The method for predicting early cycle life of lithium batteries based on degradation classification and improved CNN algorithm according to claim 1, characterized in that: The K-means clustering algorithm is: select the number of clusters n , select from the input data n Cluster centers are found, the sum of the Euclidean distances between all input data and their cluster centers is calculated, and new cluster centers are found by minimizing the cost function. The above steps are repeated until the cluster centers tend to be stable. The cost function of this algorithm is defined as follows: in, x represents the input data, usually manually extracted features, It is i Centers of categories, C i It is i categories.

4. The method for predicting early cycle life of lithium batteries based on degradation classification and improved CNN algorithm according to claim 1, characterized in that: The training and testing steps of the AL-CNN network in step 3 are as follows: Step (1) Use the convolutional network to extract features from the input data: no padding operation is performed in the convolution layer, and the convolution process is: in, E Represents the input matrix, convolution kernel W The shape is , C represents the output feature map, is located in C middle Convolution features at ; The process of maximum pooling after the convolutional layer is: Among them, the size of the pooling window is , the step size is s , X and Y represent the input and output tensors respectively, are the spatial coordinates of the tensor, c represents the channel index, max Represents the maximum function; if the input shape is ,but Y The shape is ,in , Step (2) After the convolutional network extracts features, that is, at the end of the CNN network, the flattened features and statistical features learned by the CNN network are fused using a connection operation to form a new fused feature vector; then, two fully connected layers are added for regression, and the final prediction output is a 1×1 cycle life; The nonlinear activation function used after each convolutional layer and fully connected layer during training is the rectified linear unit ReLU; The weighted summation formula of the fully connected layer is as follows: in w i are the weights in the fully connected layer, b i is the bias, x i is the input neuron; Step (3) Parameter update: First, the CORAL loss function is used for feature alignment. The CORAL loss function is defined as follows: in, C S and C T are the two-dimensional feature covariance matrices of the source domain and the target domain, represent F norm; Loss function in AL-CNN network The corrections are as follows: in, y i is the actual cycle life, To predict the cycle life, n is the number of samples, lbd is the weight of the CORAL function, which is a hyperparameter that needs to be adjusted; in, is the change in weight, is the learning rate, is the loss function Partial derivatives with respect to weights; Through back propagation, the weights of each layer can be updated according to the target loss function, so that the parameters of the model are continuously updated, thereby extracting features corresponding to the task and improving prediction accuracy; Step (4) As the model parameters are updated, the network will continue to converge. The validation set and the "early stopping mechanism" are used to determine whether the model has converged. If it has converged, the trained network model is saved. Finally, the short-life battery discharge data and statistical features in the test set are input into the AL-CNN network to obtain the cycle life prediction results of the short-life battery.

5. The method for predicting early cycle life of lithium batteries based on degradation classification and improved CNN algorithm according to claim 1, characterized in that: The loss function of the SE-CNN network is different from that of the AL-CNN network. The loss function is the mean square error (MSE). The update of the model parameters is the same as that of the AL-CNN network.

6. The method for predicting early cycle life of lithium batteries based on degradation classification and improved CNN algorithm according to claim 1, characterized in that: The compression operation in the SE-CNN network converts the tensor U The shape of Compressed into ,in C is the number of channels; through the extrusion operation, each two-dimensional channel will become a global value. The extrusion operation is described as follows: Global spatial information is squeezed into the channel descriptor through global average pooling (GAP); the correlation between channels is fully utilized, and each two-dimensional channel is converted into a one-dimensional scalar, ignoring the internal redundant spatial distribution information.

7. The method for predicting early cycle life of lithium batteries based on degradation classification and improved CNN algorithm according to claim 1, characterized in that: The excitation operation uses the information obtained from the compression operation to learn the weight of each channel. The process of the excitation operation is as follows: Where, represents the ReLU function, refers to the sigmoid function, , , r It is a hyperparameter used to reduce the number of channels and reduce the amount of computation.

8. The method for predicting early cycle life of lithium batteries based on degradation classification and improved CNN algorithm according to claim 1, characterized in that: After learning the corresponding weights of each channel, the product reassigns the weights to each channel and performs a product operation to output a new feature map. The process is described as follows: in, u c represent U The c channels, s c Represents the weight of the corresponding channel.

9. A readable storage medium, characterized in that The readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for predicting the early cycle life of a lithium battery based on a degradation classification improved CNN algorithm as described in any one of claims 1 to 8.

10. A computer program product, characterized in that The invention comprises computer executable instructions, which, when executed, are used to implement the lithium battery early cycle life prediction method based on the degradation classification improved CNN algorithm as described in any one of claims 1 to 8.

Images