Battery life prediction method fusing transfer learning and long short-term memory network

Abstract

The invention discloses a battery life prediction method fusing transfer learning and a long short-term memory network, and the method comprises the steps: obtaining the multi-feature sequence data of a lithium battery under different working conditions, obtaining a trend component and a residual component through a sequence decomposition technology, carrying out the linear layer and batch normalization processing of the decomposed data, and carrying out the prediction of the life of the lithium battery. Extracting local context information of each component and completing feature fusion of the components by using a scalar long short-term memory network sLSTM residual error stacking module; introducing transfer learning TL to construct a TL-sLSTM model, carrying out transfer learning on an sLSTM parameter model learned from a source domain battery A to help to train target domain batteries B under different working conditions, and carrying out long-term prediction on the batteries; and in the output stage, the data is converted and subjected to distribution standardization, and then a single scalar is generated by using a full-connection layer to serve as a predicted value of the remaining life of the battery. According to the method, the problems of complex degradation process, nonlinearity and multi-working-condition data sparsity of a lithium battery prediction model in practical application are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery remaining life prediction, and particularly relates to a battery life prediction method integrating transfer learning and long short-term memory network. Background Art

[0002] Lithium-ion batteries are one of the most widely used rechargeable batteries at present, and are widely used in fields such as mobile devices, electric vehicles, and energy storage systems. By predicting the remaining service life of lithium batteries, the attenuation of the batteries can be understood in a timely manner, and corresponding measures can be taken to improve the reliability of the batteries, thereby optimizing the management and maintenance of the batteries and evaluating the performance and quality of the batteries. The attenuation and failure of lithium batteries are important factors affecting battery reliability. However, with the increase in the number of charge and discharge cycles, the capacity and performance of lithium-ion batteries will gradually decline, resulting in a shortened available time and service life of the batteries. Therefore, accurately predicting the remaining service life of lithium-ion batteries can optimize battery management, improve battery reliability, enhance battery performance evaluation, and save costs and resources, which is of great significance for the development and sustainability of the battery application field.

[0003] In the existing research on the prediction of the remaining life of lithium batteries, the lithium battery RUL prediction technology is mainly divided into two categories: model-based and data-driven. However, they generally have the following two problems:

[0004] 1) Limitations of the prediction model: Since the degradation rates of lithium-ion batteries in different life stages are different, there are certain limitations in using only local degradation data for life modeling.

[0005] 2) Sparsity of multi-condition prediction: For the empirical models and semi-empirical models of lithium battery RUL prediction, the parameters are greatly affected by the sparsity of condition data and battery parameters, so the generalization ability is poor. Summary of the Invention

[0006] In order to solve the deficiencies of the existing technology and achieve the purpose of improving the prediction accuracy of the remaining life of lithium batteries under different conditions, the present invention adopts the following technical solutions:

[0007] A battery life prediction method integrating transfer learning and long short-term memory network, comprising the following steps:

[0008] Step 101: Obtain the battery capacity time series data of the battery under different conditions;

[0009] Step 102: Decompose the battery capacity time series data under each condition to obtain a trend component and a residual component;

[0010] Step 103: After reducing the dimensions of the trend component and the residual component through a linear transformation layer (Linear Layer), perform normalization operations on the activations of each layer through batch normalization (Batch Normalization) to stabilize the distribution of the network input, so that the component dimensions are reduced and the internal covariate shift is resolved;

[0011] Step 104: Use a scalar long short-term memory network (scaler Long Short-Term Memory, sLSTM) residual stacking module for the processed trend component and residual component. Through shared causal convolution and Swish activation, extract the local context information of each component. Introduce a gating mechanism through the sLSTM block to achieve weighted interaction, and achieve residual stacking through residual connection to complete the feature fusion of the trend component and the residual component;

[0012] Step 105: Introduce transfer learning technology (Transfer Learning, TL) for different working conditions to construct a transfer learning scalar long short-term memory network TL-sLSTM model, which is divided into source domain lithium battery A data and target domain lithium battery B data. Pre-train the sLSTM model with the source domain lithium battery A data, and then fine-tune the model with the target domain lithium battery B data to make the model applicable to the target domain and perform long-term prediction on the capacity of this battery.

[0013] Further, in the step 102, for the input sequence of the lithium battery data set with length L and feature number m the trend component and the residual component are extracted by applying a learnable moving average to each feature through one-dimensional convolution. The formula is as follows:

[0014]

[0015]

[0016] where represents the trend component, represents the average pooling operation, represents the padding operation, represents the input sequence, represents the residual component.

[0017] Furthermore, for the sLSTM block in step 104, the input gate (i), forget gate (f), output gate (o), and cell update gate (z) are respectively processed through a block diagonal linear layer, and each head independently learns weights, thus realizing the iterative fusion of different components; after the trend and seasonal components pass through the sLSTM block, they are combined with the Residual Connection to ensure that the model can capture the multi-scale features of the input components while avoiding information loss; the GroupNorm layer is used to normalize the features, dividing the channels into groups to calculate the normalization statistics; the dimensions of the features are adjusted through the projection up and down of the gated multi-layer perceptron (MLP), and a non-linear response is introduced through the GeLU activation function to enable the model to learn the complex mapping relationship of the battery degradation features; the residual stack is realized by introducing skip connections in the network, directly adding the input to the output to form a residual connection; through the residual stack, sLSTM can construct a deep network, improving the expression ability of the model and the ability to process complex long sequence data.

[0018] Furthermore, the scalar long short-term memory network in step 104 introduces exponential gates and normalization and stabilization operations on the long short-term memory network block, and the formula is as follows:

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] Among them, the core state (cell-state) represents the cell state at time step t, represents the cell state at time step t-1, (normalizer state) represents the state of normalizing the current time step t by combining the previous time step t-1, The (hidden-state) represents the hidden state at the current time step, which is calculated from the output gate and the normalized cell state. Computing; Gating mechanism The (input-gate), the (forget-gate), and the (output-gate) represent the input gate, the forget gate, and the output gate respectively; the candidate value The (cell-input) represents the candidate cell state input, represents the cell input activation function, represents the gate activation function. All gate activation functions are sigmoid functions, that is, σ(x) = 1 / (1 + exp(-x)), where exp represents the exponential function; the original input of the gating 、 、 、 are obtained through linear transformation; the weight vectors 、 、 、 represent the input weights between the input and the cell input, the input gate, the forget gate, and the output gate respectively. The weights 、 、 and represent the recurrent weights between the hidden state and the cell input, the input gate, the forget gate, and the output gate respectively. 、 、 and represent the corresponding bias terms respectively.

[0027] When migrating the original LSTM gating technology, that is, input or hidden state-dependent gating with bias terms, to the new architecture, the exponential activation function may lead to large numerical values and cause overflow. Therefore, the present invention adopts an additional stable state to stabilize the update gate and the forget gate:

[0028]

[0029]

[0030]

[0031] Among them, represents the stable state at the previous moment, represents the stable input gate, represents the stable forget gate.

[0032] Furthermore, memory mixing is performed through cyclic connections, allowing information sharing between different memory units.

[0033] Furthermore, with the multi-head structure, each head has its own memory mixing, but there is no cross-head memory mixing between the heads. This design provides a new way of memory mixing for sLSTM.

[0034] Furthermore, during the training process of step 105, the sLSTM model learns the common and private attributes of battery cells from all the data in the source domain. Then, the target domain training set is used to fine-tune the model to ensure that the model can capture the private attributes of the target domain battery cells and adapt to the influence of different stress conditions on battery performance.

[0035] Furthermore, the model fine-tuning in step 105 is to update the parameters of the gated multi-layer perceptron (MLP) and the output layer, while freezing the parameters of the early convolutional layers and the scalar long short-term memory network (TL-sLSTM) blocks;

[0036] First, initialize and load the parameters of the pre-trained model, freeze the parameters of the convolutional layers and the scalar long short-term memory network (TL-sLSTM) blocks, do not update their weights, and only unfreeze the parameters of the gated multi-layer perceptron (MLP) and the output layer;

[0037] Then, perform forward propagation. The frozen convolutional layers and the scalar long short-term memory network (TL-sLSTM) blocks extract general features from the input; the gated multi-layer perceptron (MLP) performs high-dimensional feature transformation on these general features and extracts features adapted to the target domain through activation functions and projection layers; in addition, perform backward propagation, calculate the prediction error of the output (such as the mean squared error, MSE), and only calculate the gradients and update the parameters of the unfrozen gated multi-layer perceptron (MLP) and the output layer.

[0038] Fine-tuning can be completed through multiple iterations. The specific number of training epochs depends on the size of the target domain dataset and the convergence of the model. Since only a small number of parameters are updated, fine-tuning is generally faster than training from scratch.

[0039] Furthermore, in the output stage of step 105, the data is further transformed through a linear layer, and then the features are distribution-normalized through instance normalization (IN) so that the output of each channel has a consistent distribution. Finally, a single scalar is generated using a fully connected layer as the predicted value of the remaining battery life.

[0040] Furthermore, the instance normalization runs independently on each channel of the time series, normalizes the data within each channel so that its mean is 0 and its variance is 1, and the formula is as follows:

[0041]

[0042] Among them, x represents the input feature, μ(x) represents the mean of the feature, and σ(x) represents the standard deviation of the feature map.

[0043] Through instance normalization, TL-sLSTMTime independently normalizes the data within each channel, improving the stability and adaptability of the model in the face of data distribution changes, and providing an efficient solution for predicting the remaining life of lithium batteries.

[0044] The advantages and beneficial effects of the present invention are as follows:

[0045] The present invention aims at obtaining multi-feature sequence data of lithium batteries under different working conditions, performing sequence decomposition on each feature sequence data, extracting trend components and residual components, respectively performing linear layer and batch normalization processing, using a stable long short-term memory network sLSTM residual stacking module to extract local context information of each component and complete the feature fusion of the components; introducing transfer learning TL to construct a TL-sLSTM model, and using the sLSTM parameter model learned from the source domain battery A by transfer learning to help train the target domain battery B under different working conditions for long-term prediction of this battery; in the output stage, the data is transformed and distribution-standardized and then a fully connected layer is used to generate a single scalar as the predicted value of the remaining life of the battery; thus forming a complete model TL-sLSTMTime. The present invention can efficiently fuse multi-dimensional features in the above process, reduce the problem of gradient disappearance in deep models, and allow the model to more effectively capture long-term dependencies, learn battery degradation features, and adapt to differences in data distribution under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the flowchart of the method of the embodiment of the present invention.

[0047] Figure 2 is a schematic diagram of decomposing each feature sequence data in the embodiment of the present invention.

[0048] Figure 3 is a schematic diagram of the scalar long short-term memory network (sLSTM) residual stacking module in the embodiment of the present invention.

[0049] Figure 4 is a schematic diagram of introducing transfer learning technology TL to construct a TL-sLSTM model in the embodiment of the present invention.

[0050] Figure 5 is a schematic diagram of fine-tuning the TL-sLSTM model in the embodiment of the present invention.

[0051] Figure 6It is a schematic diagram of the structure of the complete model TL-sLSTMTime constructed in the embodiments of the present invention. Detailed implementation manners

[0052] The following will describe in detail the specific implementation manners of the present invention with reference to the accompanying drawings. It should be understood that the specific implementation manners described herein are only for the purpose of illustrating and explaining the present invention, and are not used to limit the present invention.

[0053] In the fields of mobile devices, electric vehicles, energy storage systems, etc., the remaining service life of lithium batteries has received extensive attention. In recent years, scholars have conducted a large number of research works on the prediction of the remaining life of lithium batteries. The existing lithium battery RUL prediction technologies are mainly divided into two categories: model-based and data-driven.

[0054] However, the current prediction models generally have the following two problems: 1) Prediction model limitation: Since the degradation rates of lithium-ion batteries in different life stages are different, using only local degradation data for life modeling has certain limitations. 2) Sparse multi-condition prediction: For the empirical models and semi-empirical models of lithium-ion battery RUL prediction, the parameters are greatly affected by the sparsity of condition data and battery parameters, so the generalization ability is poor.

[0055] To address the above problems, the present application proposes a battery life prediction method that combines transfer learning and long short-term memory networks. By using sequence decomposition technology, the trend component and the residual component are obtained. The scalar long short-term memory network sLSTM residual stacking module is used to extract the local context information of each component and complete the feature fusion of the components; transfer learning TL (Transfer Learning) is introduced to construct the TL-sLSTM model, and the sLSTM parameter model learned from the source domain battery A by transfer learning is used to help train the target domain battery B under different conditions, and long-term prediction is performed on this battery; in the output stage, the data is transformed and distribution standardized, and then a fully connected layer is used to generate a single scalar as the predicted value of the remaining life of the battery. It solves the problems of complex degradation process, non-linearity, and sparse multi-condition data of the lithium battery prediction model in practical applications. As Figure 1 shown, the lithium battery life prediction method specifically includes the following steps:

[0056] Step 101: Obtain multi-feature sequence data of the lithium battery under different conditions.

[0057] For different conditions of the lithium battery, such as charge and discharge rate, working temperature, etc., which result in large differences in the degradation paths of the battery, obtain multi-feature sequence data that can reflect the lithium battery under different conditions.

[0058] Step 102: Decompose each feature sequence data by using sequence decomposition technology to obtain a trend component and a residual component.

[0059] As shown Figure 2 in the figure, for a lithium battery dataset with a length of L and a feature number of m, there is an input sequence, and the learnable moving average is applied to each feature through 1-D convolution to extract the trend component and the residual component; the calculation formula is as follows:

[0060] (1)

[0061] (2)

[0062] Among them, represents the trend component, represents the average pooling operation, represents the padding operation, represents the input sequence, and represents the residual component.

[0063] Step 103: Perform linear layer and batch normalization processing on the trend component and the residual component respectively, so that the components are dimension-reduced and the internal covariate shift is solved; the components after decomposing the lithium battery capacity time series data are dimension-reduced through the linear transformation layer (Linear Layer), and the activation of each layer is normalized through batch normalization (Batch Normalization) to stabilize the distribution of the network input.

[0064] Step 104: Use the scalar long short-term memory network (sLSTM) residual stacking module for the processed trend component and residual component, and extract the local context information of each component through the shared Swish activation and causal convolution; realize weighted interaction through the gating mechanism, and complete the feature fusion of the trend component and the residual component by combining the residual connection.

[0065] As Figure 3 shown, for the scalar long short-term memory network (sLSTM), it includes:

[0066] The sLSTM structure introduces the exponential gate as well as normalization and stabilization on the original LSTM:

[0067] Cell state cell state (3)

[0068] Normalizer state normalizer state (4)

[0069] Hidden state hidden state (5)

[0070] Cell input (6)

[0071] Input gate (7)

[0072] Forget gate (8)

[0073] Output gate (9)

[0074] Among them, the core state represents the cell state at time step t, represents the cell state at time step t - 1, represents combining the normalized state at the current time step t with the previous time step t - 1, (represents the hidden state at the current time step, calculated from the output gate and the normalized cell state ; The gating mechanism , , respectively represent the input gate, forget gate, output gate; The candidate value represents the candidate cell state input, represents the cell input activation function, represents the gate activation function, and all gate activation functions are sigmoid functions, that is f(x)=1 / (1 + exp(-x)), where exp represents the exponential function; The original input of the gating , , , are obtained through linear transformation; The weight vectors , , , respectively represent the input between the unit input, input gate, forget gate, output gate and the input weight vectors, and the weights , , and respectively represent the hidden state and the recurrent weights between the cell input, input gate, forget gate and output gate, , , and respectively represent the corresponding bias terms.

[0075] Migrate the original LSTM gating technology, i.e., input and / or hidden-dependent gating with bias terms, to the new architecture. The exponential activation function can lead to large values and cause overflow. Therefore, the present invention employs an additional state to stabilize the update gate and the forget gate:

[0076] Stabilizer state (10)

[0077] Stabilizing input gate (11)

[0078] Stabilizing forget gate (12)

[0079] Meanwhile, the following structures are also supported:

[0080] Memory Mixing: sLSTM allows memory mixing through recurrent connections, which is not possible in the original LSTM; this new memory mixing technology allows sLSTM to share information between different memory units;

[0081] Multi-Head Structure: sLSTM can have multiple heads, each with its own memory mixing, but there is no cross-head memory mixing between the heads; this design provides a new way of memory mixing for sLSTM;

[0082] For Figure 3The sLSTM residual stacking module extracts local features of each component through the shared Swish activation and causal convolution Conv4 for the trend and seasonal components, providing local context information for subsequent fusion. For the causal convolution operation, it ensures that information from future time steps is not introduced in the convolution calculation. The sLSTM block is used, and for the input (i), forget (f), output (o) gates, and cell update (z), they are processed through block diagonal linear layers respectively, and each head independently learns weights to achieve iterative fusion of different components. After the trend and seasonal components pass through the sLSTM block, they are combined with the residual connection to ensure that the model can capture multi-scale features of the input components while avoiding information loss. The GroupNorm layer is used to normalize the features, dividing the channels into groups to calculate the normalization statistics. The dimension of the features is adjusted through the projection up and down on the gated MLP (multi-layer perceptron), and a non-linear response is introduced through the GeLU activation function, enabling the model to learn the complex mapping relationship of battery degradation features. The residual stacking is achieved by introducing skip connections in the network, directly adding the input to the output to form a residual connection. Through residual stacking, sLSTM can build a deep network, improving the model's expressive ability and the ability to process complex long sequence data.

[0083] Step 105: Introduce transfer learning technology (Transfer Learning, TL) for different working conditions to construct a TL-sLSTM model. It is divided into source domain lithium battery A data and target domain lithium battery B data. The sLSTM model is pre-trained with the source domain lithium battery A data, and then the target domain lithium battery B data is used to fine-tune the model to make it applicable to the target domain for long-term prediction of the capacity of this battery.

[0084] In one embodiment, as Figure 4 shown, first, the sLSTM model is trained with the source domain data, which is specifically for the capacity prediction of the source domain battery A. This pre-trained model can then be used as the basis for transfer learning and fine-tuned through the training set of the target domain battery B data to adapt to the characteristics of the target domain battery cell. During the training process, the sLSTM model learns the common and private attributes of the battery cell through all the data in the source domain. Then, the model is fine-tuned using the training set of the target domain to ensure that the model can capture the private attributes of the target domain battery cell and adapt to the impact of different stress conditions on battery performance.

[0085] Based on the above embodiment, as Figure 5As shown in the figure, fine-tuning is performed on the sLSTM model. The focus of fine-tuning is to update the parameters of the gated MLP and the output layer, while freezing the parameters of the early convolutional layers and the sLSTM layer. First, initialize and load the parameters of the pre-trained model, freeze the parameters of the convolutional layer and the sLSTM layer, and do not update their weights. Only unfreeze the parameters of the gated MLP and the output layer. Secondly, perform forward propagation. The frozen convolutional layer and sLSTM layer extract general features from the input. The gated MLP performs high-dimensional feature transformation on these general features and extracts features adapted to the target domain through the activation function and the projection layer. In addition, perform backward propagation, calculate the prediction error of the output (such as the mean squared error MSE), and only calculate the gradients and update the parameters of the unfrozen gated MLP and the output layer. Fine-tuning can be completed through multiple iterations. The specific number of training epochs depends on the size of the target domain dataset and the convergence of the model. Since only a small number of parameters are updated, fine-tuning is generally faster than training from scratch.

[0086] In the output stage, the data is further transformed through a linear layer. Instance normalization (IN) normalizes the distribution of the features to ensure that the output of each channel has a consistent distribution. Finally, a fully connected layer is used to generate a single scalar as the predicted value of the remaining battery life, thus forming a complete model named TL-sLSTMTime.

[0087] In one embodiment, as Figure 6 shown, the model input is multi-feature sequence data of a lithium-ion battery. The trend component and the residual component in the signal are extracted through a sequence decomposition module. Then, the data is processed through a linear layer and a batch normalization layer to enhance the expression ability of the features and accelerate the training convergence. In the core part of the model, an sLSTM residual block is designed, and transfer learning technology is combined to optimize the feature extraction ability. The sLSTM network can effectively capture the long-term dependencies and non-linear changes in the battery operation data by using the gating mechanism and the time dimension information. The transfer learning module realizes knowledge transfer to the target task through pre-training on the source dataset, thereby improving the prediction ability of the model in the small sample scenario. After the last linear layer at the output end of the sLSTM, the data is further transformed to prepare for the final output through instance normalization.

[0088] Instance normalization (IN) runs independently on each channel of the time series. It normalizes the data within each channel so that its mean is 0 and its variance is 1. The formula for instance normalization of a given feature is as follows:

[0089] (13)

[0090] where x represents the input feature, μ(x) represents the mean of the feature, and σ(x) represents the standard deviation of the feature map.

[0091] Through instance normalization, TL-sLSTMTime independently normalizes the data within each channel, improving the stability and adaptability of the model in the face of data distribution changes and providing an efficient solution for predicting the remaining useful life of lithium batteries.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery life prediction method that combines transfer learning and long short-term memory network, characterized in that It includes the following steps: Step 101: Obtain the battery capacity time series data of the battery under different working conditions; Step 102: Decompose the battery capacity time series data under each working condition to obtain the trend component and the residual component; Step 103: After reducing the dimensions of the trend component and the residual component through a linear transformation layer, perform normalization operations on the activations of each layer through batch normalization; Step 104: Use the scalar long short-term memory network residual stacking module for the processed trend component and residual component. Through shared causal convolution and activation, extract the local context information of each component. Introduce a gating mechanism through the scalar long short-term memory network block to achieve weighted interaction, and achieve residual stacking through residual connection to complete the feature fusion of the trend component and the residual component; Step 105: Introduce transfer learning for different working conditions to construct a transfer learning scalar long short-term memory network model. Pre-train the model with the source domain battery data and fine-tune the model with the target domain battery data to make the model applicable to the target domain and perform long-term prediction on the capacity of the battery.

2. The battery life prediction method integrating transfer learning and long short-term memory network according to claim 1, characterized in that: In the said step 102, for the input sequence of the lithium battery data set with length L and number of features m , the trend component and the residual component are extracted by applying a learnable moving average to each feature through one-dimensional convolution. The formula is as follows: Among them, represents the trend component, represents the average pooling operation, represents the padding operation, represents the input sequence, represents the residual component.

3. A battery life prediction method integrating transfer learning and long short-term memory network according to claim 1, characterized in that: For the scalar long short-term memory network block in step 104, the input gate, forget gate, output gate, and cell update gate are respectively processed through a block diagonal linear layer, and each head independently learns the weights; after the component passes through the scalar long short-term memory network block, it is combined with the residual connection; use a group normalization layer to normalize the features, divide the channels into groups to calculate the normalization statistics; adjust the dimensions of the features through the projection up and down of the gated multi-layer perceptron, and introduce a non-linear response through the activation function to learn the complex mapping relationship of the battery degradation features; the residual stacking is achieved by introducing a skip connection in the network, directly adding the input to the output to form a residual connection.

4. A battery life prediction method integrating transfer learning and long short-term memory network according to claim 1, characterized in that: The scalar long short-term memory network in step 104 introduces an exponential gate and normalization and stabilization operations on the long short-term memory network block. In particular, the input gate and the forget gate have exponential activation functions, and the formula process is as follows: Among them, the core state represents the cell state at time step t, represents the cell state at time step t - 1, represents the state that combines and normalizes the current time step t with the previous time step t - 1, represents the hidden state at the current time step, which is calculated by the output gate and the normalized cell state ; gating mechanism 、 、 represent the input gate, forget gate, and output gate respectively; candidate value represents the candidate cell state input, represents the cell input activation function, represents the gate activation function, and all gate activation functions are sigmoid functions, that is (x)=1 / (1 + exp(-x)), where exp represents the exponential function; the original input of gating 、 、 、 are obtained through linear transformation; weight vectors 、 、 、 represent the input and the input weight vectors between the cell input, input gate, forget gate, and output gate respectively, and the weights 、 、 and represent the recurrent weights between the hidden state and the cell input, input gate, forget gate, and output gate respectively, 、 、 and represent the corresponding bias terms respectively. Adopt an additional stable state Stabilize the update gate and the forget gate: Among them, represents the stable state at the previous moment, represents the stable input gate, represents the stable forget gate.

5. A battery life prediction method integrating transfer learning and long short-term memory network according to claim 4, characterized in that: Perform memory mixing through cyclic connection, allowing information sharing between different memory units.

6. The battery life prediction method integrating transfer learning and long short-term memory network according to claim 4, characterized in that: Through a multi-head structure, each head has its own memory mixing, but there is no cross-head memory mixing between the heads.

7. A battery life prediction method integrating transfer learning and long short-term memory network according to claim 1, characterized in that: During the training process of step 105, the model learns the common attributes and private attributes of the battery cells through all the data in the source domain. Then, use the training set in the target domain to fine-tune the model to ensure that the model can capture the private attributes of the target domain battery cells and adapt to the impact of different stress conditions on the battery performance.

8. A battery life prediction method integrating transfer learning and long short-term memory network according to claim 1, characterized in that: The model fine-tuning in step 105 is to update the parameters of the gated multi-layer perceptron and the output layer, while freezing the early convolution and scalar long short-term memory network blocks; First, initialize and load the parameters of the pre-trained model, freeze the parameters of the convolution and scalar long short-term memory network blocks, do not update their weights, and only unfreeze the parameters of the gated multi-layer perceptron and the output layer; Then, in the forward propagation, the frozen convolutional and scalar long short-term memory network blocks extract general features from the input; the gated multi-layer perceptron performs high-dimensional feature transformation on these general features, and extracts features adapted to the target domain through the activation function and the projection layer; in addition, backward propagation is also performed to calculate the prediction error of the output, and gradients are calculated and updated only for the parameters of the unfrozen gated multi-layer perceptron and the output layer.

9. A battery life prediction method integrating transfer learning and long short-term memory network according to claim 1, characterized in that: In the output stage of step 105, the data is further transformed through a linear layer, and then the features are distribution-normalized through instance normalization so that the output of each channel has a consistent distribution. Finally, a single scalar is generated using a fully connected layer as the predicted value of the remaining battery life.

10. A battery life prediction method integrating transfer learning and long short-term memory network according to claim 9, characterized in that: The instance normalization runs independently on each channel of the time series, normalizes the data within each channel so that its mean is 0 and its variance is 1. For a given feature map, the formula for instance normalization is as follows: where \(x\) represents the input feature map, \(\mu(x)\) represents the mean of the feature map, and \(\sigma(x)\) represents the standard deviation of the feature map.

Images