Battery pack health state prediction method and device

Abstract

The invention discloses a battery pack health state prediction method and device, and the method comprises the steps: building a battery pack three-dimensional matrix based on collected historical data of a plurality of cells in a battery pack, carrying out the sliding window processing of the battery pack three-dimensional matrix, and generating model input data; determining a battery pack health state prediction model, wherein the battery pack health state prediction model comprises an SE-TCN network, a Performer encoder and a space-time fusion layer; and inputting the model input data into the battery pack health state prediction model to obtain a health state prediction value of the battery pack. According to the method, the time dependency characteristics of the time sequence data of the plurality of battery cells in the battery pack and the global dependency relationship among the battery cells are respectively captured, and the time sequence characteristics and the space correlation characteristics are fused, so that the model can capture the complex interaction relationship among the plurality of battery cells; and the accuracy and robustness of battery pack health state prediction are greatly improved.

Description

Technical Field

[0001] The present application relates to the field of battery detection, and more specifically, to a method and device for predicting the health status of a battery pack. Background Art

[0002] Accurately predicting the battery health status of a battery pack is critical to optimizing battery usage, extending battery life, and ensuring safety. However, the complexity and dynamic characteristics of battery management systems make it challenging to accurately predict the health status of the entire battery pack.

[0003] Existing technologies mainly obtain the health status of a single cell by processing the data of each cell separately, or by jointly evaluating it with the battery state of charge. However, in actual applications, as the scale of batteries expands, battery packs containing multiple cells are widely used in new energy vehicles and energy storage systems. There are obvious defects in predicting each cell separately:

[0004] First, the mutual influence between cells is ignored. There are mutual influences between multiple cells due to factors such as temperature, current, and mechanics. Predicting each cell individually may result in large misjudgments and fail to fully consider the complex relationship between cells.

[0005] Second, there is a model redundancy problem. Modeling each battery cell separately requires building and maintaining an independent model for each battery cell, which not only increases the complexity of the system but also brings a heavy computational burden.

[0006] Based on this, this application proposes a battery pack health status prediction solution, which can capture the complex interaction relationship between multiple battery cells and avoid the model redundancy problem. Summary of the invention

[0007] In view of this, the present application provides a method and device for predicting the health status of a battery pack, which captures the complex interaction relationship between multiple battery cells through time-space fusion modeling, improves the accuracy and robustness of the battery pack health status prediction, and avoids the problem of model redundancy.

[0008] A method for predicting a battery pack health state, comprising:

[0009] A three-dimensional matrix of the battery pack is constructed based on the collected historical data of multiple cells in the battery pack, and a sliding window processing is performed on the three-dimensional matrix of the battery pack in combination with the currently updated window size and sliding step size to generate model input data, and the window size and the sliding step size are dynamically adjusted based on the prediction performance;

[0010] Determine a battery pack health status prediction model, the battery pack health status prediction model includes an SE-TCN network, a Performer encoder and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time dependency characteristics of the time series data of multiple cells in the battery pack to generate time series features, the Performer encoder is used to capture the global dependency between the cells in the battery pack to generate spatial correlation features, and the spatiotemporal fusion layer obtains a health status prediction value by fusing the time series features and the spatial correlation feature mapping;

[0011] The model input data is input into the battery pack health state prediction model to obtain a health state prediction value of the battery pack output by the battery pack health state prediction model.

[0012] Optionally, the three-dimensional matrix shape of the battery pack is:

[0013]

[0014] The model input data shape is:

[0015]

[0016]

[0017] Among them, S is the total sample size or the number of time steps, B is the number of cells, F is the number of features, N is the number of samples, the window size is W, and the sliding step size is Δ.

[0018] Optionally, the SE-TCN network includes multiple parallel and parameter-sharing SE-TCN structures, and the SE-TCN structure is composed of a time convolution module and a SE module;

[0019] The time convolution module extracts the time-dependent characteristics of the time series data of the corresponding battery cell through dilated convolution;

[0020] The SE module performs global average pooling on the channels through a compression operation, extracts the correlation between channels through an excitation operation, recalibrates the channel weights based on the correlation, and generates time series features in combination with the time-dependent characteristics.

[0021] Optionally, the dilated convolution formula is:

[0022]

[0023] The calculation formula for global average pooling of channels through compression operation is:

[0024]

[0025] The calculation formula for extracting the correlation between channels through the excitation operation is:

[0026]

[0027]

[0028] The calibration calculation formula for recalibrating channel weights is:

[0029]

[0030] in, is the dilated convolution output sequence, is the input sequence, is the convolution kernel size, is the convolution kernel at position The value at is the expansion rate, is the output time point, is the output value of channel c after global average pooling, L is the sequence length, is the activation value of channel c at position i, z is the weight vector obtained after the excitation operation, s is the vector composed of the values ​​obtained by the global average pooling operation of each channel, C is the total number of channels, is the ReLU activation function, is the Sigmoid activation function, and is the weight matrix of the two fully connected layers in the SE module, is the eigenvalue of the channel after recalibrating the channel weights, is the weight value corresponding to channel c in the weight vector z, is the eigenvalue of the original channel c.

[0031] Optionally, the Performer encoder adopts a linear multi-head attention mechanism, wherein the linear multi-head attention mechanism includes random feature mapping, linearized self-attention calculation, and multi-head self-attention output;

[0032] The random feature mapping is used to map the input vector to the low-dimensional feature space to form a low-dimensional feature space vector;

[0033] The linearized self-attention calculation is used to calculate the attention weights between input elements, and weighted sum the value matrix according to the attention weights to obtain an output result based on the attention mechanism;

[0034] The multi-head self-attention output is used to splice the multi-head output results, and obtain the comprehensive output result through the linear transformation realized by the weight matrix.

[0035] Optionally, the calculation formula of the random feature mapping is:

[0036]

[0037] The linearized self-attention calculation formula is:

[0038]

[0039]

[0040] The calculation formula of the multi-head self-attention output is:

[0041]

[0042]

[0043] in, is a low-dimensional feature space vector, x is the input feature of each battery cell, is a random vector sampled from a specific distribution, M is the dimension of the random feature, is the linearized self-attention calculation result, is the query matrix, is the key matrix, is the value matrix, The coefficient used to normalize the attention score, 1 is a full 1 vector, Output result of multi-head self-attention, for The output result of the head, Wo is the weight matrix.

[0044] Optionally, the Performer encoder adopts residual connection and performs layer normalization operation on each layer;

[0045] The calculation formula for each layer in the Performer encoder is:

[0046]

[0047] in, is the layer calculation result, LayerNorm is the layer normalization calculation, is the output result of multi-head self-attention, and H is the hidden state vector of each battery cell at the last time step.

[0048] Optionally, the Performer encoder further includes an embedding layer;

[0049] The embedding layer generates a position embedding vector for representing the unique identity information of the battery cell according to the battery cell position of each battery cell.

[0050] Optionally, the process of determining the battery pack health status prediction model includes:

[0051] Acquire sample model input data, where the sample model input data is generated by performing sliding window processing on a three-dimensional matrix of a sample battery pack constructed based on sample historical data of battery cells;

[0052] Input the sample model input data into the constructed battery pack health status prediction model, train the model using the determined loss function and the selected optimizer, and iteratively update the model parameters until the preset training stop condition is met, wherein the training stop condition includes reaching the maximum number of iterations, the loss function value is less than the preset threshold, or the accuracy of the validation set reaches the preset value;

[0053] The loss function is:

[0054]

[0055] The update rule of the optimizer to update the model parameters is:

[0056]

[0057] in, For the The true health status value of the sample battery pack, is the predicted value of the health status of the model, N is the total number of samples, are model parameters, and are the first-order and second-order moment estimates, respectively, is the learning rate, is the weight decay coefficient, is a preset constant.

[0058] A battery pack health status prediction device, comprising:

[0059] A data processing unit, configured to construct a three-dimensional matrix of a battery pack based on the collected historical data of a plurality of cells in the battery pack, and to perform sliding window processing on the three-dimensional matrix of the battery pack in combination with a currently updated window size and a sliding step size to generate model input data, wherein the window size and the sliding step size are dynamically adjusted based on the predicted performance;

[0060] A model determination unit, used to determine a battery pack health state prediction model, wherein the battery pack health state prediction model includes an SE-TCN network, a Performer encoder, and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time dependency characteristics of the time series data of multiple cells in the battery pack to generate time series features, the Performer encoder is used to capture the global dependency between the cells in the battery pack to generate spatial correlation features, and the spatiotemporal fusion layer obtains a health state prediction value by fusing the time series features and the spatial correlation feature mapping;

[0061] The model prediction unit is used to input the model input data into the battery pack health state prediction model to obtain the health state prediction value of the battery pack output by the battery pack health state prediction model.

[0062] It can be seen from the above technical solutions that a method and device for predicting the health status of a battery pack provided in an embodiment of the present application first constructs a three-dimensional matrix of the battery pack based on the historical data of multiple cells in the battery pack collected, and performs sliding window processing on the three-dimensional matrix of the battery pack in combination with the currently updated window size and sliding step size to generate model input data, and the window size and the sliding step size are dynamically adjusted based on the prediction performance. Then, a battery pack health status prediction model is determined, and the battery pack health status prediction model includes an SE-TCN network, a Performer encoder, and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time-dependent characteristics of the time series data of multiple cells in the battery pack to generate time series features, and the Performer encoder is used to generate spatial correlation features by capturing the global dependency between the cells in the battery pack, and the spatiotemporal fusion layer obtains the health status prediction value by fusing the time series features and the spatial correlation feature mapping. Finally, the model input data is input into the battery pack health status prediction model to obtain the health status prediction value of the battery pack output by the battery pack health status prediction model.

[0063] This application can make full use of the data of multiple cells and is no longer limited to the isolated analysis of a single cell. By combining the SE-TCN network, the Performer encoder and the spatiotemporal fusion layer, the time-dependent characteristics of the time series data of multiple cells in the battery pack and the global dependencies between the cells are captured respectively, and the time series features and spatial correlation features are fused, so that the model can capture the complex interaction between multiple cells, and then make a more comprehensive and reasonable estimate, which greatly improves the accuracy and robustness of the battery pack health status prediction. This modeling method that comprehensively considers spatiotemporal factors is more in line with the actual working conditions of the battery pack, and can provide a more reliable basis and guidance for battery management and use. At the same time, by constructing a joint model, the previous way of modeling each cell separately is changed, which effectively reduces the complexity of the model. A large amount of repetitive independent model establishment and maintenance work is avoided, thereby significantly improving computing efficiency and reducing system resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0065] Figure 1 A flow chart of a method for predicting the health status of a battery pack disclosed in this application;

[0066] Figure 2 A schematic diagram of the structure of a battery pack health status prediction model disclosed in this application;

[0067] Figure 3 A schematic diagram of the analysis process of a battery pack health status prediction model disclosed in this application;

[0068] Figure 4 A schematic diagram of the analysis process of another battery pack health status prediction model disclosed in this application;

[0069] Figure 5 A schematic diagram of an example test result disclosed in this application;

[0070] Figure 6 This is a structural block diagram of a battery pack health status prediction device disclosed in this application. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0072] Next, the present application scheme is introduced. The present application proposes the following technical scheme, please see below for details.

[0073] Figure 1 This is a flow chart of a method for predicting the health status of a battery pack disclosed in an embodiment of the present application. The method for predicting the health status of a battery pack can be applied to a device for predicting the health status of a battery pack, such as Figure 1 As shown, the method may include:

[0074] Step S1: construct a three-dimensional matrix of the battery pack based on the collected historical data of multiple cells in the battery pack, and perform sliding window processing on the three-dimensional matrix of the battery pack in combination with the currently updated window size and sliding step size to generate model input data.

[0075] Specifically, various historical data of multiple cells in the battery pack are collected, including timestamp (t), voltage (V), current (I), temperature (T), number of cycles (N_cycle), cumulative discharge capacity (Q_discharge), and historical health status information. These data comprehensively reflect the status and performance of the cells at different time points and usage conditions. For example, the timestamp records the specific moment of data collection, the voltage and current reflect the electrical characteristics of the cell at work, the temperature is related to the working environment and heating conditions of the cell, the number of cycles and cumulative discharge capacity reflect the degree of use of the cell, and the health status information is an important indicator to measure the health of the cell, which will change with the use and aging of the cell. These data together constitute a three-dimensional matrix, namely the three-dimensional matrix of the battery pack, and each element of the matrix corresponds to a certain characteristic value of a specific cell at a specific time step. It should be noted that collecting multiple historical data of multiple cells in a battery pack, including collecting multiple historical data of each cell in the battery pack, can improve data integrity, help the model capture the differences between each cell, fully reflect the performance of the battery pack under different working conditions, and improve prediction accuracy. Of course, in order to improve computing efficiency, you can select multiple historical data of multiple cells in key positions in the battery pack. For example, you can select cells close to the crossbeam, which are subject to greater temperature changes and mechanical stress and are more critical to predicting the health status. You can also select cells in high current paths. Some cells may bear higher current loads and are more susceptible to performance degradation, so they are equally important for predicting the health status.

[0076] The three-dimensional matrix shape of the battery pack is:

[0077]

[0078] The model input data is generated using a sliding window method, with a window size of W and a sliding step size of Δ.

[0079] The window size W determines the data range of each input model, that is, considering multiple data points of the battery cell in the past period of time as a whole input. For example, if W is 10, then the data input to the model each time is all the feature data of the battery cell in the past 10 time steps.

[0080] The sliding step size Δ controls the interval of window movement. A smaller sliding step size will generate more samples and better data continuity, but it will also increase the amount of calculation; a larger sliding step size will reduce the number of samples and may lose some intermediate information, but the calculation efficiency will be improved.

[0081] The model input data shape is:

[0082]

[0083]

[0084] Among them, S is the total sample size or the number of time steps, B is the number of cells, F is the number of features, N is the number of samples, the window size is W, and the sliding step size is Δ.

[0085] At the same time, the window size and the sliding step size are dynamically adjusted based on the prediction performance. Different battery data characteristics and model structures may require different window sizes and sliding step sizes to achieve the best prediction effect. During the model training and verification process, by constantly trying different combinations of window sizes and sliding step sizes, observing the performance indicators of the model on the verification set (such as mean square error, mean absolute error, etc.), select the window size and sliding step size that optimize the performance indicators. For example, you can first set an initial window size and sliding step size range, and then gradually adjust them. After each adjustment, retrain and verify the model, record the performance indicators, and finally determine the best parameter combination.

[0086] Step S2: Determine a battery pack health status prediction model.

[0087] Specifically, the battery pack health status prediction model includes an SE-TCN network, a Performer encoder and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time dependence characteristics of the time series data of multiple battery cells in the battery pack to generate time series features, the Performer encoder is used to generate spatial correlation features by capturing the global dependency between the battery cells in the battery pack, and the spatiotemporal fusion layer obtains the health status prediction value by fusing the time series features and the spatial correlation feature mapping.

[0088] Specifically, TCN (Temporal Convolutional Network) is a neural network structure specifically designed to process time series data. It captures local patterns and long-term dependencies in time series through convolution operations. The SE (Squeeze-and-Excitation) module is an attention mechanism that can adaptively adjust the weights of different feature channels so that the network pays more attention to features that are important for predicting the battery health status.

[0089] In the prediction of the health status of the battery pack, the SE-TCN network is used to process the time series data of each battery cell, such as voltage, current, temperature and other data that change over time. For example, as the battery is used, the change pattern of its voltage and current during charging and discharging will change over time. The SE-TCN network can capture the time-dependent characteristics of these changes and extract valuable time series features. These features reflect the state change trend of each battery cell at different time points and are crucial for predicting the health status of the battery pack. In the battery pack, the SE-TCN network can use dynamic feature extraction methods based on the position of the battery cell to extract more targeted features, highlight important features, suppress unimportant features, and quantify the dynamic changes of the battery cell under different working conditions.

[0090] The Performer encoder is an efficient attention mechanism encoder that can capture the global dependencies between elements in the input data at a relatively low computational cost. In the battery pack scenario, the battery pack consists of multiple cells, which have mutual influences and associations between them. For example, a performance change in a cell may affect the working state of other cells, thereby affecting the health state of the entire battery pack. The Performer encoder captures this global dependency between cells and generates spatial correlation features by analyzing the feature data of each cell in the battery pack, such as the voltage difference and current distribution of different cells. These spatial correlation features can reflect the collaborative working conditions and mutual influence between cells, and provide important information for a comprehensive assessment of the health state of the battery pack.

[0091] The role of the spatiotemporal fusion layer is to fuse the time series features generated by the SE-TCN network and the spatial correlation features generated by the Performer encoder. The fusion process can be performed in a variety of ways, such as simple concatenation, weighted summation, etc., with the aim of organically combining the information in the time and space dimensions to form a more comprehensive and representative feature representation. Then, the fused features are mapped to the health status prediction value of the battery pack through a mapping function (such as a fully connected layer, etc.). This prediction value is used to evaluate the remaining service life of the battery pack, the degree of performance degradation, etc. For example, through the processing of the spatiotemporal fusion layer, the model can comprehensively consider the performance change trend of the battery cell in the time dimension and the mutual influence between the battery cells in the spatial dimension, so as to more accurately predict the health status of the battery pack at a certain moment in the future, providing an important basis for the management and maintenance of the battery.

[0092] Through the collaborative work of the above three parts, the battery pack health status prediction model can make full use of the time series data of battery cells and the spatial correlation information between cells to achieve accurate prediction of the battery pack health status and provide strong support for the efficient use and maintenance of the battery.

[0093] Step S3: input the model input data into the battery pack health state prediction model to obtain a health state prediction value of the battery pack output by the battery pack health state prediction model.

[0094] It can be seen from the above technical solutions that a method and device for predicting the health status of a battery pack provided in an embodiment of the present application first constructs a three-dimensional matrix of the battery pack based on the historical data of multiple cells in the battery pack collected, and performs sliding window processing on the three-dimensional matrix of the battery pack in combination with the currently updated window size and sliding step size to generate model input data, and the window size and the sliding step size are dynamically adjusted based on the prediction performance. Then, a battery pack health status prediction model is determined, and the battery pack health status prediction model includes an SE-TCN network, a Performer encoder, and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time-dependent characteristics of the time series data of each cell in the battery pack to generate time series features, and the Performer encoder is used to generate spatial correlation features by capturing the global dependency between the cells in the battery pack, and the spatiotemporal fusion layer obtains the health status prediction value by fusing the time series features and the spatial correlation feature mapping. Finally, the model input data is input into the battery pack health status prediction model to obtain the health status prediction value of the battery pack output by the battery pack health status prediction model.

[0095] This application can fully utilize the data of multiple cells, even the data of each cell in the battery pack, and is no longer limited to the isolated analysis of a single cell. By combining the SE-TCN network, the Performer encoder and the spatiotemporal fusion layer, the time-dependent characteristics of the time series data of each cell in the battery pack and the global dependencies between the cells are captured respectively, and the time series features and spatial correlation features are fused, so that the model can capture the complex interaction between multiple cells, and then make a more comprehensive and reasonable estimate, which greatly improves the accuracy and robustness of the battery pack health status prediction. This modeling method that comprehensively considers spatiotemporal factors is more in line with the actual working conditions of the battery pack, and can provide a more reliable basis and guidance for battery management and use. At the same time, by constructing a joint model, the previous way of modeling each cell separately is changed, which effectively reduces the complexity of the model. A large amount of repetitive independent model establishment and maintenance work is avoided, thereby significantly improving computing efficiency and reducing system resource consumption.

[0096] The above-mentioned embodiment describes the process of the battery pack health status prediction method in the present application in detail. The following will explain the process of determining the battery pack health status prediction model in the present application.

[0097] In some embodiments of the present application, determining the battery pack health status prediction model may include:

[0098] Step S21, obtaining sample model input data.

[0099] Specifically, the sample model input data is generated by performing sliding window processing on a three-dimensional matrix of a sample battery pack constructed based on sample historical data of battery cells, that is, a three-dimensional matrix is ​​constructed by collecting various historical data of battery cells, and then a sliding window method is used to generate input data suitable for model training.

[0100] Step S22: input the sample model input data into the constructed battery pack health status prediction model, train the model using the determined loss function and the selected optimizer, and iteratively update the model parameters until the preset training stop conditions are met. The training stop conditions include reaching the maximum number of iterations, the loss function value is less than a preset threshold, or the verification set accuracy reaches a preset value.

[0101] Specifically, the sample model input data is input into the constructed battery pack health status prediction model for training.

[0102] Wherein, the loss function is:

[0103]

[0104] The update rule of the optimizer to update the model parameters is:

[0105]

[0106] in, For the The true health status value of the sample battery pack, is the predicted value of the health status of the model, N is the total number of samples, are model parameters, and are the first-order and second-order moment estimates, respectively, is the learning rate, is the weight decay coefficient, is a preset constant.

[0107] In some embodiments of the present application, the structure and function of the battery pack health status prediction model of the present application are specifically introduced, such as Figure 2 As shown, the battery pack health status prediction model includes SE-TCN network A, Performer encoder B and spatiotemporal fusion layer C. Figure 3 and Figure 4 , the SE-TCN network A, Performer encoder B and spatiotemporal fusion layer C are introduced in detail in turn.

[0108] SE-TCN Network A:

[0109] Optionally, the SE-TCN network A includes multiple parallel and parameter-sharing SE-TCN structures, and the SE-TCN structure is composed of a temporal convolution module and a SE module.

[0110] The time convolution module extracts the time-dependent characteristics of the time series data of the corresponding battery cell through dilated convolution;

[0111] The SE module performs global average pooling on the channels through a compression operation, extracts the correlation between channels through an excitation operation, recalibrates the channel weights based on the correlation, and generates time series features in combination with the time-dependent characteristics.

[0112] Specifically, a parallel SE-TCN structure is designed for the time series data of each battery cell, and parameters are shared. The purpose of this is to capture the time-dependent characteristics of each battery cell. In this way, the model can process the data of multiple batteries at the same time, which can greatly reduce the number of parameters of the model and avoid overfitting. Parameter sharing is used to reduce the number of parameters of the model and improve training efficiency. At the same time, it can effectively capture the feature changes and dependencies of each battery cell in the time dimension. Among them, the input shape of the SE-TCN structure is , W is the window size, B is the number of cells, and F is the feature dimension.

[0113] This application uses dilated convolution to extract temporal features. Dilated convolution can expand the receptive field of the convolution kernel without adding too many parameters, thereby better capturing long-distance dependencies in time series. For example, when the dilation rate is large, the convolution kernel can skip some input values, thereby covering a wider time range.

[0114] The dilated convolution formula is:

[0115]

[0116] The SE module first performs global average pooling on the channels through compression operations. This step compresses the feature map of each channel into a scalar to obtain the global information of the channel. For example, for a battery cell data containing multiple features (such as voltage, current, temperature, etc.), each feature can be regarded as a channel. Through global average pooling, the information of each feature in the entire time series is compressed into one value, thereby obtaining a global summary of the feature.

[0117] Then, the correlation between channels is extracted through the excitation operation. This is achieved through the fully connected layer, which can learn the weight relationship between different channels, that is, the correlation between channels. For example, there is a certain intrinsic correlation between the voltage channel and the current channel. The excitation operation can capture this correlation and assign a weight to each channel to indicate the importance of the channel.

[0118] The calculation formula for global average pooling of channels through compression operation is:

[0119]

[0120] The calculation formula for extracting the correlation between channels through the excitation operation is:

[0121]

[0122]

[0123] Finally, the channel weights are recalibrated based on the learned correlation between channels, and the time-dependent characteristics extracted by the time convolution module are combined to generate time series features. Recalibrating the channel weights is to multiply the previously learned weights with the original channel features, so that the model can pay more attention to important channel features and suppress unimportant channel features. After combining the time-dependent characteristics, the generated time series features not only contain the change information of the battery cell in the time dimension, but also take into account the importance differences between different feature channels, providing more valuable input features for subsequent tasks such as battery health status prediction.

[0124] The calibration calculation formula for recalibrating channel weights is:

[0125]

[0126] in, is the dilated convolution output sequence, is the input sequence, is the convolution kernel size, is the convolution kernel at position The value at is the expansion rate, is the output time point, is the output value of channel c after global average pooling, L is the sequence length, is the activation value of channel c at position i, z is the weight vector obtained after the excitation operation, s is the vector composed of the values ​​obtained by the global average pooling operation of each channel, C is the total number of channels, is the ReLU activation function, is the Sigmoid activation function, and is the weight matrix of the two fully connected layers in the SE module, is the eigenvalue of the channel after recalibrating the channel weights, is the weight value corresponding to channel c in the weight vector z, is the eigenvalue of the original channel c.

[0127] Performer Encoder B:

[0128] Optionally, the Performer encoder B adopts a linear multi-head attention mechanism, which includes random feature mapping, linearized self-attention calculation, and multi-head self-attention output.

[0129] The random feature mapping is used to map the input vector to the low-dimensional feature space to form a low-dimensional feature space vector;

[0130] The linearized self-attention calculation is used to calculate the attention weights between input elements, and weighted sum the value matrix according to the attention weights to obtain an output result based on the attention mechanism;

[0131] The multi-head self-attention output is used to splice the multi-head output results, and obtain the comprehensive output result through the linear transformation realized by the weight matrix.

[0132] Specifically, the role of random feature mapping is to map the input vector to a low-dimensional feature space to form a low-dimensional feature space vector. The advantage of doing so is that it can reduce the dimension of the data, reduce the amount of calculation and the complexity of the model. For example, for the high-dimensional feature data of a large number of cells in a battery pack, random feature mapping can be used to convert it into a low-dimensional representation while retaining important information, making subsequent processing more efficient.

[0133] The calculation formula of the random feature map is:

[0134]

[0135] The linearized self-attention calculation is used to calculate the attention weights between input elements and perform weighted summation on the value matrix according to these attention weights to obtain the output result based on the attention mechanism. The traditional self-attention calculation complexity is , while Performer uses random feature mapping to approximate self-attention, so that the calculation can be linearized and the complexity is reduced to This is very important for processing long sequence data (such as large-scale cell data in battery packs), because it can greatly improve the computational efficiency of the model, allowing the model to quickly process large amounts of data without causing excessive training and prediction time due to excessive computation.

[0136] Specifically, through this linearized self-attention calculation, the model can automatically learn the importance of input elements, that is, the attention weight. For example, in a battery pack, different cells may have different degrees of mutual influence, and the state of some cells may have a greater impact on other cells. The linearized self-attention calculation can capture these relationships and perform a weighted summation of the cell features according to the weights, highlighting important information and ignoring unimportant information.

[0137] The linearized self-attention calculation formula is:

[0138]

[0139]

[0140] The multi-head self-attention mechanism can integrate the expressions of the relationship between multiple cells. The multi-head self-attention output first splices the multi-head output results, which means that the model calculates the attention weights and output results from multiple different angles. Each head has its own parameters and calculation methods, which allows the model to capture richer information and different feature patterns. Then, through the linear transformation realized by the weight matrix, the comprehensive output result is obtained. This comprehensive output result integrates the information of multiple heads and can provide a more comprehensive and accurate representation.

[0141] The calculation formula for multi-head self-attention output is:

[0142]

[0143]

[0144] in, is a low-dimensional feature space vector, x is the input feature of each battery cell, is a random vector sampled from a specific distribution, M is the dimension of the random feature, is the linearized self-attention calculation result, is the query matrix, is the key matrix, is the value matrix, The coefficient used to normalize the attention score, 1 is a full 1 vector, Output result of multi-head self-attention, for The output result of the head, Wo is the weight matrix.

[0145] The concatenated output is transformed by a linear transformation (the weight matrix is ), the dimension is , d head It is a custom parameter that defines the number of heads of the multi-head attention mechanism, indicating the number of types of expressions of cell relationships.

[0146] Furthermore, the Performer encoder adopts residual connection and performs layer normalization operation on each layer;

[0147] The calculation formula for each layer in the Performer encoder is:

[0148]

[0149] in, is the layer calculation result, is the output result of multi-head self-attention, H is the hidden state vector of each cell at the last time step. LayerNorm is the layer normalization calculation, which is used to normalize the output of each layer to make the flow of data in the network more stable.

[0150] In addition, the Performer encoder may further include an embedding layer;

[0151] The embedding layer generates a position embedding vector for representing the unique identity information of the battery cell according to the battery cell position of each battery cell.

[0152] Specifically, when processing data related to battery cells, the location information of the battery cells has an important impact on the performance of the model. The embedding layer can generate a position embedding vector specifically used to represent the unique identity information of each battery cell based on the location of the battery cell. Through the position embedding vector generated by the embedding layer in this application, the Performer encoder can better capture and utilize the location features of the battery cells, thereby improving the prediction and analysis of the battery cell status.

[0153] Space-time fusion layer C:

[0154] The spatiotemporal fusion layer obtains the health status prediction value by fusing the time series features and the spatial correlation feature mapping, that is, integrating the features of different dimensions (time and space) to realize the prediction of health status.

[0155] Flatten the output of the Performer encoder into a vector:

[0156]

[0157] Mapped to the health status prediction value through the fully connected layer:

[0158]

[0159] in, is the activation function (such as ReLU).

[0160] The following is an example of this application:

[0161] 1. Collect the historical charge / discharge timestamps, maximum current, cut-off voltage, average temperature, equivalent cycle number, test duration of multiple cells in a battery pack; and the corresponding SOH test results of the entire pack.

[0162] 2. Load the above data and divide it into training set and test set according to the proportion; the training set is used to train the model, and the test set is used to conduct real-time evaluation of the trained model on a large amount of data to verify its generalization ability.

[0163] 3. Construct a sliding window dataset by combining the data of each battery cell according to the technical solution introduction and perform model training.

[0164] 4. Save the trained model parameters and read the test set to verify the generalization ability of the model.

[0165] 5. Deploy the verified model and use it to predict single-point whole-package SOH or continuous prediction through the input of actual data.

[0166] Specific test results:

[0167] The 2000+ data sets of 12 cells are divided into training set and test set in a ratio of 8:2. Figure 5 And as shown in Table 1.

[0168] Figure 5 is a schematic diagram of the test results of the example, where Figure 5The middle figure (a) shows the comparison between the predicted SOH value of the battery pack of the training set and the true value, figure (b) shows the comparison between the predicted SOH value of the battery pack of the test set and the true value, figure (c) shows the distribution of the SOH prediction error of the battery pack of the training set, and figure (d) shows the distribution of the SOH prediction error of the battery pack of the test set.

[0169]

[0170] Table 1

[0171] The battery pack health status prediction device provided in an embodiment of the present application is described below. The battery pack health status prediction device described below and the battery pack health status prediction method described above can be referenced to each other.

[0172] See also Figure 6 , Figure 6 This is a structural block diagram of a battery pack health status prediction device disclosed in an embodiment of the present application.

[0173] like Figure 6 As shown, the battery pack health status prediction device may include:

[0174] The data processing unit 110 is used to construct a three-dimensional matrix of the battery pack based on the historical data of the multiple cells in the battery pack collected, and perform sliding window processing on the three-dimensional matrix of the battery pack in combination with the currently updated window size and sliding step size to generate model input data, wherein the window size and the sliding step size are dynamically adjusted based on the prediction performance;

[0175] A model determination unit 120 is used to determine a battery pack health state prediction model, wherein the battery pack health state prediction model includes an SE-TCN network, a Performer encoder, and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time dependency characteristics of the time series data of multiple cells in the battery pack to generate time series features, the Performer encoder is used to capture the global dependency between the cells in the battery pack to generate spatial correlation features, and the spatiotemporal fusion layer obtains a health state prediction value by fusing the time series features and the spatial correlation feature mapping;

[0176] The model prediction unit 130 is used to input the model input data into the battery pack health state prediction model to obtain the health state prediction value of the battery pack output by the battery pack health state prediction model.

[0177] It can be seen from the above technical solutions that a method and device for predicting the health status of a battery pack provided in an embodiment of the present application first constructs a three-dimensional matrix of the battery pack based on the historical data of multiple cells in the battery pack collected, and performs sliding window processing on the three-dimensional matrix of the battery pack in combination with the currently updated window size and sliding step size to generate model input data, and the window size and the sliding step size are dynamically adjusted based on the prediction performance. Then, a battery pack health status prediction model is determined, and the battery pack health status prediction model includes an SE-TCN network, a Performer encoder, and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time-dependent characteristics of the time series data of multiple cells in the battery pack to generate time series features, and the Performer encoder is used to generate spatial correlation features by capturing the global dependency between the cells in the battery pack, and the spatiotemporal fusion layer obtains the health status prediction value by fusing the time series features and the spatial correlation feature mapping. Finally, the model input data is input into the battery pack health status prediction model to obtain the health status prediction value of the battery pack output by the battery pack health status prediction model.

[0178] This application can make full use of the data of each battery cell and is no longer limited to the isolated analysis of a single battery cell. By combining the SE-TCN network, the Performer encoder and the spatiotemporal fusion layer, the time-dependent characteristics of the time series data of multiple batteries in the battery pack and the global dependencies between the batteries are captured respectively, and the time series features and spatial correlation features are fused, so that the model can capture the complex interaction between multiple batteries, and then make a more comprehensive and reasonable estimate, which greatly improves the accuracy and robustness of the battery pack health status prediction. This modeling method that comprehensively considers spatiotemporal factors is more in line with the actual working conditions of the battery pack, and can provide a more reliable basis and guidance for battery management and use. At the same time, by constructing a joint model, the previous method of modeling each battery cell separately is changed, which effectively reduces the complexity of the model. A large amount of repetitive independent model establishment and maintenance work is avoided, thereby significantly improving computing efficiency and reducing system resource consumption.

[0179] Optionally, the three-dimensional matrix shape of the battery pack is:

[0180]

[0181] The model input data shape is:

[0182]

[0183]

[0184] Among them, S is the total sample size or the number of time steps, B is the number of cells, F is the number of features, N is the number of samples, the window size is W, and the sliding step size is Δ.

[0185] Optionally, the SE-TCN network includes multiple parallel and parameter-sharing SE-TCN structures, and the SE-TCN structure is composed of a time convolution module and a SE module;

[0186] The time convolution module extracts the time-dependent characteristics of the time series data of the corresponding battery cell through dilated convolution;

[0187] The SE module performs global average pooling on the channels through a compression operation, extracts the correlation between channels through an excitation operation, recalibrates the channel weights based on the correlation, and generates time series features in combination with the time-dependent characteristics.

[0188] Optionally, the dilated convolution formula is:

[0189]

[0190] The calculation formula for global average pooling of channels through compression operation is:

[0191]

[0192] The calculation formula for extracting the correlation between channels through the excitation operation is:

[0193]

[0194]

[0195] The calibration calculation formula for recalibrating channel weights is:

[0196]

[0197] in, is the dilated convolution output sequence, is the input sequence, is the convolution kernel size, is the convolution kernel at position The value at is the expansion rate, is the output time point, is the output value of channel c after global average pooling, L is the sequence length, is the activation value of channel c at position i, z is the weight vector obtained after the excitation operation, s is the vector composed of the values ​​obtained by the global average pooling operation of each channel, C is the total number of channels, is the ReLU activation function, is the Sigmoid activation function, and is the weight matrix of the two fully connected layers in the SE module, is the eigenvalue of the channel after recalibrating the channel weights, is the weight value corresponding to channel c in the weight vector z, is the eigenvalue of the original channel c.

[0198] Optionally, the Performer encoder adopts a linear multi-head attention mechanism, wherein the linear multi-head attention mechanism includes random feature mapping, linearized self-attention calculation, and multi-head self-attention output;

[0199] The random feature mapping is used to map the input vector to the low-dimensional feature space to form a low-dimensional feature space vector;

[0200] The linearized self-attention calculation is used to calculate the attention weights between input elements, and weighted sum the value matrix according to the attention weights to obtain an output result based on the attention mechanism;

[0201] The multi-head self-attention output is used to splice the multi-head output results, and obtain the comprehensive output result through the linear transformation realized by the weight matrix.

[0202] Optionally, the calculation formula of the random feature mapping is:

[0203]

[0204] The linearized self-attention calculation formula is:

[0205]

[0206]

[0207] The calculation formula of the multi-head self-attention output is:

[0208]

[0209]

[0210] in, is a low-dimensional feature space vector, x is the input feature of each battery cell, is a random vector sampled from a specific distribution, M is the dimension of the random feature, is the linearized self-attention calculation result, is the query matrix, is the key matrix, is the value matrix, The coefficient used to normalize the attention score, 1 is a full 1 vector, Output result of multi-head self-attention, for The output result of the head, Wo is the weight matrix.

[0211] Optionally, the Performer encoder adopts residual connection and performs layer normalization operation on each layer;

[0212] The calculation formula for each layer in the Performer encoder is:

[0213]

[0214] in, is the layer calculation result, LayerNorm is the layer normalization calculation, is the output result of multi-head self-attention, and H is the hidden state vector of each battery cell at the last time step.

[0215] Optionally, the Performer encoder further includes an embedding layer;

[0216] The embedding layer generates a position embedding vector for representing the unique identity information of the battery cell according to the battery cell position of each battery cell.

[0217] Optionally, the model determination unit performs a process of determining a battery pack health status prediction model, including:

[0218] Acquire sample model input data, where the sample model input data is generated by performing sliding window processing on a three-dimensional matrix of a sample battery pack constructed based on sample historical data of battery cells;

[0219] Input the sample model input data into the constructed battery pack health status prediction model, train the model using the determined loss function and the selected optimizer, and iteratively update the model parameters until the preset training stop condition is met, wherein the training stop condition includes reaching the maximum number of iterations, the loss function value is less than the preset threshold, or the accuracy of the validation set reaches the preset value;

[0220] The loss function is:

[0221]

[0222] The update rule of the optimizer to update the model parameters is:

[0223]

[0224] in, For the The true health status value of the sample battery pack, is the predicted value of the health status of the model, N is the total number of samples, are model parameters, and are the first-order and second-order moment estimates, respectively, is the learning rate, is the weight decay coefficient, is a preset constant.

[0225] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0226] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0227] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the health status of a battery pack, characterized in that: include: A three-dimensional matrix of the battery pack is constructed based on the collected historical data of multiple cells in the battery pack, and a sliding window processing is performed on the three-dimensional matrix of the battery pack in combination with the currently updated window size and sliding step size to generate model input data, and the window size and the sliding step size are dynamically adjusted based on the prediction performance; Determine a battery pack health status prediction model, the battery pack health status prediction model includes an SE-TCN network, a Performer encoder and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time dependency characteristics of the time series data of multiple cells in the battery pack to generate time series features, the Performer encoder is used to capture the global dependency between the cells in the battery pack to generate spatial correlation features, and the spatiotemporal fusion layer obtains a health status prediction value by fusing the time series features and the spatial correlation feature mapping; The model input data is input into the battery pack health state prediction model to obtain a health state prediction value of the battery pack output by the battery pack health state prediction model.

2. The method according to claim 1, characterized in that The three-dimensional matrix shape of the battery pack is: The model input data shape is: Among them, S is the total sample size or the number of time steps, B is the number of cells, F is the number of features, N is the number of samples, the window size is W, and the sliding step size is Δ.

3. The method according to claim 1, characterized in that The SE-TCN network includes multiple parallel and parameter-sharing SE-TCN structures, and the SE-TCN structure is composed of a temporal convolution module and a SE module; The time convolution module extracts the time-dependent characteristics of the time series data of the corresponding battery cell through dilated convolution; The SE module performs global average pooling on the channels through a compression operation, extracts the correlation between channels through an excitation operation, recalibrates the channel weights based on the correlation, and generates time series features in combination with the time-dependent characteristics.

4. The method according to claim 3, characterized in that The dilated convolution formula is: The calculation formula for global average pooling of channels through compression operation is: The calculation formula for extracting the correlation between channels through the excitation operation is: The calibration calculation formula for recalibrating channel weights is: in, is the dilated convolution output sequence, is the input sequence, is the convolution kernel size, is the convolution kernel at position The value at is the expansion rate, is the output time point, is the output value of channel c after global average pooling, L is the sequence length, is the activation value of channel c at position i, z is the weight vector obtained after the excitation operation, s is the vector composed of the values ​​obtained by the global average pooling operation of each channel, C is the total number of channels, is the ReLU activation function, is the Sigmoid activation function, and is the weight matrix of the two fully connected layers in the SE module, is the eigenvalue of the channel after recalibrating the channel weights, is the weight value corresponding to channel c in the weight vector z, is the eigenvalue of the original channel c.

5. The method according to claim 1, characterized in that The Performer encoder adopts a linear multi-head attention mechanism, which includes random feature mapping, linearized self-attention calculation, and multi-head self-attention output; The random feature mapping is used to map the input vector to the low-dimensional feature space to form a low-dimensional feature space vector; The linearized self-attention calculation is used to calculate the attention weights between input elements, and weighted sum the value matrix according to the attention weights to obtain an output result based on the attention mechanism; The multi-head self-attention output is used to splice the multi-head output results, and obtain the comprehensive output result through the linear transformation realized by the weight matrix.

6. The method according to claim 5, characterized in that The calculation formula of the random feature map is: The linearized self-attention calculation formula is: The calculation formula of the multi-head self-attention output is: in, is a low-dimensional feature space vector, x is the input feature of each battery cell, is a random vector sampled from a specific distribution, M is the dimension of the random feature, is the linearized self-attention calculation result, is the query matrix, is the key matrix, is the value matrix, The coefficient used to normalize the attention score, 1 is a full 1 vector, Output result of multi-head self-attention, for The output result of the head, Wo is the weight matrix.

7. The method according to claim 1, characterized in that The Performer encoder uses residual connection and performs layer normalization operation on each layer; The calculation formula for each layer in the Performer encoder is: in, is the layer calculation result, LayerNorm is the layer normalization calculation, is the output result of multi-head self-attention, and H is the hidden state vector of each battery cell at the last time step.

8. The method according to claim 1, characterized in that The Performer encoder also includes an embedding layer; The embedding layer generates a position embedding vector for representing the unique identity information of the battery cell according to the battery cell position of each battery cell.

9. The method according to claim 1, characterized in that: The process of determining the battery pack health status prediction model includes: Acquire sample model input data, where the sample model input data is generated by performing sliding window processing on a three-dimensional matrix of a sample battery pack constructed based on sample historical data of battery cells; Input the sample model input data into the constructed battery pack health status prediction model, train the model using the determined loss function and the selected optimizer, and iteratively update the model parameters until the preset training stop condition is met, wherein the training stop condition includes reaching the maximum number of iterations, the loss function value is less than the preset threshold, or the accuracy of the validation set reaches the preset value; The loss function is: The update rule of the optimizer to update the model parameters is: in, For the The true health status value of the sample battery pack, is the predicted value of the health status of the model, N is the total number of samples, are model parameters, and are the first-order and second-order moment estimates, respectively, is the learning rate, is the weight decay coefficient, is a preset constant.

10. A battery pack health status prediction device, characterized in that: include: A data processing unit, configured to construct a three-dimensional matrix of a battery pack based on the collected historical data of a plurality of cells in the battery pack, and to perform sliding window processing on the three-dimensional matrix of the battery pack in combination with a currently updated window size and a sliding step size to generate model input data, wherein the window size and the sliding step size are dynamically adjusted based on the predicted performance; A model determination unit, used to determine a battery pack health state prediction model, wherein the battery pack health state prediction model includes an SE-TCN network, a Performer encoder, and a spatiotemporal fusion layer, wherein the SE-TCN network is used to capture the time dependency characteristics of the time series data of multiple cells in the battery pack to generate time series features, the Performer encoder is used to capture the global dependency between the cells in the battery pack to generate spatial correlation features, and the spatiotemporal fusion layer obtains a health state prediction value by fusing the time series features and the spatial correlation feature mapping; The model prediction unit is used to input the model input data into the battery pack health state prediction model to obtain the health state prediction value of the battery pack output by the battery pack health state prediction model.

Images