A lithium battery health state estimation method based on stability characteristics and AS-TCN model

Abstract

The present invention relates to an improvement of a lithium battery state of health assessment method, specifically a lithium battery state of health estimation method based on stable characteristics and an AS-TCN model, which can realize accurate and efficient monitoring of SOH, and comprises the following steps: step S1: selecting a lithium battery to be tested for a charge and discharge experiment, wherein the technical parameters to be collected in the experiment include the voltage value, temperature value, collection time point and battery capacity value of the corresponding cycle under constant current charging conditions; step S2: performing curve fitting and data preprocessing on the voltage and temperature data of each group of cyclic charge and discharge cycles at the same time node; step S3: clustering the collected data using a dynamic time warping centroid averaging algorithm, including a dynamic time warping distance algorithm to determine the optimal alignment of two sequences in the case of time misalignment, and then dividing the clustered voltage and temperature sequences into a training data set and a test data set for model prediction; step S4: building an attention separable temporal convolutional network model.

Description

Technical Field

[0001] The present invention relates to an improvement on a lithium battery health status assessment method, specifically a lithium battery health status estimation method based on stability characteristics and an AS-TCN model. Background Art

[0002] Lithium-ion batteries have been widely used in our daily lives due to their fast charging speed, low self-discharge, and long service life. Their applications have expanded beyond portable electronic devices (such as mobile phones, cameras, and laptops) to larger applications (such as new energy vehicles, drones, and satellites). They are also essential for larger industrial equipment and energy storage. During cycling, lithium-ion batteries experience a decrease in total capacity and an increase in internal resistance. This phenomenon indicates a decline in the battery's energy storage capacity, a characteristic commonly referred to as battery aging. The state of health (SOH) is a metric used to assess the extent of battery aging, essentially reflecting the degree of degradation and damage. Battery aging is the process by which various battery functions deteriorate. It is a complex, nonlinear process involving a wide range of factors. Generally, a battery's service life ends when its actual capacity drops to 70%-80% of its rated capacity. Continued use at this point can compromise the overall system's performance, potentially leading to catastrophic consequences.

[0003] Given the current state of research, equivalent circuit models, which simulate battery characteristics by forming circuits with electronic devices, are easier to model than electrochemical models, have stronger feasibility, and exhibit stronger dynamic response. However, the essence of equivalent circuits is approximation, and parameter deviations in some models can lead to large prediction errors. Electrochemical models can accurately estimate the battery's state of health (SOH) by primarily studying the electrochemical reaction processes within the battery. However, health estimates based on electrochemical models are difficult to model and face challenges in practical application. Methods based on mechanistic models are difficult to implement and lack universal applicability. Therefore, more research is currently focused on data-driven models, which are categorized as empirical models and rely on training on large amounts of data to achieve accurate predictions.

[0004] According to the current research status of data-driven models, the data-driven model consists of two parts: feature extraction and model estimation. Regarding the feature extraction part, the current research status shows that:

[0005] 1. Different feature extraction methods produce varying degrees of correlation for batteries under different working conditions. Currently, there is no high-precision feature extraction method that is applicable to all batteries.

[0006] 2. Current feature extraction methods only analyze the importance of features based on the correlation between characteristic factors and capacity degradation trends, while ignoring the impact of improving high-correlation features on prediction results from the essence of the model. In particular, many current feature extraction methods still have the limitation of relying on the entire charging and discharging process.

[0007] 3. There is great instability in the discharge process.

[0008] Regarding the model estimation part, we can find from the current research status that:

[0009] The models that can be improved by the health feature-based SOH estimation method are mostly concentrated in traditional algorithms or models that have been proven to perform well in the field of battery performance, but there is still much room for improvement.

[0010] Therefore, it is important to design a method that can accurately and efficiently monitor SOH to ensure that batteries are replaced before they fail. Summary of the Invention

[0011] In order to solve this problem, the present invention proposes a lithium battery state of health estimation method based on stability characteristics and AS-TCN model, which can realize accurate and efficient monitoring of SOH.

[0012] To solve the above technical problems, the technical solution adopted by the present invention is: a lithium battery health status estimation method based on stability characteristics and AS-TCN model, comprising the following steps:

[0013] Step S1: Select a lithium battery to be tested for charge and discharge experiments. The technical parameters that need to be collected in the experiment include voltage value, temperature value, collection time point and battery capacity value of the corresponding cycle under constant current charging conditions;

[0014] Step S2: performing curve fitting and data preprocessing on the voltage and temperature data of each set of cyclic charge and discharge cycles at the same time node;

[0015] Step S3: Clustering the collected data using the Dynamic Time Warping (DTW) centroid averaging algorithm, which includes a Dynamic Time Warping (DTW) distance algorithm to determine the optimal alignment of two sequences in the case of time misalignment. The clustered voltage and temperature series are then divided into training and test datasets for model prediction.

[0016] Step S4: Build an attention separable temporal convolutional network model, including a depth-wise separable convolutional improved structure and a convolutional attention mechanism model;

[0017] Step S5: Estimate the health status of the lithium battery based on the established AS-TCN model.

[0018] Preferably, the curve fitting and data preprocessing of the data in step S2 specifically include: fitting the corresponding time series of the two technical parameter sequence groups of the collected lithium battery voltage and temperature at the same collection time node, and the result is a time series of the numerical changes of the temperature and voltage at the same charging time node as the battery ages under different charge and discharge cycles.

[0019] Preferably, in step S3, clustering the collected data using the dynamic time warping centroid averaging algorithm specifically includes:

[0020] Step S31: randomly selecting a sequence from the current sequence set S as an average sequence;

[0021] Step S32: Calculate the DTW distance between the selected average sequence and each sequence in the sequence set, and match the coordinates of the average sequence with the coordinates of other sequences in the sequence set;

[0022] Step S33: According to the DTW distance algorithm, assume that there is a sequence A=(a1, a2, ..., a m ), B=(b1,b2,...,b n ), then the DTW distance between A and B is recorded as D(A i ,B j ), represents the DTW distance between the i-th time point of sequence A and the j-th time point of sequence B;

[0023] Step S34: Compare with the first calculated DTW distance. If they are the same, execute step S36; otherwise, execute step S35.

[0024] Step S35: Determine whether the sum of squares of the DTW distances decreases. If so, proceed to step S36. If not, the current average sequence is the result.

[0025] Step S36: Update each coordinate of the average sequence to the mean value of the coordinates matching it in the sequence set, and then return to step S32 to continue iteration.

[0026] Preferably, the DTW distance algorithm in step S33 determines the optimal alignment of the two sequences under the condition of time misalignment, specifically including:

[0027] (1) Definition of boundary conditions: Given w1 = (1, 1), w k =(m,n), the starting and ending points of the curved path must be the first and last points of the aligned time series;

[0028] (2) Definition of monotonicity condition: Given w k =(x k ,y k ), w k-1 =(xk-1 ,y k-1 ), where x k -x k-1 ≥0,y k -y k-1 ≥0;

[0029] (3) Definition of step size condition: Given w k =(x k ,y k ), w k-1 =(x k-1 ,y k-1 ), where x k -x k-1 ≤1,y k -y k-1 ≤1, the basic step length condition is w k -w k-1 ∈{(1,1),(1,0),(0,1)};

[0030] (4) Cumulative distance d w The calculation formula for (A, B) is:

[0031]

[0032] Where, is the set of all possible paths;

[0033] (5) The minimum calculation formula of the cumulative distance is:

[0034]

[0035] (6) The cumulative distance matrix is ​​D, D(i,j) is the set of all elements of the distance matrix D, and the calculation formula is:

[0036]

[0037]

[0038] D(i,j)=C ij +min{C i-1,j-1 ,C i-1,j ,C i,j-1},i∈[1:m],j∈[1:n] (5).

[0039] Preferably, in step S4, the depthwise separable convolution improved structure includes:

[0040] (1) The parameter calculation formula of the optimized temporal convolutional network based on depthwise separable convolution is: S×K×1+S×1×1×n;

[0041] Among them, n×S is the dimension of the input, S is the length of the sequence, and n is the number of sequences; K×1 represents the size of the convolution kernel;

[0042] (2) The input dimension of the temporal convolutional network based on depthwise separable convolution is: a×b;

[0043] Where a and b represent the number and length of the input sequence, and the value range of a and b increases, a ≥ 1, b ≥ 1;

[0044] (3) The specific position of depthwise separable convolution in the temporal convolutional network structure: Depthwise convolution of depthwise separable convolution exists at the input position of the network for automatic feature extraction, and Pointwise convolution is used to reorganize the extracted features before the fully connected layer;

[0045] (4) Select the activation function as Leaky Relu;

[0046] Preferably, in step S4, the convolutional attention mechanism model includes:

[0047] (1) The channel attention calculation formula is:

[0048]

[0049] Among them, Mc represents the channel attention feature, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, W0∈R C / r×1 , W1∈R 1×C / r , σ represents the Sigmoid activation operation, F represents the feature map, C represents the size of the third dimension of the feature map, and r is the reduction rate;

[0050] (2) The calculation formula of spatial attention features is:

[0051]

[0052] Among them, Ms represents the spatial attention feature, 7×7 represents the convolution kernel size, and s represents the size of the second dimension of the feature map;

[0053] (3) The calculation formula of the final generated feature F″ is as follows:

[0054]

[0055] (4) The specific location of the convolutional attention mechanism in the AS-TCN model is: after batch normalization and before the Leaky Relu activation function.

[0056] Preferably, in step S4, each residual module of the AS-TCN model consists of two sub-modules, which are connected by a residual structure in the middle. The sub-module structure is: hybrid dilated causal convolution, batch normalization, convolutional attention mechanism, LeakyReLu activation function and dropout; the overall structure is: input layer, Depthwise convolution operation, residual module, Ponitwise convolution operation and output layer.

[0057] Preferably, in step S5, estimating the health status of the lithium battery based on the established AS-TCN model specifically includes the following steps:

[0058] Step S51: Determine the technical parameters that need to be collected: current, voltage, time node and lithium battery capacity;

[0059] Step S52: Data preprocessing: including data denoising and missing value supplementation, fitting the corresponding time series for the two technical parameter sequence groups of lithium battery voltage and temperature collected at the same collection time node;

[0060] Step S53: Data input: cluster the pre-processed voltage and temperature data using the DBA algorithm to remove data spatial coupling, and input any segment current and voltage data;

[0061] Step S54: training the model: dividing the data into training data and test data at a ratio of 7:3, and adjusting the hyperparameters in the model by the control variable method;

[0062] Step S55: Selection of regression evaluation criteria: selecting mean square error and root mean square error as indicators for evaluating model accuracy;

[0063] Step S56: Model verification: Use 30% of the test data to verify the model and prove the quality of the model through regression evaluation indicators.

[0064] The lithium battery health status estimation method based on stability characteristics and AS-TCN model of the present invention can achieve the following beneficial effects:

[0065] (1) The automatic feature extraction mode replaces manual feature extraction and uses a depth-wise separable convolutional feature extraction layer to automatically extract features. The features obtained in this way are more stable and more universally applicable, and are not restricted to the type of battery.

[0066] (2) Only part of the data is needed to complete the prediction. The key point is that there is no need for correlation analysis to judge the quality of the features. The original data and prediction model can be adjusted directly based on the prediction results, rather than adjusting the manually extracted features, which is more reliable.

[0067] (3) The voltage and temperature values ​​of the battery at different aging degrees are characterized by the acquisition time. It is found that there is a large amount of redundant data. It can be seen from the figure that there is a stable stage in the voltage and temperature data. The data in this stage are highly similar in space and are mainly in the stable stage. The voltage and temperature series in the stable stage are clustered by the DBA algorithm, which retains the original trend of the data and reduces the computational burden of the prediction model.

[0068] (4) Build an improved AS-TCN model, which is based on the lightweight model TCN, with small computational complexity and high accuracy. It overcomes the shortcomings of the original TCN for multi-sequence prediction models and is easier to use in practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is a flow chart of the lithium battery health status estimation method of the present invention;

[0070] Figure 2 It is the time series of the temperature and voltage values ​​at the same charging time node as the battery ages;

[0071] Figure 3 This is the result of DBA clustering of some voltage and temperature data;

[0072] Figure 4 Schematic diagram of the AS-TCN model structure and the technologies used in each layer of the network;

[0073] Figure 5 Schematic diagram of offline estimation results of unknown battery capacity. DETAILED DESCRIPTION

[0074] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0075] A lithium battery state-of-health (SOH) estimation method based on stable features and an AS-TCN neural network is proposed. This method is based on the stable degradation trend of capacity during charging, which is reflected in the stable degradation trend of technical parameters, namely voltage and temperature, with battery aging. This trend is directly used as the basis for determining the remaining battery capacity, resulting in good stability. A deep separable convolution layer is then used to automatically extract features, which avoids the errors associated with manually extracting health features and is more universally applicable. The DBA algorithm is then used to cluster health features based on time nodes to compress data without changing the overall trend. Finally, an AS-TCN estimation model is established. During model construction, the TCN is used as the original model. Depthwise convolution, a technique used in deep separable convolution, is used to automatically extract multi-dimensional health features from the input. The Leaky Relu activation function is used to address the problem of excessive neuron death caused by small feature changes. The convolutional attention mechanism improves the overall prediction accuracy of the model, captures subtle features, and suppresses redundant features that cannot be manually removed. Finally, pointwise convolution is used to aggregate all features into a fully connected layer for final prediction.

[0076] In the specific implementation, the process is as follows Figure 1 As shown, the specific steps are as follows:

[0077] Step S1: Select a lithium battery to be tested for charge and discharge experiments. The technical parameters that need to be collected in the experiment include voltage value, temperature value, collection time point and battery capacity value of the corresponding cycle under constant current charging conditions;

[0078] Step S2: curve fitting and data preprocessing are performed on the voltage and temperature data of each set of charge and discharge cycles at the same time node. The data preprocessing includes de-duplication, de-noising and zero filling. The two technical parameter sequence groups of the collected lithium battery voltage and temperature are fitted with the corresponding time series at the same collection time node. The result is the time series of the temperature and voltage values ​​at the same charging time node as the battery ages under different charge and discharge cycles. Figure 2 As shown in the figure, it can be found that there is a stable stage in the voltage and temperature data, and the data in this stage are highly similar in space;

[0079] Step S3: Clustering the spatially highly similar data between voltage and temperature using the Dynamic Time Warping Barycenter Averaging (DBA) algorithm, specifically including:

[0080] Step S31: randomly selecting a sequence from the current sequence set S as an average sequence;

[0081] Step S32: Calculate the DTW distance between the selected average sequence and each sequence in the sequence set, and match the coordinates of the average sequence with the coordinates of other sequences in the sequence set.

[0082] Step S33: According to the DTW distance algorithm, assume that there is a sequence A=(a1, a2, ..., a m ), B=(b1,b2,...,b n ), then the DTW distance between A and B is recorded as D(A i ,B j ), represents the DTW distance between the i-th time point of sequence A and the j-th time point of sequence B;

[0083] Step S34: Compare with the first calculated DTW distance. If they are the same, execute step S36; otherwise, execute step S35.

[0084] Step S35: Determine whether the sum of squares of the DTW distances decreases. If so, proceed to step S36. If not, the current average sequence is the result.

[0085] Step S36: Update each coordinate of the average sequence to the mean value of the coordinates matching it in the sequence set, and then return to step S32 to continue iteration.

[0086] Step S33 of the DBA algorithm includes a dynamic time warping (DTW) distance algorithm to determine the optimal alignment of two sequences in the case of time misalignment, which is specifically as follows:

[0087] (1) Definition of boundary conditions: Given w1 = (1, 1), w k =(m,n), the starting point and the ending point of the curved path must be the first point and the last point of the aligned time series.

[0088] (2) Definition of monotonicity condition: Given w k =(x k ,y k ), w k-1 =(x k-1 ,y k-1 ), where x k -x k-1 ≥0,y k -y k-1 ≥0, which makes the points in W monotonically spaced in time.

[0089] (3) Definition of step size condition: Given w k =(x k ,y k ), w k-1 =(x k-1 ,yk-1 ), where x k -x k-1 ≤1,y k -y k-1 ≤1, the basic step length condition is w k -w k-1 ∈{(1,1),(1,0),(0,1)}, which restricts long-distance jumps in the curved path during sequence alignment.

[0090] (4) Cumulative distance d w The calculation formula for (A, B) is:

[0091]

[0092] Where w∈W=(w1,w2,...,w k ), is the set of all possible paths

[0093] (5) The minimum calculation formula of the cumulative distance is:

[0094]

[0095] (6) The cumulative distance matrix is ​​D, D(i,j) is the set of all elements of the distance matrix D, and the calculation formula is:

[0096]

[0097]

[0098] D(i,j)=C ij +min{C i-1,j-1 ,C i-1,j ,C i,j-1},i∈[1:m],j∈[1:n] (5)

[0099] The clustered voltage and temperature series are then divided into training and testing datasets for model prediction. Figure 3 This is the result of DBA clustering of some voltage and temperature data. From the figure, we can see that the DBA algorithm completely extracts the overall trend of this part of the data.

[0100] Step S4: Build the Attention Separable Temporal Convolutional Networks (AS-TCN) model, including the depth-wise separable convolutional improved structure and convolutional attention mechanism model, the AS-TCN model structure and the technology used in each layer of the network. Figure 4As shown in the figure, after the original data is input, it first goes through the initial feature extraction stage of Depthwise. From the perspective of the model structure, the input value and calculation process of the improved model includes:

[0101] (1) The input dimension of the temporal convolutional network based on depthwise separable convolution is: a×b, where a and b represent the number and length of the input sequence, and the value range of a and b increases, a≥1, b≥1. In this study, voltage and temperature are selected as the input sequence, that is, the value of a is 2.

[0102] (2) The parameter calculation formula for the optimized temporal convolutional network based on depthwise separable convolution is: S×K×1+S×1×1×n. The parameter calculation formula for ordinary convolution is S×K×1×n, where n×S is the input dimension, S is the length of the sequence, and n is the number of sequences; K×1 represents the size of the convolution kernel. From the formula, it can be seen that the greater the difference between the input and output dimensions n of the model, the fewer the parameters. Therefore, this method has the advantage of low computational complexity for voltage and temperature data with large data volumes.

[0103] (3) The specific location of the depthwise convolution in the AS-TCN structure: The depthwise convolution of the depthwise separable convolution is located at the input of the network. This has two advantages: first, it is used for automatic feature extraction, thereby obtaining stable features; second, it improves the TCN network into a new structure for multivariate sequence prediction. Pointwise convolution is used to reorganize the extracted features before the fully connected layer.

[0104] After the Depthwise feature is automatically extracted, the feature is further optimized through the convolutional attention mechanism model:

[0105] (1) Channel Attention:

[0106] In order to improve the model's ability to pay attention to important information, and focus on the factors that really affect the prediction results, rather than just from the correlation analysis. The lightweight convolutional attention module is integrated into the model. As a method to improve the representation ability of the CNN network, it simply and efficiently improves the model's ability to learn important features. This study converts it from the three-dimensional image processing field to two-dimensional to process sequence data, and embeds it into the TCN structure. The lightweight convolutional attention module consists of a channel attention mechanism and a spatial attention mechanism, which are used to increase the expressive power of features. Compared with most attention mechanisms that only focus on channels, the effect is better. The model first performs average pooling operations and maximum pooling downsampling on the intermediate feature map in the channel dimension, compresses the global spatial information into a one-dimensional vector channel descriptor, and puts it into the multi-layer perceptron for adjustment. Finally, the sigmoid activation operation is performed to generate the final channel attention feature map Mc

[0107]

[0108] Among them, Mc represents the channel attention feature, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, W0∈R C / r×1 , W1∈R 1×C / r , σ represents the Sigmoid activation operation, F represents the feature map, C represents the size of the third dimension of the feature map, and r is the reduction rate.

[0109] (2) The channel attention feature map is element-wise multiplied with the input feature map to generate the input features required by the spatial attention module. Then, in the spatial dimension, the feature map processed in the channel dimension is processed by global average pooling and global maximum pooling based on the channel, and the result is connected based on the channel. Then, the dimension is reduced by a convolution operation, and finally the spatial attention feature Ms is output through the sigmoid activation function. The calculation formula of the spatial attention feature is:

[0110]

[0111] Among them, Ms represents the spatial attention feature, 7×7 represents the convolution kernel size, and s represents the size of the second dimension of the feature map.

[0112] By multiplying Ms by the input features of the module, the final generated feature F″ can be obtained. The features after denoising by the convolutional attention mechanism can be regularized before regression prediction through pointwise convolution.

[0113]

[0114] Secondly, it is important to note that the convolutional attention mechanism is placed in the AS-TCN model after batch normalization and before the Leaky Relu activation function. The Leaky Relu activation function was chosen through a combination of theoretical and experimental methods, addressing the issue of neuron death during actual estimation by setting the negative interval to zero.

[0115] Step S5: Estimate the health status of the lithium battery based on the established AS-TCN model.

[0116] The topology of the SOH estimation based on the AS-TCN model is as follows: First, each residual module in the AS-TCN model consists of two submodules connected by a residual structure. The submodule structure includes: hybrid dilated causal convolution, batch normalization, convolutional attention mechanism, LeakyReLu activation function, and dropout. The overall structure is: input layer, depthwise convolution operation layer, five residual modules, positwise convolution regularization layer, and output layer.

[0117] The specific steps include:

[0118] Step S51: Determine the model input parameters and their dimensions: select voltage and temperature as inputs, and the corresponding capacity as input labels;

[0119] Step S52: data preprocessing before automatic feature extraction: including data denoising and missing value supplementation, and data normalization;

[0120] Step S53: Model input: cluster the pre-processed voltage and temperature data using the DBA algorithm to remove data spatial coupling, and input any segment current and voltage data;

[0121] Step S54: training the model: the data is divided into training data and test data in a ratio of 7:3, and the hyperparameters in the model are adjusted by the control variable method. The loss function is selected as the mean square error function, and the optimizer is selected as the Adam optimization algorithm;

[0122] Step S55: Selection of regression evaluation criteria: selecting mean absolute error and root mean square error as indicators for evaluating model accuracy;

[0123] Step S56: Model verification: Use 30% of the test data to verify the model and use regression evaluation indicators to prove the quality of the model. Figure 5 This is the offline estimation result of the unknown battery capacity, with a root mean square error of 0.012 and a mean absolute error of 0.007.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

[0125] The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A lithium battery health status estimation method based on stability characteristics and AS-TCN model, characterized in that: The following steps are involved: Step S1: Select a lithium battery to be tested for charge and discharge experiments. The technical parameters that need to be collected in the experiment include voltage value, temperature value, collection time point and battery capacity value of the corresponding cycle under constant current charging conditions; Step S2: performing curve fitting and data preprocessing on the voltage and temperature data of each set of cyclic charge and discharge cycles at the same time node; Step S3: Clustering the collected data using the Dynamic Time Warping (DTW) centroid averaging algorithm, which includes a Dynamic Time Warping (DTW) distance algorithm to determine the optimal alignment of two sequences in the case of time misalignment. The clustered voltage and temperature series are then divided into training and test datasets for model prediction. Step S4: Build an attention separable temporal convolutional network model, including a depth-wise separable convolutional improved structure and a convolutional attention mechanism model; Step S5: estimating the health status of the lithium battery based on the established AS-TCN model; In step S3, the dynamic time warping centroid averaging algorithm is used to cluster the collected data, specifically including: Step S31: randomly selecting a sequence from the current sequence set S as an average sequence; Step S32: Calculate the DTW distance between the selected average sequence and each sequence in the sequence set, and match the coordinates of the average sequence with the coordinates of other sequences in the sequence set; Step S33: According to the DTW distance algorithm, assume that there is a sequence A=(a1, a2, ..., a m ), B=(b1,b2,...,b n ), then the DTW distance between A and B is recorded as D(A i ,B j ), represents the DTW distance between the i-th time point of sequence A and the j-th time point of sequence B; Step S34: Compare with the first calculated DTW distance. If they are the same, execute step S36; otherwise, execute step S35. Step S35: Determine whether the sum of squares of the DTW distances decreases. If so, proceed to step S36. If not, the current average sequence is the result. Step S36: Update each coordinate of the average sequence to the mean value of the coordinates matching it in the sequence set, and then return to step S32 to continue iteration; In step S33, the DTW distance algorithm determines the optimal alignment of the two sequences under the condition of time misalignment, specifically including: (1) Definition of boundary conditions: Given w1 = (1, 1), w k =(m,n), the starting and ending points of the curved path must be the first and last points of the aligned time series; (2) Definition of monotonicity condition: Given w k =(x k ,y k ), w k-1 =(x k-1 ,y k-1 ), where x k -x k-1 ≥0,y k -y k-1 ≥0; (3) Definition of step size condition: Given w k =(x k ,y k ), w k-1 =(x k-1 ,y k-1 ), where x k -x k-1 ≤1,y k -y k-1 ≤1, the basic step length condition is w k -w k-1 ∈{(1,1),(1,0),(0,1)}; (4) Cumulative distance d w The calculation formula for (A, B) is: Where, is the set of all possible paths; (5) The minimum calculation formula of the cumulative distance is: (6) The cumulative distance matrix is ​​D, D(i,j) is the set of all elements of the distance matrix D, and the calculation formula is: D(i,j)=C ij +min{C i-1,j-1 ,C i-1,j ,C i,j-1 },i∈[1:m],j∈[1:n](5); In step S4, the depthwise separable convolution improved structure includes: (1) The parameter calculation formula of the optimized temporal convolutional network based on depthwise separable convolution is: S×K×1+S×1×1×n; Among them, n×S is the dimension of the input, S is the length of the sequence, and n is the number of sequences; K×1 represents the size of the convolution kernel; (2) The input dimension of the temporal convolutional network based on depthwise separable convolution is: a×b; Where a and b represent the number and length of the input sequence, and the value range of a and b increases, a ≥ 1, b ≥ 1; (3) The specific position of depthwise separable convolution in the temporal convolutional network structure: Depthwise convolution of depthwise separable convolution exists at the input position of the network for automatic feature extraction, and Pointwise convolution is used to reorganize the extracted features before the fully connected layer; (4) Select the activation function as Leaky Relu; In step S4, the convolutional attention mechanism model includes: (1) The channel attention calculation formula is: Among them, Mc represents the channel attention feature, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, W0∈R C / r×1 , W1∈R 1×C / r , σ represents the Sigmoid activation operation, F represents the feature map, C represents the size of the third dimension of the feature map, and r is the reduction rate; (2) The calculation formula of spatial attention features is: Among them, Ms represents the spatial attention feature, 7×7 represents the convolution kernel size, and s represents the size of the second dimension of the feature map; (3) The calculation formula of the final generated feature F″ is as follows: (4) The specific location of the convolutional attention mechanism in the AS-TCN model is: after batch normalization and before the Leaky Relu activation function.

2. The lithium battery health status estimation method based on stability characteristics and AS-TCN model according to claim 1 is characterized in that: The curve fitting and data preprocessing of the data in step S2 specifically include: fitting the corresponding time series of the two technical parameter sequence groups of the collected lithium battery voltage and temperature at the same collection time node, and the result is a time series of the numerical changes of temperature and voltage at the same charging time node as the battery ages under different charge and discharge cycles.

3. The method for estimating the health status of a lithium battery based on stability characteristics and an AS-TCN neural network according to claim 1, wherein: In step S4, each residual module of the AS-TCN model consists of two submodules connected by a residual structure. The submodule structure is: hybrid dilated causal convolution, batch normalization, convolutional attention mechanism, LeakyReLu activation function and dropout; the overall structure is: input layer, depthwise convolution operation, residual module, pointwise convolution operation and output layer.

4. The method for estimating the health status of a lithium battery based on stability characteristics and an AS-TCN neural network according to claim 1, wherein: In step S5, the health status of the lithium battery is estimated based on the established AS-TCN model, which specifically includes the following steps: Step S51: Determine the technical parameters that need to be collected: current, voltage, time node and lithium battery capacity; Step S52: Data preprocessing: including data denoising and missing value supplementation, fitting the corresponding time series for the two technical parameter sequence groups of lithium battery voltage and temperature collected at the same collection time node; Step S53: Data input: cluster the pre-processed voltage and temperature data using the DBA algorithm to remove data spatial coupling, and input any segment current and voltage data; Step S54: training the model: dividing the data into training data and test data at a ratio of 7:3, and adjusting the hyperparameters in the model by the control variable method; Step S55: Selection of regression evaluation criteria: selecting mean square error and root mean square error as indicators for evaluating model accuracy; Step S56: Model verification: Use 30% of the test data to verify the model and prove the quality of the model through regression evaluation indicators.

Images