A battery deformation state monitoring method based on machine learning

Abstract

The present application relates to the technical field of battery, especially to a battery deformation state monitoring method based on machine learning, comprising collecting original capacitance signals in the process of battery charging and discharging through an implantable capacitance array sensor; after filtering and denoising through a signal preprocessing module, using a multi-scale time-frequency feature extraction network and parallel hole convolution kernel to extract multi-scale features; through an adaptive feature decoupling module, separating deformation features from interference features based on a channel attention mechanism and a variational autoencoder architecture; based on deformation feature components, constructing a deformation variable prediction model to obtain the current deformation variable, and combining a deformation-health mapping model to evaluate the battery health state, and outputting a warning signal when the battery health state is lower than a preset threshold. The present application can effectively capture the complex time sequence characteristics of the battery deformation signal and accurately evaluate the health state, providing technical support for battery safety management.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and in particular to a method for monitoring battery deformation state based on machine learning. Background Technology

[0002] Implantable capacitive array sensors, by being embedded inside the battery or closely attached to the battery casing, can directly acquire data on minute deformations caused by side reactions such as lithium deposition and electrolyte decomposition during the charge-discharge cycle, providing an important source of internal characteristics for assessing battery health and predicting lifespan.

[0003] The complex chemical environment inside the battery and the dynamic coupling effect of electrode materials during electrochemical reactions inevitably cause the acquired raw capacitance signal to be mixed with multiple sources of interference, such as electric field changes, chemical concentration fluctuations, and temperature gradients. How to accurately separate the characteristic components directly related to deformation from these highly coupled mixed signals has become a key bottleneck restricting the accurate perception of the battery's internal state.

[0004] Common machine learning-based battery monitoring methods typically employ deep learning architectures such as convolutional neural networks, recurrent neural networks, or attention mechanisms to extract temporal features or fuse time-frequency domain features from the original sensor signals. These extracted features are then used to assess the battery's health status. However, when processing highly coupled, low signal-to-noise ratio signals captured by implanted capacitor array sensors, the feature extraction network struggles to effectively distinguish deformation-related features from interference components. This results in insufficient regression accuracy of the monitoring model for the actual deformation within the battery, making it impossible to reliably assess the battery's health status and lifespan. Summary of the Invention

[0005] The purpose of this invention is to provide a battery deformation state monitoring method based on machine learning to solve the technical problems mentioned in the background.

[0006] This invention provides a battery deformation state monitoring method based on machine learning, comprising: The raw capacitance signal data collected by the implanted capacitance array sensor during the battery charge and discharge cycle is obtained, and the raw capacitance signal data is input into the signal preprocessing module for filtering and noise reduction to obtain the preprocessed capacitance signal. The preprocessed capacitance signal is input into a multi-scale time-frequency feature extraction network for feature extraction to obtain a multi-scale time-frequency feature representation. The time-frequency feature representations at each scale are input into the adaptive feature decoupling module for feature decoupling processing to obtain deformation-related feature components and interference feature components. A battery internal deformation prediction model is constructed based on the aforementioned deformation characteristic components, and the current deformation of the battery is obtained through regression analysis. Based on the current deformation and the pre-established deformation-health mapping model, the battery health status is assessed to obtain the battery health status assessment result. The health status assessment result is compared with a preset battery life termination threshold. When the health status assessment result is lower than the preset threshold, a battery life warning signal is output.

[0007] In some embodiments, the signal preprocessing module includes: a pre-filtering unit, a baseline drift correction unit, and a signal enhancement unit connected in sequence; The pre-filtering unit uses a bandpass filter to perform preliminary filtering on the original capacitor signal data to remove high-frequency noise and power frequency interference; The baseline drift correction unit uses an adaptive Kalman filter algorithm to correct the baseline drift of the filtered signal, so as to eliminate the baseline offset caused by sensor zero drift and changes in ambient temperature. The signal enhancement unit employs a wavelet transform domain threshold denoising method to further process the corrected signal, thereby suppressing random noise while preserving the effective components of the deformed signal.

[0008] In some embodiments, the multi-scale time-frequency feature extraction network includes: a first parallel convolutional branch, a second parallel convolutional branch, and a third parallel convolutional branch; The first parallel convolutional branch is equipped with a dilated convolutional kernel with a first dilation rate, which is used to extract low-frequency slowly varying feature components in the preprocessed capacitance signal. The second parallel convolution branch is equipped with a dilated convolution kernel with a second dilation rate, which is used to extract the mid-frequency feature components in the preprocessed capacitance signal. The third parallel convolutional branch is equipped with a dilated convolutional kernel with a third dilation rate, which is used to extract high-frequency transient feature components from the preprocessed capacitance signal. The first void ratio is less than the second void ratio, and the second void ratio is less than the third void ratio.

[0009] In some embodiments, the adaptive feature decoupling module includes: a channel attention subnetwork, a feature separation subnetwork, and a feature reconstruction subnetwork; The channel attention subnetwork is constructed based on an improved compressed excitation network structure, which is used to learn the contribution weights of each feature channel to deformation information and generate channel weight vectors. The feature separation subnetwork is built on a variational autoencoder architecture, which maps multi-scale time-frequency feature representations to a latent feature space, and achieves preliminary separation of deformation-related features and interference features in the latent space. The feature reconstruction subnetwork is constructed based on fully connected layers and deconvolutional layers, and reconstructs the separated feature vectors into deformation feature components and interference feature components respectively. The output of the adaptive feature decoupling module is a concatenated feature vector of deformation feature components and interference feature components.

[0010] In some embodiments, the battery internal deformation prediction model is composed of a cascaded deformation feature extraction subnetwork and a deformation regression subnetwork; The deformation feature extraction subnetwork is constructed using a multilayer perceptron structure, including an input layer, at least two hidden layers, and an output layer; each hidden layer is equipped with a modified linear unit activation function and a batch normalization layer. The deformation regression subnetwork is constructed using a fully connected regression layer. It receives the high-level deformation feature vector output by the deformation feature extraction subnetwork and obtains the current deformation prediction value of the battery through linear mapping.

[0011] In some embodiments, the process of establishing the deformation-health mapping model includes: Deformation data and corresponding capacity decay data of multiple battery samples during long-term charge-discharge cycles are collected; the deformation data and capacity decay data are aligned to obtain a deformation-capacity paired dataset. The deformation-capacity pairing data are fitted using a multinomial regression algorithm or a support vector regression algorithm to establish a mapping relationship between deformation and battery capacity retention rate; Based on the pre-set battery life termination threshold, the corresponding deformation threshold is determined as the basis for judging battery life warning.

[0012] In some embodiments, prior to the step of acquiring raw capacitance signal data collected by the implanted capacitor array sensor during battery charge-discharge cycles, the method further includes: The implantable capacitor array sensor is embedded inside the battery or close to the inner side of the battery casing. The implantable capacitor array sensor consists of multiple capacitor sensing units arranged in a preset array pattern. Each capacitor sensing unit is used to sense the deformation information of the battery at different locations.

[0013] In some embodiments, the capacitance signals collected by multiple capacitance sensing units are aggregated by a multiplexing circuit, converted into digital signals by an analog-to-digital converter, and transmitted to a data processing terminal. The data processing terminal uses a graphics processing unit to process the digital signals in parallel.

[0014] In some embodiments, the outputs of the first parallel convolutional branch, the second parallel convolutional branch, and the third parallel convolutional branch are integrated through a feature fusion layer. The feature fusion layer adopts a channel splicing strategy to associate the feature vectors extracted by each branch in the channel dimension, and realizes cross-channel interaction and comprehensive expression of features through pointwise convolution operations.

[0015] In some embodiments, during the operation of the adaptive feature decoupling module: The channel attention subnetwork compresses each feature channel into a channel-level scalar statistic through global average pooling, and then learns the dependencies between channels through fully connected layers and non-linear activation functions. The encoder part of the feature separation subnetwork gradually reduces the dimensionality of the high-dimensional input features to the mean and variance parameters of the latent variables through multiple fully connected layers, while the decoder part samples and reconstructs the original features from the latent variables. In the latent space, deformation-related features and interference features occupy different regions, and the two types of features are initially separated through feature clustering analysis.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. By fully utilizing the capacitance signal variation characteristics at different time scales through a multi-scale time-frequency feature extraction network, the complex temporal characteristics of battery deformation signals can be effectively captured; 2. An adaptive feature decoupling module, based on an improved channel attention mechanism and variational autoencoder architecture, effectively separates deformation-related features from interference components, solving the problem of feature extraction for highly coupled, low signal-to-noise ratio signals; 3. Establish a quantitative relationship between deformation and battery health status through a deformation-health mapping model to achieve reliable assessment of battery health status and lifespan; 4. Significantly improves the regression accuracy of implanted capacitive array sensors in battery internal state monitoring, providing an effective technical means for battery safety management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the implantable capacitor array sensor and battery deformation monitoring method of the present invention; Figure 2 This is a schematic diagram of the adaptive feature decoupling module of the present invention. Detailed Implementation

[0019] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Before implementing the implanted capacitor array sensor and battery deformation monitoring method, a sensor deployment step is required first, embedding the implanted capacitor array sensor inside the battery or close to the inner side of the battery casing. The implanted capacitor array sensor consists of multiple capacitive sensing units arranged in a preset array pattern. Each capacitive sensing unit is used to sense the deformation information of the battery at different locations. The array arrangement can be either row-column or circular to achieve comprehensive perception of the overall deformation distribution of the battery.

[0021] Example First, the raw capacitance signal data collected by the implanted capacitor array sensor during the battery charge and discharge cycle is acquired, and the raw capacitance signal data is input into the signal preprocessing module for filtering and noise reduction to obtain the preprocessed capacitance signal.

[0022] Specifically, the raw capacitance signal data includes deformation information caused by side reactions such as lithium deposition and electrolyte decomposition within the battery, as well as interference components from changes in electric field, fluctuations in chemical concentration, and temperature gradients. The capacitance signals collected by multiple capacitance sensing units are aggregated by a multiplexing circuit, converted into digital signals by an analog-to-digital converter, and transmitted to the data processing terminal. The data processing terminal uses a graphics processing unit to process the digital signals in parallel to meet the computational performance requirements of real-time monitoring. The data processing terminal also includes a data storage module to store historically collected capacitance signal data and corresponding battery status tags to support online model updates and continuous optimization.

[0023] The signal preprocessing module comprises a pre-filtering unit, a baseline drift correction unit, and a signal enhancement unit connected in sequence. The pre-filtering unit uses a bandpass filter to initially filter the raw capacitance signal data to remove high-frequency noise and power frequency interference. The baseline drift correction unit uses an adaptive Kalman filter algorithm to correct the baseline drift of the filtered signal, eliminating baseline shift caused by sensor zero-point drift and ambient temperature changes. The signal enhancement unit uses a wavelet transform domain threshold denoising method to further process the corrected signal, suppressing random noise to the greatest extent possible while preserving the effective components of the deformation signal.

[0024] After processing by the signal preprocessing module, the noise and interference components in the original capacitance signal data are effectively suppressed, and the preprocessed capacitance signal exhibits a higher signal-to-noise ratio in both the time and frequency domains.

[0025] Next, the preprocessed capacitance signal is input into a multi-scale time-frequency feature extraction network for feature extraction to obtain a multi-scale time-frequency feature representation.

[0026] Specifically, the multi-scale time-frequency feature extraction network performs convolution operations on the preprocessed capacitance signal using dilated convolution kernels with different dilation rates in parallel to capture signal variation patterns at different time scales. The multi-scale time-frequency feature extraction network includes a first parallel convolution branch, a second parallel convolution branch, and a third parallel convolution branch. The first parallel convolution branch uses a dilated convolution kernel with a first dilation rate to extract low-frequency slowly varying feature components from the preprocessed capacitance signal. The second parallel convolution branch uses a dilated convolution kernel with a second dilation rate to extract mid-frequency feature components from the preprocessed capacitance signal. The third parallel convolution branch uses a dilated convolution kernel with a third dilation rate to extract high-frequency transient feature components from the preprocessed capacitance signal. The first dilation rate is less than the second dilation rate, and the second dilation rate is less than the third dilation rate. The output features of the parallel convolution branches are fused through a feature fusion layer to obtain the fused multi-scale time-frequency feature representation.

[0027] In the actual operation of the multi-scale time-frequency feature extraction network, the first parallel convolutional branch adopts a small hole ratio configuration. Its hole convolutional kernel can extract the slow changing trend components in the capacitance signal within a large receptive field. These components mainly correspond to the macroscopic volume changes and long-term deformation accumulation effects inside the battery.

[0028] The second parallel convolutional branch employs a moderate void ratio configuration, and its voided convolutional kernel captures the periodic fluctuation characteristics of the capacitance signal within a moderate receptive field. These components are closely related to the regular deformation response during battery charging and discharging.

[0029] The third parallel convolutional branch employs a larger hole ratio configuration, and its holed convolutional kernel extracts abrupt components and high-frequency transient responses in the capacitance signal within a compact receptive field. These components correspond to rapid side reactions and sudden deformation events occurring inside the battery.

[0030] The outputs of the three parallel convolutional branches are integrated through a feature fusion layer. The feature fusion layer uses a channel concatenation strategy to concatenate the feature vectors extracted by each branch in the channel dimension, and realizes cross-channel interaction and comprehensive expression of features through pointwise convolution operations to generate a fused multi-scale time-frequency feature representation.

[0031] Next, the time-frequency feature representations at each scale are input into the adaptive feature decoupling module for feature decoupling processing to obtain deformation-related feature components and interference feature components.

[0032] Specifically, the adaptive feature decoupling module learns the contribution weights of each feature channel to deformation information based on an attention mechanism, and uses a feature reconstruction strategy to separate highly coupled mixed features into mutually independent deformation features and interference features. The adaptive feature decoupling module includes a channel attention subnetwork, a feature separation subnetwork, and a feature reconstruction subnetwork.

[0033] The system comprises several sub-networks: a channel attention sub-network based on an improved compressed excitation network structure, used to learn the contribution weights of each feature channel to deformation information and generate channel weight vectors; a feature separation sub-network based on a variational autoencoder architecture, mapping multi-scale time-frequency feature representations to a latent feature space, achieving preliminary separation of deformation-related features from interference features in the latent space; a feature reconstruction sub-network based on fully connected layers and deconvolutional layers, reconstructing the separated feature vectors into deformation feature components and interference feature components; and an adaptive feature decoupling module outputting a concatenated feature vector of deformation and interference feature components.

[0034] During the operation of the adaptive feature decoupling module, the channel attention subnetwork first receives the fused feature representation from the multi-scale time-frequency feature extraction network. This subnetwork compresses each feature channel into a channel-level scalar statistic through global average pooling. Subsequently, it learns the dependencies between channels through fully connected layers and nonlinear activation functions, generating channel weight vectors that reflect the contribution of each channel to deformation information. The channel weight vectors are then multiplied channel-by-channel with the input feature representation to enhance deformation-related feature channels and suppress interfering feature channels.

[0035] The feature separation subnetwork, building upon the feature representation enhanced by channel attention, further maps the feature space to a low-dimensional latent feature space. The encoder portion of this subnetwork progressively reduces the dimensionality of the high-dimensional input features to the mean and variance parameters of the latent variables through multiple fully connected layers, while the decoder portion samples and reconstructs the original features from the latent variables. In the latent space, deformation-related features and interference features occupy different regions. By performing feature clustering analysis in this space, preliminary separation of the two types of features can be achieved.

[0036] The feature reconstruction sub-network receives the separated feature vectors and maps them back to the original feature space through independent reconstruction paths, outputting a spliced ​​feature vector of deformation feature components and interference feature components.

[0037] Then, a prediction model of the battery's internal deformation is constructed based on the deformation characteristic components, and the current deformation of the battery is obtained through regression analysis.

[0038] Specifically, the battery internal deformation prediction model consists of a cascaded deformation feature extraction subnetwork and a deformation regression subnetwork. The deformation feature extraction subnetwork is constructed using a multilayer perceptron structure, including an input layer, at least two hidden layers, and an output layer. Each hidden layer is equipped with a modified linear unit activation function and a batch normalization layer to accelerate network convergence and improve feature representation capability. The deformation regression subnetwork is constructed using fully connected regression layers, receiving the high-level deformation feature vector output by the deformation feature extraction subnetwork, and obtaining the predicted value of the battery's current deformation through linear mapping. The battery internal deformation prediction model is trained using a mean squared error loss function, and the training samples include battery charge-discharge experimental data with known deformation.

[0039] In the construction of the battery internal deformation prediction model, the deformation feature extraction sub-network receives deformation feature components from the adaptive feature decoupling module as input. The number of neurons in the input layer matches the dimension of the deformation feature components, completing the initial feature reading. The hidden layer adopts a multilayer perceptron architecture, with each hidden layer containing multiple neurons, and all neurons are fully connected. A modified linear unit activation function introduces nonlinear transformation capabilities, enabling the network to learn the complex mapping relationships between deformation features. The batch normalization layer standardizes each batch of input data, normalizing the data distribution to a standard normal distribution with a mean of zero and a variance of one, effectively alleviating the internal covariate bias problem, accelerating network training convergence, and enhancing the model's generalization ability.

[0040] The output layer maps the high-level deformation feature vectors extracted by the hidden layer to a feature space of a preset dimension, outputting the results of the deformation feature extraction subnetwork. The deformation regression subnetwork receives the high-level deformation feature vectors as input, performs linear combination and nonlinear transformation through a fully connected regression layer, and finally outputs the predicted value of the battery's current deformation. This predicted value reflects the actual deformation of the battery relative to its initial state under the current charge / discharge state.

[0041] Then, based on the current deformation variables and the pre-established deformation-health mapping model, the battery health status is assessed to obtain the battery health status assessment results.

[0042] Specifically, the deformation-health mapping model is trained using historical charge-discharge cycle data and characterizes the quantitative relationship between battery deformation and battery health state. The process of establishing the deformation-health mapping model includes the following steps: Collect deformation data and corresponding battery capacity decay data of multiple battery samples during long-term charge-discharge cycles; The deformation data and capacity decay data are aligned to obtain a deformation-capacity paired dataset; Multinomial regression or support vector regression algorithms are used to fit the deformation-capacity paired data to establish a mapping relationship between deformation and battery capacity retention rate. Based on the pre-set battery life termination threshold, the corresponding deformation threshold is determined as the basis for judging battery life warning.

[0043] In the process of establishing the deformation-health mapping model, firstly, a large number of battery samples were collected for deformation measurement data and corresponding battery capacity measurement data under different charge and discharge cycle stages. These data cover the complete life cycle of the battery from brand new to the end of its life, and cover a variety of operating conditions and environmental conditions.

[0044] By aligning the timestamps, deformation and capacity data from the same moment are paired and combined to form a deformation-capacity paired dataset. When using the polynomial regression algorithm, a polynomial function of appropriate order is selected for fitting based on the data distribution characteristics. The polynomial coefficients are solved using the least squares method to minimize the sum of squared residuals between the fitted curve and the actual data points.

[0045] When using the support vector regression algorithm, a suitable kernel function is selected to map the input data to a high-dimensional feature space, and the optimal hyperplane is found in this space to realize the regression prediction from deformation to capacity preservation.

[0046] The above method establishes a quantitative mapping relationship between deformation and battery capacity retention rate. Based on the capacity retention rate corresponding to the pre-set battery life end threshold, the corresponding deformation threshold is calculated in reverse. This deformation threshold serves as the critical value for determining whether the battery has reached the end of its life.

[0047] During the battery health status assessment process, the current deformation is used as input and substituted into a pre-established deformation-health mapping model for calculation to obtain the current battery health status assessment result. This assessment result is expressed in the form of capacity retention rate, reflecting the degree to which the battery retains its initial capacity in its current state.

[0048] Finally, the health status assessment result is compared with the preset battery life termination threshold. When the health status assessment result is lower than the preset threshold, a battery life warning signal is output.

[0049] Specifically, the obtained battery health status assessment results are compared point by point with the preset battery life termination threshold. When the health status assessment result is higher than or equal to the preset threshold, it indicates that the battery is still in normal service and the monitoring system continues to perform the routine monitoring process. If the health status assessment result is lower than the preset threshold, the battery life warning mechanism is triggered and the system outputs a battery life warning signal.

[0050] The warning signal includes key parameters such as the battery's current health status, current deformation value, and cumulative charge-discharge cycle count, and prompts the user to pay attention or replace the battery in a timely manner through the human-machine interface.

[0051] In practical applications, the machine learning-based battery deformation state monitoring method provided in this embodiment can be applied to the real-time monitoring of power battery systems in new energy vehicles. During vehicle operation, an embedded capacitor array sensor continuously collects deformation signal data from inside the battery. The data processing terminal executes the processing flow of this embodiment in real time to continuously assess the battery's health status. When the battery health status assessment result is detected to be lower than a preset threshold, a warning signal is immediately sent to the vehicle's dashboard, prompting the driver to perform timely battery maintenance or replacement to ensure driving safety.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0053] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for monitoring battery deformation state based on machine learning, characterized in that, include: The raw capacitance signal data collected by the implanted capacitance array sensor during the battery charge and discharge cycle is obtained, and the raw capacitance signal data is input into the signal preprocessing module for filtering and noise reduction to obtain the preprocessed capacitance signal. The preprocessed capacitance signal is input into a multi-scale time-frequency feature extraction network for feature extraction to obtain a multi-scale time-frequency feature representation. The time-frequency feature representations at each scale are input into the adaptive feature decoupling module for feature decoupling processing to obtain deformation-related feature components and interference feature components. A battery internal deformation prediction model is constructed based on the aforementioned deformation characteristic components, and the current deformation of the battery is obtained through regression analysis. Based on the current deformation and the pre-established deformation-health mapping model, the battery health status is assessed to obtain the battery health status assessment result. The health status assessment result is compared with a preset battery life termination threshold. When the health status assessment result is lower than the preset threshold, a battery life warning signal is output.

2. The method according to claim 1, characterized in that, The signal preprocessing module includes: a pre-filtering unit, a baseline drift correction unit, and a signal enhancement unit connected in sequence; The pre-filtering unit uses a bandpass filter to perform preliminary filtering on the original capacitor signal data to remove high-frequency noise and power frequency interference; The baseline drift correction unit uses an adaptive Kalman filter algorithm to correct the baseline drift of the filtered signal, so as to eliminate the baseline offset caused by sensor zero drift and changes in ambient temperature. The signal enhancement unit employs a wavelet transform domain threshold denoising method to further process the corrected signal, thereby suppressing random noise while preserving the effective components of the deformed signal.

3. The method according to claim 1, characterized in that, The multi-scale time-frequency feature extraction network includes: a first parallel convolutional branch, a second parallel convolutional branch, and a third parallel convolutional branch; The first parallel convolutional branch is equipped with a dilated convolutional kernel with a first dilation rate, which is used to extract low-frequency slowly varying feature components in the preprocessed capacitance signal. The second parallel convolution branch is equipped with a dilated convolution kernel with a second dilation rate, which is used to extract the mid-frequency feature components in the preprocessed capacitance signal. The third parallel convolutional branch is equipped with a dilated convolutional kernel with a third dilation rate, which is used to extract high-frequency transient feature components from the preprocessed capacitance signal. The first void ratio is less than the second void ratio, and the second void ratio is less than the third void ratio.

4. The method according to claim 1, characterized in that, The adaptive feature decoupling module includes: a channel attention subnetwork, a feature separation subnetwork, and a feature reconstruction subnetwork; The channel attention subnetwork is constructed based on an improved compressed excitation network structure, which is used to learn the contribution weights of each feature channel to deformation information and generate channel weight vectors. The feature separation subnetwork is built on a variational autoencoder architecture, which maps multi-scale time-frequency feature representations to a latent feature space, and achieves preliminary separation of deformation-related features and interference features in the latent space. The feature reconstruction subnetwork is constructed based on fully connected layers and deconvolutional layers, and reconstructs the separated feature vectors into deformation feature components and interference feature components respectively. The output of the adaptive feature decoupling module is a concatenated feature vector of deformation feature components and interference feature components.

5. The method according to claim 1, characterized in that, The battery internal deformation prediction model is composed of a cascaded deformation feature extraction subnetwork and a deformation regression subnetwork. The deformation feature extraction subnetwork is constructed using a multilayer perceptron structure, including an input layer, at least two hidden layers, and an output layer; each hidden layer is equipped with a modified linear unit activation function and a batch normalization layer; The deformation regression subnetwork is constructed using a fully connected regression layer. It receives the high-level deformation feature vector output by the deformation feature extraction subnetwork and obtains the current deformation prediction value of the battery through linear mapping.

6. The method according to claim 1, characterized in that, The process of establishing the deformation-health mapping model includes: Deformation data and corresponding capacity decay data of multiple battery samples during long-term charge-discharge cycles are collected; the deformation data and capacity decay data are aligned to obtain a deformation-capacity paired dataset. The deformation-capacity pairing data are fitted using a multinomial regression algorithm or a support vector regression algorithm to establish a mapping relationship between deformation and battery capacity retention rate; Based on the pre-set battery life termination threshold, the corresponding deformation threshold is determined as the basis for judging battery life warning.

7. The method according to claim 1, characterized in that, Before the step of acquiring the raw capacitance signal data collected by the implanted capacitor array sensor during battery charge-discharge cycles, the following steps are also included: The implantable capacitor array sensor is embedded inside the battery or close to the inner side of the battery casing. The implantable capacitor array sensor consists of multiple capacitor sensing units arranged in a preset array pattern. Each capacitor sensing unit is used to sense the deformation information of the battery at different locations.

8. The method according to claim 7, characterized in that, The capacitance signals collected by multiple capacitance sensing units are aggregated by a multiplexing circuit, converted into digital signals by an analog-to-digital converter, and transmitted to a data processing terminal. The data processing terminal uses a graphics processing unit to process the digital signals in parallel.

9. The method according to claim 3, characterized in that, The outputs of the first, second, and third parallel convolutional branches are integrated through a feature fusion layer. The feature fusion layer adopts a channel splicing strategy to associate the feature vectors extracted from each branch in the channel dimension, and realizes cross-channel interaction and comprehensive expression of features through pointwise convolution operations.

10. The method according to claim 4, characterized in that, During the operation of the adaptive feature decoupling module: The channel attention subnetwork compresses each feature channel into a channel-level scalar statistic through global average pooling, and then learns the dependencies between channels through fully connected layers and non-linear activation functions. The encoder part of the feature separation subnetwork gradually reduces the dimensionality of the high-dimensional input features to the mean and variance parameters of the latent variables through multiple fully connected layers, while the decoder part samples and reconstructs the original features from the latent variables. In the latent space, deformation-related features and interference features occupy different regions, and the two types of features are initially separated through feature clustering analysis.

Images