A lithium battery electrochemical impedance spectroscopy prediction method and system

Abstract

This invention discloses a method and system for predicting the electrochemical impedance spectroscopy (EIS) of lithium-ion batteries, belonging to the field of battery management technology. The method includes: collecting constant-current charging curves of lithium-ion batteries under different aging conditions; extracting capacity, temperature, and expansion displacement sequences as multi-source input features; targeting the EIS sequence under a fixed state of charge; constructing a dual-attention convolutional long short-term memory (LSTM) network prediction model, which integrates channel attention and spatial attention mechanisms, combined with hierarchical convolution and LSM units; training the model and applying it to unknown charging curves, outputting the predicted EIS results. This invention integrates force, heat, and electrical multi-source signals, enabling high-precision EIS estimation under different states of charge and charging rates, and provides interpretability through SHAP analysis, significantly improving the accuracy and adaptability of battery state monitoring.

Description

Technical Field

[0001] This invention relates to the fields of battery management systems and deep learning technology, specifically to a method and system for predicting the electrochemical impedance spectroscopy of lithium batteries based on machine learning of multi-source feature signals. Background Technology

[0002] Electrochemical impedance spectroscopy (EIS) is widely used to estimate key battery performance indicators such as state of charge (SOC), state of health (SOH), and remaining useful life (RUL). However, traditional EIS testing methods suffer from three major bottlenecks: high equipment cost, lengthy testing cycles, and the need for a long period of rest before testing to allow the battery to reach a steady state, a time that is difficult to reduce.

[0003] To address these issues, researchers have developed online soft measurement techniques for EIS based on the dynamic characteristics of conventional charging and discharging. For example, Duan et al. constructed a deep neural network model that mapped from the constant current charging curve to the full-frequency EIS; Guo et al. estimated the impedance spectrum using the local charging voltage curve; Sun et al. estimated the EIS using a combination of partial constant current charging voltage-capacity curves and Seq2Seq; Ko et al. used the current signal in the constant voltage charging or relaxation voltage as input; and Zhu et al. proposed a few-shot learning method.

[0004] Regarding existing patents, CN118501713A discloses a semi-supervised learning-based method for predicting the electrochemical impedance spectroscopy of lithium batteries, but it only applies to impedance spectrum estimation at a fully charged state (100% SOC), ignoring performance changes at different SOCs. CN116520174A discloses a method for predicting battery impedance spectra based on current, voltage, and temperature, but it requires complete charging data, which is difficult to obtain in actual dynamic operating environments. Furthermore, existing technologies primarily rely on charge and discharge data (voltage, current), failing to fully utilize other important physical signals generated during battery operation, such as temperature and mechanical signals.

[0005] Therefore, there is an urgent need for a lithium battery electrochemical impedance spectroscopy prediction method that integrates multiple physical signals, is applicable to different SOCs and rates, and does not require complete charging data. Summary of the Invention

[0006] This invention aims to overcome the shortcomings of existing technologies that rely on complete charging data, are only applicable to the fully charged state, and do not fully utilize multi-source signals of force, heat, and electricity. It provides a machine learning method based on multi-source characteristic signals such as pressure, capacity, and temperature for predicting the electrochemical impedance spectroscopy of lithium batteries, thereby improving the accuracy and reliability of battery state estimation and enhancing its adaptability to practical application scenarios.

[0007] To achieve the above objectives, the technical solution provided by the present invention is as follows: A first aspect of the present invention provides a method for predicting the electrochemical impedance spectroscopy of lithium batteries, comprising the following steps: Step 1: Collect constant current charging curves of lithium batteries under different aging conditions, divide them, use the divided charging curves as the input of the training set, and set the target as the electrochemical impedance spectroscopy sequence of the battery under fixed charge state. Step 2: Construct a prediction model using a dual-attention convolutional long short-term memory network; Step 3: Train the dual-attention convolutional long short-term memory network prediction model; Step 4: Apply the trained deep learning model to predict the electrochemical impedance spectroscopy of the unknown charging curve.

[0008] Preferably, in step 1, three types of physical quantities collected during the constant current charging process are extracted from the measurement data as model inputs: capacity sequence, temperature sequence, and expansion displacement sequence; wherein the expansion displacement sequence is offset with the initial capacity as a reference benchmark to remove the dependence on the initial capacity value.

[0009] Preferably, the dual-attention convolutional long short-term memory network model includes a channel attention module, a spatial attention module, a hierarchical convolutional structure, long short-term memory units, and a fully connected network; the channel attention module is used to learn the dependencies between sensor channels, the spatial attention module is used to locate key time steps and local context regions, the hierarchical convolutional structure is used for hierarchical extraction of local patterns, the long short-term memory units are used for recursive memorization of global temporal evolution, and the fully connected network is used to map features to the target spectral space of electrochemical impedance spectroscopy.

[0010] Preferably, in the dual-attention convolutional long short-term memory network model, data passes sequentially through three cascaded blocks, each block consisting of a one-dimensional convolutional layer, a convolutional block attention module, and a max-pooling layer; the number of convolutional channels in the three cascaded blocks are 32, 64, and 128, respectively; at the end of the feature extraction backbone, an adaptive average pooling layer maps the variable-length time series input to a fixed 256-dimensional feature vector; this 256-dimensional feature vector is fed into the long short-term memory unit for time series modeling, and then mapped to the 72-dimensional electrochemical impedance spectroscopy target spectral space through two fully connected network layers.

[0011] Preferably, in step 3, the model training uses the Adam optimizer with an initial learning rate of 0.01, employs a mini-batch gradient descent strategy with a batch size of 64, and a maximum number of iterations of 5000; an early stopping mechanism is introduced, which terminates training early when the validation set loss does not improve within 500 consecutive iterations; and the loss function combines the mean square error of the real and imaginary parts of the electrochemical impedance spectroscopy.

[0012] Preferably, the method is capable of estimating the electrochemical impedance spectroscopy of lithium batteries under different states of charge and different charging rates; and also includes combining SHAP analysis to reveal the degree of influence of each input feature on the electrochemical impedance spectroscopy prediction results.

[0013] Preferably, in step 1, the discretized data of all battery cycles are preprocessed using Z-score standardization, and both the validation set and the test set are standardized using the parameters of the training set; in step 4, the trained model is saved locally, and when used later, the battery data is obtained and the model is directly loaded for prediction to obtain the frequency and its corresponding impedance value.

[0014] Preferably, the spatial attention module constructs a dual-channel input by performing a dot product between the vector and the feature map in space, and combining the results of max pooling and average pooling; the one-dimensional convolutional layer uses a 3×1 kernel with symmetrical padding.

[0015] Preferably, the dual-attention convolutional long short-term memory network prediction model constructed in step 2 takes a standardized multi-source time-series signal as input and outputs an electrochemical impedance spectral sequence under a fixed charge state.

[0016] A second aspect of the present invention provides a lithium battery electrochemical impedance spectroscopy prediction system, comprising: Data acquisition and preprocessing module: used to collect constant current charging curves of lithium batteries under different aging conditions, extract capacity sequence, temperature sequence and expansion displacement sequence collected during constant current charging, and perform standardized preprocessing; Model building module: used to build a prediction model of dual-attention convolutional long short-term memory network. The model includes a channel attention module, a spatial attention module, a hierarchical convolutional structure, long short-term memory units, and a fully connected network. The hierarchical convolutional structure contains three cascaded blocks. Each block consists of a one-dimensional convolutional layer, a convolutional block attention module, and a max pooling layer. The number of convolutional channels is 32, 64, and 128, respectively. An adaptive average pooling layer is set at the end of the feature extraction backbone to map the variable-length time series input to a fixed 256-dimensional feature vector. This 256-dimensional feature vector is then mapped to the 72-dimensional electrochemical impedance spectroscopy target spectral space through the long short-term memory unit and two fully connected network layers. Model training module: used to train the prediction model using the training set data; Prediction module: used to apply the trained model to the unknown charging curve and output the electrochemical impedance spectroscopy prediction results.

[0017] Compared with the prior art, the present invention has the following significant advantages: 1) Multi-source feature fusion: Simultaneously utilizing three types of physical signals—capacity (electrical), temperature (thermal), and expansion displacement (force)—makes up for the deficiency that single electrical information cannot fully characterize the changes in the internal state of the battery.

[0018] 2) No complete charging data required: EIS prediction can be achieved using specific data segments (constant current charging segments), which is more adaptable to actual dynamic operating environments.

[0019] 3) Applicable to the entire SOC range: It can effectively estimate the EIS of lithium batteries under different states of charge and different charging rates, overcoming the limitation of existing technologies that are only applicable to the fully charged state.

[0020] 4) Dual attention mechanism: The dual collaboration of channel attention and spatial attention enables the model to simultaneously capture the local fine features and global long-term dependencies of the data, significantly improving feature extraction capability and prediction accuracy.

[0021] 5) Interpretability: SHAP analysis is introduced to clarify the contribution of each input feature to the prediction results, providing a basis for model optimization and battery mechanism analysis. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application 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 recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic flowchart of a lithium battery electrochemical impedance spectroscopy prediction method provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the Dual Attention Convolutional Long Short-Term Memory Network (DAC-LSTM) model of the present invention.

[0025] Figure 3 The diagram shows the EIS estimation error distribution of the DAC-LSTM model of this invention under different discharge SOCs, where (a) is the RMSE error distribution and (b) is the RMSE of the real and imaginary parts of the impedance.

[0026] Figure 4The diagram shows the RMSE distribution of the EIS estimation of the DAC-LSTM model of this invention under different charging rates. Detailed Implementation

[0027] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Please see Figure 1 In a first aspect, the present invention provides a method for predicting the electrochemical impedance spectroscopy of lithium batteries, comprising the following steps: Step S101: Collect the constant current charging curves of lithium batteries under different aging conditions, divide them, use the divided charging curves as the input of the training set, and set the target as the electrochemical impedance spectroscopy sequence of the battery under a fixed charge state. In the preprocessing stage, three types of physical quantities collected during the constant current charging process are extracted from the measurement data as model inputs: capacity sequence. Temperature sequence T, expansion displacement sequence F. Specifically, let Q... }、 , There are three corresponding data sequences, among which, Indicates the first Capacity value for each data point Taken directly from the initial measurement. With initial capacity Offset processing is performed using a reference benchmark to remove the bias. This data construction method is closer to the input segment of the actual battery charging process because the model can make estimates without relying on the exact initial capacity value. Indicates the first Temperature values ​​of each data point. Indicates the first The expansion displacement value of each data point.

[0029] To improve the model's generalization ability and estimation accuracy, this chapter performs standardized preprocessing on the discretized data for all battery cycles. Specifically, the input variables are processed using the Z-score standardization method, and the validation and test sets are standardized using the same parameters as the training set.

[0030] Step S102: Construct a prediction model for a dual-attention convolutional long short-term memory network; Single attention mechanisms have limitations in mining deep spatiotemporal correlations of multi-source heterogeneous signals. To address this, this paper proposes a dual-attention convolutional long short-term memory network (DAC-LSTM). This model combines improved channel attention, spatial attention, and LSTM temporal modeling units to process temporal data from multiple sensors. While traditional channel attention modules can enhance the representation of one-dimensional signals by adaptively recalibrating to filter key channels and suppress noise, they do not adequately address the dynamic evolution of features over time. To solve this problem, DAC-LSTM introduces a dual channel-spatiotemporal collaboration mechanism: in the channel domain, the improved channel attention module learns the dependencies between sensor channels, focusing on high-information feature subspaces; in the spatiotemporal domain, the temporal spatial attention module locates key time steps and local context regions to capture transient abrupt changes in long-range dependencies. The core of the model consists of three main parts: a dual-attention collaboration module responsible for weighted filtering of multi-dimensional features, a hierarchical convolutional structure responsible for progressive extraction of local patterns, and an LSTM unit responsible for recursive memorization of global temporal evolution. This design enables the model to simultaneously capture both fine local features and long-term global dependencies of the data, achieving a dual focus on information channels and key time steps. Experimental results show that, compared with the baseline model, DAC-LSTM exhibits better feature extraction capabilities and prediction accuracy when processing high-dimensional, nonlinear, and strongly coupled multi-source data.

[0031] like Figure 2 As shown, the data is sequentially processed through three cascaded blocks, each consisting of a three-layer sequence: one-dimensional convolution, a Convolutional Block Attention Module (CBAM), and max pooling. CBAM enhances important features by performing a dot product between the vector and the feature map in space, and constructs a dual-channel input by combining the results of max pooling and average pooling. Different sized convolutional kernels (kernels of 5, 3, and 3 respectively) are used to adjust the local context awareness range. The number of convolutional channels in these blocks are 32, 64, and 128, respectively. Each convolutional layer uses a 3×1 kernel with symmetric padding (padding = 1). At the end of this feature extraction backbone, an adaptive average pooling layer maps the variable-length time-series input to a fixed 256-dimensional feature vector.

[0032] The 256-dimensional vector after feature extraction is fed into an LSTM for time-series modeling, and finally mapped to a 72-dimensional EIS target spectral space through a two-layer fully connected network. The regression head consists of two cascaded linear layers: the first layer maps the 256-dimensional input to 64 neurons, and the second layer further maps it to a 72-dimensional output. A ReLU activation function is used between both layers to introduce a nonlinear transformation, ultimately outputting reconstructed electrochemical impedance spectroscopy data.

[0033] Step S103: Train the dual-attention convolutional long short-term memory network prediction model; In terms of model construction, 20% of the dataset was reserved as a test set, while the remaining subset was used for training. The Adam optimizer was used for model training, with an initial learning rate of 0.01, employing a mini-batch gradient descent strategy (batch size = 64), and a maximum of 5000 iterations. An early stopping mechanism was also introduced to suppress overfitting; training was terminated early if the validation set loss did not show effective improvement within 500 consecutive iterations. The loss function for the entire impedance spectrum combines the real and imaginary parts: in and These represent the real and imaginary parts of the EIS measurement, respectively. and Let these represent the real and imaginary part estimates of the EIS, respectively. This refers to the batch size. Model selection is based on the minimum MSE on the validation set.

[0034] Here, RMSE is set to ,in and Let represent the real and imaginary parts of the EIS measurement, respectively, and n represent the total sample size. This invention is tested on a dataset, where... Figure 3 The objective is to estimate EIS under different SOCs. To examine the generalization ability of the embodiments of the present invention under different SOCs, DAC-LSTM models for each SOC level were trained using constant current charging data and corresponding EIS labels for multiple SOCs. Figure 3 As can be seen in (a), the median RMSE of each SOC point is basically stable in the range of 0.9 to 2.0 mΩ, which indicates that the model has a relatively robust estimation ability under different SOC conditions. Figure 3 (b) further demonstrates the variation of the estimation errors of the real part (Z') and imaginary part (Z'') of the impedance with SOC, thereby observing their sensitivity to SOC. Real part It fluctuates between 0.9 and 2.0 mΩ, while the imaginary part... The value is between 0.5 and 1.0 mΩ, indicating that the model performs well in predicting different SOCs.

[0035] Figure 4The objective is to estimate the EIS (Electronic Information Performance) of data at different charging rates as input. Unlike the training and testing strategies in previous sections, this experiment uses 80% of the data from all charging rates as the training set, while the remaining 20% ​​of the data at the target charging rate is used as an independent test set to verify the model's generalization ability under different charging rate inputs. Finally, RMSE (Recovery Rate Estimate) is used as the error metric, and the EIS estimation errors at each charging rate are shown below. Figure 4 As shown in the figure. Experiments show that the model maintains good performance at different charging rates, with RMSE below 4 mΩ at most test points.

[0036] Step S104: Apply the trained deep learning model to predict the electrochemical impedance spectroscopy of the unknown charging curve.

[0037] Save the trained model locally. When using it later, obtain battery data and directly load the model to make predictions, and obtain the frequency and its corresponding impedance value.

[0038] This invention provides a method for estimating the electrochemical impedance spectroscopy (EIS) of lithium-ion batteries under different states of charge and charging rates. Furthermore, it incorporates SHAP (Shapley Additive Explanations) analysis to reveal the degree of influence of each input feature on the EIS prediction results, providing a direct theoretical basis for subsequent model optimization and data improvement. This invention also provides a lithium battery electrochemical impedance spectroscopy prediction system, comprising: Data acquisition and preprocessing module: used to collect constant current charging curves of lithium batteries under different aging conditions, extract capacity sequence, temperature sequence and expansion displacement sequence collected during constant current charging, and perform standardized preprocessing; Model building module: used to build a prediction model of dual-attention convolutional long short-term memory network. The model includes a channel attention module, a spatial attention module, a hierarchical convolutional structure, long short-term memory units, and a fully connected network. The hierarchical convolutional structure contains three cascaded blocks. Each block consists of a one-dimensional convolutional layer, a convolutional block attention module, and a max pooling layer. The number of convolutional channels is 32, 64, and 128, respectively. An adaptive average pooling layer is set at the end of the feature extraction backbone to map the variable-length time series input to a fixed 256-dimensional feature vector. This 256-dimensional feature vector is then mapped to the 72-dimensional electrochemical impedance spectroscopy target spectral space through the long short-term memory unit and two fully connected network layers. Model training module: used to train the prediction model using the training set data; Prediction module: used to apply the trained model to the unknown charging curve and output the electrochemical impedance spectroscopy prediction results.

[0039] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily 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 invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lithium battery electrochemical impedance spectroscopy prediction method, characterized in that, Includes the following steps: Collect constant current charging curves of lithium batteries under different aging conditions, divide them, use the divided charging curves as input to the training set, and set the target as the electrochemical impedance spectroscopy sequence of the battery under fixed charge state. Construct a prediction model for a dual-attention convolutional long short-term memory network; Train the dual-attention convolutional long short-term memory network prediction model; The trained deep learning model is used to predict the electrochemical impedance spectroscopy of unknown charging curves.

2. The lithium battery electrochemical impedance spectroscopy prediction method of claim 1, wherein, In step 1, three types of physical quantities collected during the constant current charging process are extracted from the measurement data as model inputs: capacity sequence, temperature sequence, and expansion displacement sequence; wherein the expansion displacement sequence is offset with the initial capacity as a reference benchmark to remove the dependence on the initial capacity value. 3.The lithium battery electrochemical impedance spectroscopy prediction method of claim 1, wherein, The dual-attention convolutional long short-term memory network model includes a channel attention module, a spatial attention module, a hierarchical convolutional structure, long short-term memory units, and a fully connected network. The channel attention module is used to learn the dependencies between sensor channels, the spatial attention module is used to locate key time steps and local context regions, the hierarchical convolutional structure is used for hierarchical extraction of local patterns, the long short-term memory units are used for recursive memorization of global temporal evolution, and the fully connected network is used to map features to the target spectral space of electrochemical impedance spectroscopy.

4. The lithium battery electrochemical impedance spectroscopy prediction method of claim 3, wherein, In the dual-attention convolutional long short-term memory network model, data passes sequentially through three cascaded blocks, each consisting of a one-dimensional convolutional layer, a convolutional block attention module, and a max-pooling layer; the number of convolutional channels in the three cascaded blocks are 32, 64, and 128, respectively; at the end of the feature extraction backbone, an adaptive average pooling layer maps the variable-length time series input to a fixed 256-dimensional feature vector; this 256-dimensional feature vector is fed into the long short-term memory unit for time series modeling, and then mapped to the 72-dimensional electrochemical impedance spectroscopy target spectral space through two fully connected network layers.

5. The lithium battery electrochemical impedance spectroscopy prediction method of claim 1, wherein, In step 3, the model training uses the Adam optimizer with an initial learning rate of 0.01, employs a mini-batch gradient descent strategy with a batch size of 64, and a maximum number of iterations of 5000. An early stopping mechanism is introduced, which terminates training early when the validation set loss does not improve within 500 consecutive iterations. The loss function combines the mean square error of the real and imaginary parts of the electrochemical impedance spectroscopy. 6.The lithium battery electrochemical impedance spectroscopy prediction method of claim 1, wherein, The method can estimate the electrochemical impedance spectroscopy of lithium batteries under different states of charge and different charging rates; and also includes combining SHAP analysis to reveal the degree of influence of each input feature on the electrochemical impedance spectroscopy prediction results.

7. The method for predicting the electrochemical impedance spectroscopy of lithium batteries according to claim 1, characterized in that, In step 1, the discretized data of all battery cycles are preprocessed using Z-score standardization, and the validation set and test set are standardized using the parameters of the training set. In step 4, the trained model is saved locally. When used later, the battery data is obtained and the model is directly loaded for prediction to obtain the frequency and its corresponding impedance value.

8. The lithium battery electrochemical impedance spectroscopy prediction method of claim 3, wherein, The spatial attention module constructs a dual-channel input by performing a dot product between the vector and the feature map in space, and combining the results of max pooling and average pooling; the one-dimensional convolutional layer uses a 3×1 kernel with symmetrical padding. 9.The lithium battery electrochemical impedance spectroscopy prediction method of claim 1, wherein, The dual-attention convolutional long short-term memory network prediction model constructed in step 2 takes a standardized multi-source time-series signal as input and outputs an electrochemical impedance spectral sequence under a fixed charge state.

10. A lithium battery electrochemical impedance spectroscopy prediction system, characterized in that, include: Data acquisition and preprocessing module: used to collect constant current charging curves of lithium batteries under different aging conditions, extract capacity sequence, temperature sequence and expansion displacement sequence collected during constant current charging, and perform standardized preprocessing; Model building module: used to build a prediction model of dual-attention convolutional long short-term memory network. The model includes a channel attention module, a spatial attention module, a hierarchical convolutional structure, long short-term memory units, and a fully connected network. The hierarchical convolutional structure contains three cascaded blocks. Each block consists of a one-dimensional convolutional layer, a convolutional block attention module, and a max pooling layer. The number of convolutional channels is 32, 64, and 128, respectively. An adaptive average pooling layer is set at the end of the feature extraction backbone to map the variable-length time series input to a fixed 256-dimensional feature vector. This 256-dimensional feature vector is then mapped to the 72-dimensional electrochemical impedance spectroscopy target spectral space through the long short-term memory unit and two fully connected network layers. Model training module: used to train the prediction model using the training set data; Prediction module: used to apply the trained model to the unknown charging curve and output the electrochemical impedance spectroscopy prediction results.

Images