Battery health state estimation method and device based on convolution automatic encoder
By extracting battery aging features from electrochemical impedance spectroscopy data using a convolutional autoencoder and a deep learning model, the problem of insufficient accuracy and robustness in SOH estimation in existing technologies is solved, and high-precision battery health state estimation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies struggle to effectively extract time-frequency features of battery aging from high-dimensional electrochemical impedance spectroscopy data, and neglect time-series dependence and correlations between features, resulting in insufficient accuracy and robustness in SOH estimation.
A convolutional autoencoder was used to extract overcomplete features from electrochemical impedance spectroscopy data. Combined with convolutional neural networks and recurrent neural networks, a time-series feature sequence was constructed. Local correlation and time-dependent features were fused, and the SOH of the battery was estimated through a deep learning model.
It achieves high-precision and robust battery SOH estimation, overcomes the limitations of traditional methods, and improves the continuity and accuracy of the estimation.
Smart Images
Figure CN121955796A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of battery health status assessment technology, and more specifically relates to a battery health status estimation method and apparatus based on a convolutional autoencoder. Background Technology
[0002] Lithium-ion batteries, as electrochemical energy storage devices, are widely used in electric transportation, energy storage power stations, portable electronic devices, and industrial power supplies due to their high energy density, high charge-discharge efficiency, and wide range of applications. However, during long-term charge-discharge processes, battery performance gradually degrades due to complex internal electrochemical reactions and external operating conditions, leading to a decline in safety and economic efficiency. Battery state of health (SOH) is a key parameter for measuring the degree of battery aging, typically defined as the ratio of current usable capacity to rated capacity. Therefore, achieving accurate real-time estimation of SOH is crucial for ensuring the safe operation of battery systems, extending their effective lifespan, and optimizing system operation strategies.
[0003] Currently, the mainstream methods for SOH estimation mainly include direct measurement methods, model-based methods, and data-driven methods.
[0004] The direct measurement method mainly obtains the physical quantities related to capacity directly through coulomb counting or internal resistance measurement. Although it does not require complex algorithms or assumed models, its measurement results are easily affected by test conditions and noise interference, making it difficult to achieve high-precision online estimation under actual dynamic working conditions.
[0005] Model-based methods mainly include electrochemical models and equivalent circuit models. Electrochemical models, starting from the internal reaction mechanisms of the battery, can deeply reveal the aging process. However, these models involve a large number of partial differential equations and physicochemical parameters, making parameter identification difficult and computationally complex, thus limiting their applicability to practical online estimation scenarios. Equivalent circuit models simulate the external characteristics of the battery through circuit elements such as resistors and capacitors, achieving a certain balance between computational efficiency and accuracy. However, their estimation performance heavily depends on the accurate identification of the model structure and parameters, and the mapping relationship between model parameters and state of equilibrium (SOH) usually exhibits strong nonlinear and time-varying characteristics, making it difficult to maintain stable accuracy throughout the battery's entire lifespan.
[0006] Data-driven approaches do not rely on precise battery mechanism models, but instead learn the mapping relationship between aging characteristics and state of equilibrium (SOH) from historical data through machine learning algorithms. Electrochemical impedance spectroscopy (EIS) can reflect multi-scale, multi-frequency dynamic response information within the battery and is considered an effective data source for characterizing battery aging. However, EIS data is characterized by high dimensionality and strong nonlinearity, making it challenging to directly extract robust features strongly correlated with aging from raw EIS data. Furthermore, existing data-driven methods often ignore the temporal dependencies present during battery aging and do not fully consider the local correlations and global coupling relationships between impedance characteristics at different frequencies, resulting in limited generalization ability of the constructed models and decreased estimation accuracy under complex operating conditions and differences between multiple battery batches.
[0007] Therefore, how to effectively extract time-frequency features that are strongly correlated with battery aging from high-dimensional EIS data, and on this basis, construct an estimation model that can integrate time-series dependence and the correlation between features, is a key technical challenge for achieving high-precision and high-robustness SOH estimation. Summary of the Invention
[0008] To address the shortcomings and improvement needs of existing technologies, this invention provides a method and apparatus for estimating battery state of health based on a convolutional autoencoder. The purpose is to effectively extract time-frequency features strongly correlated with battery aging from high-dimensional EIS data, and on this basis, construct an estimation model that can integrate time-series dependence and the correlation between features to achieve high-precision and highly robust battery SOH estimation.
[0009] To achieve the above objectives, according to one aspect of the present invention, a battery health state estimation method based on a convolutional autoencoder is provided, comprising: S1: Collect electrochemical impedance spectroscopy (EIS) data of the battery, process the EIS data, and use a convolutional autoencoder to automatically extract overcomplete features from the EIS data. S2: Stack overcomplete features from multiple consecutive cycles to form a short-term time-dependent temporal feature sequence; S3: Input the time-series feature sequence into the convolutional neural network model and the recurrent neural network model respectively. The convolutional neural network model extracts the local correlation features of the time-series feature sequence, and the recurrent neural network model is used to extract the time-dependent features of the time-series feature sequence. S4: Concatenate the locally relevant feature vectors and time-dependent feature vectors output by the convolutional neural network model and the recurrent neural network model respectively, and input the concatenated feature vector into the DNN model to estimate the current battery's maximum available capacity. S5: Estimate the SOH of the battery by using the ratio of the estimated current maximum available capacity of the battery to the rated capacity of the battery.
[0010] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects: This invention achieves automated feature extraction from raw EIS data. By introducing a convolutional autoencoder, it can automatically learn and extract effective overcomplete features directly from high-dimensional, nonlinear raw EIS data. This method completely avoids the tedious process and subjective limitations of manual feature design relying on expert prior knowledge in traditional technical approaches, automating feature engineering and extracting more comprehensive features that better characterize the essential information of battery aging.
[0011] This invention effectively models the temporal dynamics and state dependence of the battery aging process. It creatively constructs extracted features as time-series data as model input, enabling the entire estimation model to explicitly learn and capture the dynamic degradation trajectory and state dependence of battery performance during long-term cycling. This design overcomes the static modeling deficiency of most existing data-driven methods that treat each measurement as an independent sample, thus significantly improving the continuity and accuracy of SOH estimation.
[0012] This invention employs a joint learning mechanism that deeply integrates local feature correlation and long-term temporal dependencies. By combining a hybrid architecture of convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, it achieves efficient collaborative mining of multi-level information from EIS data. The CNN module excels at capturing local correlation features and spatial patterns reflected by adjacent frequency points in the EIS spectrum; the LSTM module specializes in learning the long-term evolution and dependencies of these feature sequences over time. This organic combination endows the model with both powerful spatial feature abstraction and temporal dynamic modeling capabilities, resulting in more accurate and robust battery SOH estimation. Attached Figure Description
[0013] Figure 1 The figure shown is an EIS Nyquist plot of a lithium-ion battery at different SOH stages according to an embodiment of the present invention.
[0014] Figure 2 The diagram shown is a schematic diagram of a convolutional autoencoder structure provided according to an embodiment of the present invention.
[0015] Figure 3 The diagram shown is a structural schematic of a combined convolutional neural network model and a recurrent neural network model provided according to an embodiment of the present invention.
[0016] Figure 4 The figure shown is a comparison between the estimated SOH of the test battery and the actual value according to an embodiment of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0018] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.
[0019] Example 1: This invention discloses a battery health state estimation method based on a convolutional autoencoder, comprising: S1: Acquire electrochemical impedance spectroscopy (EIS) data from the battery and process the data. A convolutional autoencoder (CAE) is used to automatically extract overcomplete features from the EIS data. Specifically, acquiring the EIS data includes: acquiring EIS data over multiple consecutive charge-discharge cycles, and acquiring the real and imaginary parts of the impedance when the battery reaches its electrochemical steady state. Processing the EIS data includes: performing standardization preprocessing on the acquired EIS data; calculating the maximum, minimum, mean, and standard deviation of the real and imaginary parts, and concatenating them with the acquired EIS data to enhance information representation; reshaping the concatenated data into a four-dimensional tensor to adapt to the input of the CAE. Automatically extracting overcomplete features from the EIS data using a CAE includes: automatically extracting features from the four-dimensional tensor using a trained CAE, and extracting overcomplete feature vectors from the encoder portion of the CAE. Furthermore, the convolutional autoencoder of this invention includes an encoder and a decoder. The encoder includes multiple two-dimensional convolutional layers for layer-by-layer feature extraction of the input four-dimensional tensor features. The kernel size of each convolutional layer is selected according to the structure of the input four-dimensional tensor features, and the kernel size covers the main spatial range of the input four-dimensional tensor features. The number of filters in the convolutional layers is set according to the requirements of the input four-dimensional tensor features. Each convolutional layer uses a non-linear activation function. After processing by the encoder, the input four-dimensional tensor features are mapped to four-dimensional overcomplete features, which are then transformed into one-dimensional feature vectors through feature unrolling operations for use in temporal modeling and health status estimation.
[0020] S2: Stack overcomplete features from multiple consecutive cycles to form a short-term time-dependent temporal feature sequence; S3: Input the temporal feature sequence into a convolutional neural network model and a recurrent neural network model respectively. The convolutional model extracts the local correlation features of the temporal feature sequence, and the recurrent neural network model is used to extract the time dependence features of the temporal feature sequence. Specifically, the convolutional neural network model includes at least two convolutional layers and corresponding pooling layers to extract local correlation features from the input temporal feature sequence. The convolutional neural network model includes a two-dimensional convolutional neural network, a residual neural network, or a variant thereof. The recurrent neural network includes a multi-layer recurrent neural network structure. The number of layers and neurons in each layer are set according to the complexity of the temporal features and the modeling settings. The recurrent neural network includes a long short-term memory network (LSTM), a gated recurrent unit (GRU), or a variant thereof.
[0021] S4: Concatenate the locally relevant feature vectors and time-dependent feature vectors output by the convolutional model and the recurrent neural network model, respectively, and input the concatenated feature vector into the DNN model to estimate the maximum available capacity of the current battery. Specifically, the DNN model includes at least one fully connected layer, which is used to fuse features from different model branches and estimate the battery capacity. The number of layers and neurons in each fully connected layer is set according to the feature fusion requirements. The output layer of the DNN model is used to generate the estimated value of the battery capacity.
[0022] S5: Estimate the SOH of the battery by using the ratio of the estimated current maximum available capacity of the battery to the rated capacity of the battery.
[0023] Furthermore, in order to better understand the process of battery SOH estimation in this invention, specific embodiments will be described below.
[0024] In one embodiment, using a lithium battery as an example (though those skilled in the art should understand that this is merely an example of a lithium battery, and the method of the present invention is not limited to lithium batteries), the lithium-ion battery is first subjected to a cycle aging test. Each cycle involves charging in a 1C constant current / constant voltage mode to a cutoff voltage of 4.2V, followed by discharging in a 2C constant current mode to 3V. After charging is complete, the circuit is disconnected and the battery is allowed to stand to allow the polarization effect to fully relax. After standing for 15 minutes, EIS measurements are performed. The measurement frequency range is set from 0.02Hz to 20kHz, with 10 frequency points recorded every ten octaves, for a total of 60 frequency points of electrochemical impedance spectroscopy data. Figure 1 The Nyquist plot shown indicates that there are significant differences in the EIS curves under different SOH conditions, demonstrating the effectiveness of EIS data as a health indicator.
[0025] After collecting the electrochemical impedance spectroscopy (EIS) data of the battery and performing preprocessing steps, it was used as input to the model. First, the real and imaginary parts of the EIS data for each sample, for example, 60 points each, were concatenated along the vertical axis to form a 120-dimensional vector. For n cycles, the data shape was (n, 120). The concatenated data was standardized using the z-score method to eliminate the influence of dimensions. Further, the standardized data was separated into real and imaginary channels, and the data shape was reshaped to (n, 60, 2). Four statistical features—minimum, maximum, mean, and standard deviation—were calculated for each sample's real and imaginary data. These four features were then concatenated along the vertical axis as additional features, expanding the data shape to (n, 64, 2). Finally, the data was reshaped into a four-dimensional tensor feature with a shape of (n, 8, 8, 2), which served as input to the subsequent two-dimensional convolutional layer.
[0026] like Figure 1 The figure shown is the EIS Nyquist plot of the electrochemical impedance spectroscopy data of the lithium-ion battery at different SOH stages according to an embodiment of the present invention.
[0027] Figure 2 The diagram illustrates a convolutional autoencoder structure according to an embodiment of the present invention. Specifically, the convolutional autoencoder includes an encoder and a decoder. The encoder performs feature compression on the input tensor, mapping an input tensor of shape (n, 8, 8, 2) to an overcomplete feature representation of shape (n, 8, 8, 4) through multiple convolutional operations. The decoder then reconstructs this latent representation, restoring it to a shape consistent with the original input. During the feature extraction stage, the (n, 8, 8, 4) overcomplete features output by the encoder are flattened into a one-dimensional feature vector through a Flatten layer, with a vector dimension of 8 × 8 × 4 = 256. To further incorporate temporal information, the 256-dimensional feature vectors obtained from each loop are stacked in cyclic order. In this embodiment, features from five consecutive loops are selected to form a temporal feature sequence of shape (n, 5, 256), which will serve as the input to the subsequent combined model.
[0028] Figure 3The diagram illustrates the structure of a combined convolutional neural network (CNN) model and a recurrent neural network (LSTM) model according to an embodiment of the present invention. Specifically, the CNN model is a CNN model, and the LSTM model is an LSTM model. In this embodiment, a temporal feature sequence of (n, 5, 256) is input into the CNN model and the LSTM model, respectively. In the CNN model, a channel dimension is first added to the input sequence to reshape it to a shape of (n, 5, 256, 1) to meet the input requirements of a two-dimensional convolutional layer. The processed sequence is then sequentially input into two convolutional layers for feature extraction, and flattened by a Flatten layer to output a feature vector of dimension 512. In the LSTM model, the input sequence is directly input into two LSTM layers, ultimately outputting a feature vector of dimension 64. The 512-dimensional vector output by the CNN model is then concatenated with the 64-dimensional vector output by the LSTM model, and the resulting combined feature vector is input into a DNN model composed of two fully connected layers. Finally, the output layer consists of a single neuron, which uses a linear activation function to estimate the battery capacity. The estimated state of charge (SOH) of the battery is then calculated by the ratio of the current capacity to the rated capacity.
[0029] Figure 4 The figure shows a comparison between the estimated SOH of the battery according to an embodiment of the present invention and the actual value. It can be seen from the figure that the SOH estimation curve output by the combined model is highly consistent with the actual SOH degradation curve of the battery. Furthermore, the method of the present invention was compared with baseline methods such as those using only a CNN model and those using only an LSTM model on the same dataset, and the results are shown in Table 1. The method of the present invention outperforms the comparative methods in several key evaluation indicators, including root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²), demonstrating higher estimation accuracy and reliability.
[0030] Table 1
[0031] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to imply the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0032] Example 2: The present invention also discloses an apparatus for implementing the battery health state estimation method based on a convolutional autoencoder in the embodiments, comprising: Data acquisition module: used to acquire electrochemical impedance spectroscopy (EIS) data of the battery and process the EIS data. It uses a convolutional autoencoder to automatically extract complete features from the EIS data. Feature fusion module: used to stack overcomplete features from multiple consecutive cycles to form a short-term time-dependent temporal feature sequence; Feature extraction module: Used to input time-series feature sequences into convolutional neural network model and recurrent neural network model respectively. The convolutional model extracts local correlation features of time-series feature sequences, and the recurrent neural network model is used to extract time-dependent features of time-series feature sequences. Feature concatenation module: This module concatenates the locally relevant feature vectors and time-dependent feature vectors output by the convolutional model and the recurrent neural network model, respectively, and inputs the concatenated feature vector into the DNN model to estimate the current battery's maximum available capacity. Calculation module: Used to estimate the SOH of the battery by taking the ratio of the estimated current maximum available capacity of the battery to the rated capacity of the battery.
[0033] The specific process is described in Example 1, and will not be repeated here to avoid redundancy.
[0034] In summary, the battery health state estimation method and apparatus based on a convolutional autoencoder disclosed in this invention achieves automated and intelligent feature extraction from battery electrochemical impedance spectroscopy (EIS) data. By introducing a convolutional autoencoder, this invention can directly and automatically learn and extract effective overcomplete features from the raw EIS data. This method completely avoids the tedious process and subjective limitations of manual feature design relying on expert prior knowledge in traditional technical approaches, automating feature engineering and extracting more comprehensive features that better characterize the essential information of battery aging.
[0035] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A battery health state estimation method based on a convolutional autoencoder, characterized in that, include: S1: Collect electrochemical impedance spectroscopy (EIS) data of the battery, process the EIS data, and use a convolutional autoencoder to automatically extract overcomplete features from the EIS data. S2: Stack overcomplete features from multiple consecutive cycles to form a short-term time-dependent temporal feature sequence; S3: Input the time-series feature sequence into the convolutional neural network model and the recurrent neural network model respectively. The convolutional neural network model extracts the local correlation features of the time-series feature sequence, and the recurrent neural network model is used to extract the time-dependent features of the time-series feature sequence. S4: Concatenate the locally relevant feature vectors and time-dependent feature vectors output by the convolutional neural network model and the recurrent neural network model respectively, and input the concatenated feature vector into the DNN model to estimate the current battery's maximum available capacity. S5: Estimate the SOH of the battery by using the ratio of the estimated current maximum available capacity of the battery to the rated capacity of the battery.
2. The battery health state estimation method based on a convolutional autoencoder according to claim 1, characterized in that, The specific steps of collecting the electrochemical impedance spectroscopy data of the battery in step S1 include: collecting the electrochemical impedance spectroscopy data of the battery during multiple consecutive charge-discharge cycles, and collecting the real and imaginary parts of its impedance data when the battery reaches an electrochemically stable state.
3. The battery health state estimation method based on a convolutional autoencoder according to claim 1, characterized in that, The specific processing of electrochemical impedance spectroscopy data in step S1 includes: performing standardization preprocessing on the acquired electrochemical impedance spectroscopy data; calculating the maximum, minimum, mean, and standard deviation of the real and imaginary parts of the data respectively, and concatenating them with the acquired electrochemical impedance spectroscopy data to enhance information expression; and reshaping the concatenated data into a four-dimensional tensor to adapt to the input of the convolutional autoencoder.
4. The battery health state estimation method based on a convolutional autoencoder according to claim 3, characterized in that, Step S1 uses a convolutional autoencoder to automatically extract overcomplete features from the electrochemical impedance spectroscopy data. Specifically, this includes: using a trained convolutional autoencoder to automatically extract features from the four-dimensional tensor, and extracting overcomplete feature vectors from the encoder part of the convolutional autoencoder.
5. The battery health state estimation method based on a convolutional autoencoder according to claim 4, characterized in that, The convolutional autoencoder includes an encoder and a decoder. The encoder comprises multiple two-dimensional convolutional layers for layer-by-layer feature extraction from the input four-dimensional tensor features. The kernel size of each convolutional layer is selected based on the structure of the input four-dimensional tensor features, covering the main spatial range of the input four-dimensional tensor features. The number of filters in the convolutional layers is set according to the representation requirements of the input four-dimensional tensor features. Each convolutional layer employs a non-linear activation function. After processing by the encoder, the input four-dimensional tensor features are mapped to four-dimensional overcomplete features. The feature vector is transformed into a one-dimensional feature vector through feature expansion, which is then used for time series modeling and health status estimation.
6. The battery health state estimation method based on a convolutional autoencoder according to claim 4, characterized in that, The decoder includes multiple transposed convolutional layers and convolutional layers, used to recover the spatial structure of the input four-dimensional tensor features layer by layer. The kernel size and number of filters of each transposed convolutional layer are matched with the encoder structure.
7. The battery health state estimation method based on a convolutional autoencoder according to claim 1, characterized in that, The convolutional neural network model described in step S3 includes at least two convolutional layers and pooling layers corresponding to the convolutional layers, used to extract local relevant features from the input temporal feature sequence. The convolutional neural network model includes a two-dimensional convolutional neural network, a residual neural network, or a variant thereof.
8. The battery health state estimation method based on a convolutional autoencoder according to claim 1, characterized in that, The recurrent neural network described in step S3 includes a multi-layer recurrent neural network structure. The number of layers and neurons in each layer are set according to the complexity of the temporal features and the modeling settings. The recurrent neural network includes a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU), or a variant thereof.
9. The battery health state estimation method based on a convolutional autoencoder according to claim 1, characterized in that, In step S4, the DNN model includes at least one fully connected layer, which is used to fuse features from different model branches and estimate the battery capacity. The number of layers and neurons in each fully connected layer is set according to the feature fusion requirements. The output layer of the DNN model is used to generate the estimated battery capacity.
10. An apparatus for implementing the battery health state estimation method based on a convolutional autoencoder as described in any one of claims 1-9, characterized in that, include: Data acquisition module: used to acquire electrochemical impedance spectroscopy (EIS) data of the battery and process the EIS data. It uses a convolutional autoencoder to automatically extract complete features from the EIS data. Feature fusion module: used to stack overcomplete features from multiple consecutive cycles to form a short-term time-dependent temporal feature sequence; Feature extraction module: Used to input time-series feature sequences into convolutional neural network model and recurrent neural network model respectively. Convolutional neural network model extracts local correlation features of time-series feature sequences, and recurrent neural network model is used to extract time-dependent features of time-series feature sequences. Feature concatenation module: This module concatenates the locally relevant feature vectors and time-dependent feature vectors output by the convolutional neural network model and the recurrent neural network model, respectively, and inputs the concatenated feature vector into the DNN model to estimate the current battery's maximum available capacity. Calculation module: Used to estimate the SOH of the battery by taking the ratio of the estimated current maximum available capacity of the battery to the rated capacity of the battery.