Lithium battery health state assessment method and system

Abstract

The invention belongs to the technical field of battery health state evaluation, and discloses a lithium battery health state evaluation method, which comprises the following steps of: acquiring charging and discharging data of a lithium battery and preprocessing; carrying out feature extraction on the preprocessed charging and discharging data to obtain a plurality of original features representing the state of the battery, and carrying out filtering processing on the original features to obtain filtered features; carrying out importance evaluation on the filtered features, screening out a key feature subset, and dividing the key feature subset into a training set and a test set; constructing a hybrid prediction model, and training the hybrid prediction model by using the training set to obtain a trained lithium battery health state prediction model; and verifying the trained lithium battery health state prediction model by using the test set, and evaluating the health state of the lithium battery to be tested by using the verified model. According to the method, the features can be adaptively screened, the nonlinear degradation process can be accurately modeled, and the accuracy and robustness of lithium battery health state evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of battery health status assessment technology, and in particular to a method and system for assessing the health status of lithium batteries. Background Technology

[0002] Lithium-ion batteries, due to their high energy density and long cycle life, have become the core power source for electric vehicles, energy storage systems, and portable electronic devices. However, batteries undergo performance degradation during cycling, primarily manifested as capacity decay and increased internal resistance, directly impacting system safety and reliability. State of health, as a key indicator for quantifying the degree of battery degradation, is crucial for accurate battery management.

[0003] Existing SOH assessment methods are mainly divided into three categories: direct measurement methods, model-based methods, and data-driven methods, all of which have significant limitations. While direct measurement methods are computationally simple, they rely on high-precision sensors, making online implementation difficult and unable to capture the dynamic changes in the internal electrochemical processes of the battery. Model-based methods (such as electrochemical models and equivalent circuit models) have numerous parameters, complex calculations, poor robustness, insufficient generalization ability, and are sensitive to fluctuations in operating conditions. Data-driven methods, especially those based on machine learning, have become the mainstream, but they also face two major bottlenecks: first, feature selection defects—features selected by traditional linear methods are difficult to adapt to the strong nonlinearity of neural networks, potentially leading to the omission of key features or the retention of redundant features, reducing model efficiency; second, insufficient handling of capacity recovery phenomena—the transient capacity rebound caused by SEI film repair and other effects during battery aging is weak, nonlinear, and transient, making it difficult for traditional models to capture such patterns, resulting in significantly increased prediction errors during the recovery phase.

[0004] Therefore, there is an urgent need to develop an innovative method that can adaptively screen high-contribution features and accurately model nonlinear degradation dynamics in order to achieve high-precision and robust SOH assessment. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method and system for assessing the health status of lithium batteries. This method can adaptively select features and accurately model nonlinear degradation processes, thereby improving the accuracy and robustness of lithium battery health status assessment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for assessing the health status of a lithium battery, comprising the following steps: Acquire the charge and discharge data of the lithium battery and preprocess the charge and discharge data; Feature extraction is performed on the preprocessed charge and discharge data to obtain multiple original features characterizing the battery state. The original features are then filtered to obtain filtered features. The importance of filtered features is evaluated based on the model class dependency method, a subset of key features is selected, and then divided into training and test sets. A hybrid prediction model comprising a convolutional neural network module, a temporal network module, and an adaptive state optimization module is constructed. The hybrid prediction model is trained using a training set to obtain a trained lithium battery health state prediction model. The trained lithium battery health status prediction model was validated using a test set, and the validated model was used to evaluate the health status of the lithium battery under test.

[0007] As an optional implementation method, charge and discharge data of the lithium battery are acquired, and the charge and discharge data are preprocessed, specifically including: Acquire charging and discharging data of lithium batteries during the constant current charging phase, when the voltage is within a preset range; The charging and discharging data are interpolated or resampled to a fixed number of sampling points; Based on the unified data, multiple features were calculated, including voltage sequence, charging time sequence, temperature sequence, incremental capacity curve, and differential voltage analysis curve.

[0008] As an alternative implementation, the original features are filtered, specifically by using the Savitzky-Golay filtering method to smooth and reduce noise on the incremental capacity curve and the differential voltage analysis curve, respectively.

[0009] As an alternative implementation, the importance of the filtered features is evaluated based on a model class dependency method to select a subset of key features, specifically including: The time-series-divided training set is input into the hybrid prediction model for training; The baseline prediction error was calculated using the test set without feature replacement. Construct a replacement feature matrix, and replace each feature in the test set individually in turn, calculating the prediction error after each replacement; Based on the change in prediction error after feature replacement relative to the baseline prediction error, the importance index of each feature is calculated and ranked accordingly. Based on the ranking results, the features with the highest importance ranking are selected to form a subset of key features.

[0010] As an alternative implementation, in the hybrid prediction model, the temporal network module is a bidirectional long short-term memory network, and the adaptive state optimization module is an adaptive state extractor. The adaptive state extractor processes historical and future information separately and uses a gating mechanism to adaptively fuse information in order to optimize feature representation and strengthen core features related to health status.

[0011] As an alternative implementation, the training process of the hybrid prediction model employs a dynamically adjusted learning rate scheduling strategy and an early stopping mechanism, and uses mean squared error as the loss function.

[0012] Secondly, the present invention provides a lithium battery health status assessment system, comprising: The data acquisition and preprocessing module is configured to acquire the charging and discharging data of the lithium battery and preprocess the charging and discharging data. The feature extraction module is configured to: extract features from the preprocessed charge and discharge data to obtain multiple original features characterizing the battery state, and filter the original features to obtain filtered features. The feature selection module is configured to: evaluate the importance of filtered features based on the model class dependency method, select a subset of key features, and divide them into training and test sets; The model building and training module is configured to: build a hybrid prediction model that includes a convolutional neural network module, a temporal network module and an adaptive state optimization module; train the hybrid prediction model using the training set to obtain the trained lithium battery health state prediction model. The health status assessment module is configured to: validate the trained lithium battery health status prediction model using a test set, and use the validated model to assess the health status of the lithium battery under test.

[0013] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0014] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0015] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The proposed lithium battery health status assessment method creatively integrates the Model Class Dependency (MCR) feature selection method with the Adaptive State Refiner (ASR) module. The MCR method, by simulating the model's decision-making process, adaptively selects a subset of key features that contribute most to SOH prediction from numerous features, effectively eliminating redundant and noisy features. This provides the model with a cleaner, more information-dense input, fundamentally improving the model's expressive power and generalization potential. The ASR module, through its unique bidirectional independent path and gating mechanism, can precisely capture the temporal dependencies during battery aging, particularly demonstrating excellent modeling capabilities for weak, nonlinear capacity recovery phenomena caused by electrochemical relaxation effects, which are difficult for traditional models to handle. The synergistic effect of these two core technologies enables the hybrid model of this invention to maintain high-accuracy SOH assessment even under complex operating conditions and individual battery differences, significantly enhancing the model's robustness.

[0017] The lithium battery health status assessment method proposed in this invention significantly reduces the dimensionality of input data while maintaining prediction accuracy through MCR feature selection, thereby alleviating the computational burden on the model and improving training and inference speed. Simultaneously, the model architecture (CNN-BiLSTM-ASR) is rationally designed. The CNN module efficiently extracts local features, the BiLSTM module fully mines sequential information, and the ASR module refines features for SOH prediction. Each module has a clear division of labor, resulting in high overall computational efficiency. The training strategies employed, such as dynamic learning rate, early stopping, Dropout, and L2 regularization, effectively prevent overfitting, ensuring the stability and reliability of the model in practical deployments, and providing a feasible technical solution for online or embedded BMS applications.

[0018] The proposed lithium battery health status assessment method is independent of specific battery models or complex electrochemical model parameters. Its data-driven nature allows it to learn general patterns from historical battery data from various sources (such as public datasets or experimental data). Provided with standardized charging segment data, this method can be trained and applied to SOH assessment of different types of lithium batteries, demonstrating good versatility and potential for widespread application.

[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1The present invention provides a flowchart of a lithium battery health status assessment method. Figure 1 ; Figure 2 The present invention provides a flowchart of a lithium battery health status assessment method. Figure 2 . Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed description is exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Example 1 like Figures 1 to 2 As shown, this embodiment provides a method for assessing the health status of a lithium battery, including the following steps: Step 1: Obtain the charge and discharge data of the lithium battery and preprocess the charge and discharge data; Step 2: Extract features from the preprocessed charge and discharge data to obtain multiple original features characterizing the battery state, and then filter the original features to obtain filtered features. Step 3: Evaluate the importance of the filtered features based on the model class dependency method, select a subset of key features, and divide them into training and test sets; Step 4: Construct a hybrid prediction model that includes a convolutional neural network module, a temporal network module, and an adaptive state optimization module. Train the hybrid prediction model using the training set to obtain the trained lithium battery health state prediction model. Step 5: Validate the trained lithium battery health status prediction model using the test set, and use the validated model to evaluate the health status of the lithium battery under test.

[0027] The specific solution of the present invention is as follows: Step 1 specifically includes: Publicly available datasets (such as NASA's B0005, B0006, B0007, and B0018 battery data) should be used preferentially. These datasets must contain complete charge-discharge cycle records. If experimental data is used, high-precision acquisition equipment (voltage sensor accuracy no less than ±0.1mV, current sensor accuracy ±0.05%) must be used, and the data must be collected in a constant temperature environment (24℃±0.5℃).

[0028] Data from the "constant current charging stage" was extracted. Using voltage as the criterion, the data segment with voltage ranging from 3.2V to 4.2V was selected, and data from the constant voltage stage was removed. Finally, the number of sampling points was calibrated. Since subsequent feature calculations require a unified dimension, the constant current charging data for each effective cycle was adjusted to 320 sampling points. If the original data points exceeded 320, they were extracted at equal intervals from beginning to end. If there were fewer than 320 points, linear interpolation based on voltage was used to supplement the data, ensuring that the final data still completely covers the 3.2V-4.2V range. The specific values ​​mentioned above were obtained experimentally.

[0029] Step 2 specifically includes: First, based on the filtered valid data, five features are calculated sequentially: The first feature is the voltage sequence (V k The data is derived from the voltage data during the constant current charging phase. After extracting the raw voltage data of this cyclic constant current phase, sampling is performed according to the rule of "0.003125V equal interval": starting from 3.2V, a corresponding voltage value is selected for every 0.003125V increase, and so on until 4.2V, ensuring that the voltage difference between each adjacent sampling point is 0.003125V, finally obtaining 320 voltage values ​​arranged in chronological and voltage increment order; The second feature is the charging time series (t) k The data is synchronized with the voltage sequence and comes from the cumulative time data of the same cycle. First, the cumulative time value corresponding to each of the 320 sampling points of the voltage sequence is extracted (that is, the cumulative charging time corresponding to each voltage sampling point with a spacing of 0.003125V is recorded synchronously).

[0030] The third feature is the temperature sequence (T) k The data comes from battery surface temperature data of the same cycle. The temperature values ​​corresponding to the 320 sampling points with a spacing of 0.003125V in the voltage sequence are extracted.

[0031] The fourth characteristic is the incremental capacity curve (IC), which needs to be calculated by combining the voltage sequence, charging time sequence, and constant current stage current data of the same cycle. The core of incremental capacity is the ratio of the change in charging capacity to the change in voltage. First, the average value of the current in the constant current stage is calculated (approximately 1.5A). Then, the time difference between 320 adjacent sampling points in the charging time sequence is calculated. The capacity change is obtained by "current × time difference" and converted to milliampere-hours. Since the voltage sequence has been extracted at equal intervals, the voltage change between adjacent points is fixed at 0.003125V. The 319 IC values ​​are obtained by "capacity change / voltage change". The formula is as follows: ; in, Indicates when the voltage from k +1 becomes k Changes in capacity over time Indicates from k Voltage change from sampling point +1 to sampling point k Indicates from k +1 to k The time interval between sampling points, I Represents current. k Indicates the number of samples. t Indicates the sampling time.

[0032] The fifth feature is the differential voltage analysis curve (DVA), the core of which is the "differentiation of voltage with capacity." It needs to be calculated by combining the voltage sequence and cumulative capacity data of the same cycle to reflect the rate of voltage change with capacity. First, the cumulative capacity of each sampling point is calculated: based on the average current during the constant current phase (approximately 1.5A) and the cumulative charging time, it is converted to milliampere-hours (mAh) using the formula "current × cumulative charging time / 3.6," resulting in a cumulative capacity sequence corresponding to 320 voltage sampling points. Then, the voltage difference and capacity difference between adjacent sampling points are calculated, and 319 DVA values ​​are calculated using the formula. ; Secondly, the data to be filtered is loaded and verified. From the generated five feature datasets, the incremental capacity curve and differential voltage analysis curve data for each effective cycle are extracted. It needs to be confirmed that both types of curves are one-dimensional numerical sequences containing 319 data points, and each data point corresponds one-to-one with the voltage sequence sampling point. If the curve data for a certain cycle is missing, it is filled in by the mean of the corresponding data from the two adjacent cycles of that battery, ensuring that there are no data gaps in subsequent filtering calculations. Then, the filtering parameters are configured using a fixed and experimentally verified combination: the window length is set to 41, the polynomial order is set to 4, and the boundary processing mode is determined to be "interpolation completion" to properly handle the numerical distortion problem that may be caused by incomplete windows at the beginning and end of the curve, laying the foundation for the stability and repeatability of the filtering effect.

[0033] Next, the incremental capacity curve is filtered. The Savitzky-Golay filtering module in the scientific computing software is called, the incremental capacity curve data is imported, and the configured parameters are entered. During the filtering process, a sliding window is formed by selecting 20 data points before and after each data point. A fourth-order polynomial is used to perform least-squares fitting on the data within the window. The original data values ​​are then replaced with the values ​​of the fitted curve at the center data point. After traversing 319 data points point by point, the filtered incremental capacity curve is generated.

[0034] The filtering process for the differential voltage analysis curve is the same as that for the incremental capacity curve. The filtering module is called in the same way, and the input data is replaced with the differential voltage analysis curve data. The parameters and operation process remain unchanged. Through fourth-order polynomial fitting and sliding window traversal, the filtering calculation of 319 data points is completed to obtain the filtered differential voltage analysis curve.

[0035] Step 3 specifically includes: For the complete feature data (including voltage sequence, charging time sequence, temperature sequence, incremental capacity curve, and differential voltage analysis curve) after filtering in step 2, the training set and test set are divided into time series, strictly following the principle of "time series continuity": First, the data is sorted in the order of "battery number → original cycle number" to ensure that all samples completely match the actual aging process from the early cycle to the late cycle of the battery; then, the data is truncated at a fixed ratio of 70% / 30%, with the first 70% of the time series samples used as the training set and the last 30% of the time series samples used as the test set.

[0036] The base model (i.e., the hybrid prediction model) is trained and the baseline error is calculated. A CNN-BiLSTM architecture is used to build the base neural network model (i.e., the hybrid prediction model). The 70% training set is input into the model, and the root mean square error (RMSE) is used as the loss function. The batch size is set to 32, the initial learning rate is 0.001, and the maximum number of training epochs is 100. An early stopping mechanism is enabled (training is terminated if the loss does not decrease for 10 consecutive epochs) to ensure that the model fully learns the mapping relationship between "features and SOH" in the training set. After the base model is trained, the 30% original test set without any replacement is input into the model, and forward propagation is performed to obtain the SOH prediction value. The error value eorigin (baseline error) is calculated according to the RMSE formula. This value serves as the core reference for subsequent feature importance evaluation.

[0037] ; Subsequently, a replacement feature matrix was constructed and feature replacement was performed. Feature data from the fifth cycle was extracted from the original data in the NASA dataset and integrated into a replacement matrix H in a "319×5" format, with each of the five columns corresponding to one of the five features. Independent replacements were performed on each of the five features in the test set, changing only one target feature at a time while retaining the original test data for the other four features: the voltage sequence in the test set was replaced with the first column of H to generate the replacement test set X1; the charging time sequence in the test set was replaced with the second column of H to generate X2; the temperature sequence in the test set was replaced with the third column of H to generate X3; the incremental capacity curve column in the test set was replaced with the fourth column of H to generate X4; and the differential voltage analysis curve column in the test set was replaced with the fifth column of H to generate X5.

[0038] After the replacement is completed, the error after each replacement is calculated. X1 to X5 are input into the trained base model respectively, and forward propagation is performed under the same test environment. The five replacement error values ​​e1 to e5 are calculated sequentially according to the RMSE formula. The error after each feature replacement is recorded as a percentage of the baseline error e. origin The range of change.

[0039] Then, the feature importance index is calculated and sorted according to the formula Ip=(e j -e origin ) / e origin ×100% Quantify the importance of each feature. Based on the experimental results of the NASA dataset, the Ip ranking of the five features is as follows: incremental capacity curve > voltage sequence > charging time series > temperature sequence > differential voltage analysis curve. Based on this, a feature importance ranking table is established.

[0040] Finally, key features were selected and the preprocessed dataset was divided. The top three most important features were retained, forming a preprocessed dataset in the format of "number of samples × 319 × 3" (feature channels are voltage sequence, charging time sequence, and incremental capacity curve in order). Temperature sequence and differential voltage analysis curve were removed (the Ip value is significantly low, the information is redundant, and the improvement to model accuracy is limited). Then, according to the previously determined ratio, the preprocessed dataset was divided into training set and test set.

[0041] Step 4 specifically includes: Step 4-1, detailed design of the model architecture.

[0042] The input layer strictly adapts to the preprocessed dataset from step 3. The input data consists of three key features after filtering (voltage sequence, charging time sequence, and incremental capacity curve), with a fixed data dimension of (number of samples × 319 × 3). The input data must first undergo Z-score standardization, using the following formula: ,in and These are the mean and standard deviation of each feature on the training set, respectively, to eliminate dimensional differences and ensure the stability of model training convergence.

[0043] The Convolutional Neural Network (CNN) module is responsible for extracting local features: a one-dimensional convolutional layer is configured with 32 convolutional kernels, a kernel size of 5, a stride of 1, and uses "same" padding to maintain the same output dimension as the input. The ReLU activation function is chosen to introduce non-linearity. The convolution operation formula is as follows: ,in, x c Indicates the first c One channel (c=1 for voltage, c=2 for time, c=3 for IC value). w j It is the first j The weight matrix of each convolutional kernel has dimensions (s, 3). b j It is the first j Bias terms for each convolution kernel. Z j ( v ) represents the linear output of the convolution operation. v This represents the time step at the current time. s This represents the kernel size. A one-dimensional max-pooling layer is then applied, with a pooling window size of 2 and a stride of 2, to reduce the dimensionality of the convolutional features and enhance their robustness.

[0044] The Bidirectional Long Short-Term Memory (BiLSTM) module is used to capture temporal dependencies: the forward and backward LSTM units are symmetrical in parameters, each with 64 units, and the hidden state dimension is 128. The gating mechanism is calculated using the following formula: Forget Gate Input gate Candidate cell status Cell state update Output gate Unit output The forward and backward hidden states are concatenated at each time step, and the final output is a feature sequence. f t It is the result of the Gate of Oblivion. i t It is the result of the input gate. σ This is the sigmoid function, which compresses the output to the (0,1) interval. ω f It is the weight matrix of the forget gate. ω i It is the weight matrix of the input gate. b f It is the bias vector of the forget gate. b iIt is the bias vector of the input gate. x t It is the input value at the current moment. Y t 1 is the output of the previous adaptive refining unit.

[0045] The Adaptive State Refiner (ASR) module optimizes features through bidirectional independent paths: the forward path processes historical information, and the reverse path processes future information, avoiding early feature confusion. The gating system calculates the following formula: Reset gate (controls the degree of historical information retention). Update gate (determines the importance of candidate states) Candidate status Final output This module enhances the ability to capture capacity recovery phenomena by weighted fusion of the original state and candidate state, filtering out redundant information, strengthening SOH-related features, and improving the ability to capture capacity recovery phenomena. r t This is the result of resetting the door. z t This is the result of updating the door. σ This is the sigmoid function, which compresses the output to the (0,1) interval. ω r It is the weight matrix of the reset gate. ω z It updates the weight matrix of the gate. b r It is the bias vector for resetting the gate. b z It is the bias vector for updating the gate. h t It is the output of the LSTM unit.

[0046] The output regression layer is responsible for feature integration and prediction: the fully connected layer has 64 neurons with ReLU activation function, performing non-linear integration of the features output by the ASR module; the final output layer uses single-neuron linear activation, directly outputting the SOH percentage value in the range of 0-1, and the loss function is mean squared error (MSE), as shown in the formula. It is used for supervised optimization of the training process.

[0047] Step 4-2, Training Parameter Configuration and Optimization Strategies The model training employs a combination of the Adam optimizer and the MSE loss function to balance convergence speed and stability. The learning rate scheduling strategy is dynamically adjusted based on the validation set performance: the initial learning rate is set to 0.001, and when the validation loss does not decrease for 10 consecutive rounds, the learning rate decays to 0.0001, with a lower bound set to 1e-6 to avoid gradient vanishing.

[0048] The hyperparameters were fixed after experimental optimization: the batch size was 32 to ensure the temporal continuity of data within the batch and training efficiency; the maximum number of training epochs was 200, and an early stopping mechanism was enabled. If the validation loss did not decrease for 15 consecutive epochs, training was automatically terminated to avoid resource waste and overfitting. Regularization strategies included Dropout (with a rate of 0.2, applied to the output of CNN-BiLSTM layers) and L2 weight decay (coefficient 1e-4, added to the loss function) to doubly suppress overfitting.

[0049] Step 4-3, Training Execution Process Load the training set divided in step 3, divide the training set into multiple batches according to the preset batch size, and each batch contains complete charging cycle data; perform Z-score standardization based on the pre-calculated mean and standard deviation of the training set to avoid data leakage.

[0050] Batch data is sequentially input into each module of the model for forward propagation: the CNN module extracts key local information of features, and then pooling operations are used to reduce dimensionality and enhance robustness; then the feature sequence is input into the BiLSTM module, where forward and backward units work together to capture bidirectional temporal dependencies; finally, the output of BiLSTM is sent to the ASR module, where a gating mechanism is used to adaptively optimize feature representation, filter out redundant information, and strengthen core features related to SOH.

[0051] The mean squared error (MSE) loss function is used to calculate the deviation between the predicted SOH and the true SOH, while a regularization strategy is introduced to suppress overfitting. An automatic differentiation tool is used to calculate the gradient of the loss with respect to all trainable weights of the model. The Adam optimizer is used to update the parameters of the CNN, BiLSTM, and ASR modules, completing one round of parameter iteration. The training loss is recorded after each batch processing, and traversing the entire training set completes one training round. After each training round, the model performance is evaluated using a validation set partitioned from the training set, and core metrics such as RMSE and R² are calculated. A checkpointing mechanism is employed to save only the model parameters with the best performance on the validation set, while also recording metadata such as the training round, loss value, and hyperparameters to ensure training reproducibility. If the validation loss does not decrease for several consecutive rounds, an early stopping mechanism is triggered to terminate training, avoiding overtraining and resource waste.

[0052] Step 5 specifically involves: First, the optimal model parameters saved in step 4 and the test set data generated in step 3 are loaded. Before testing, the data is preprocessed, performing Z-score standardization identical to the training set, using the pre-calculated mean μ and standard deviation σ from the training set to ensure data uniformity and avoid affecting validation results due to preprocessing differences. The validation process is executed on a specified hardware platform (Intel Core i7-10750H processor, 16GB RAM, NVIDIA GTX 1650Ti graphics card) and software environment (Python 3.10.18, TensorFlow 2.8.0), using batch processing with a batch size of 32 to balance efficiency and stability.

[0053] Then, the forward propagation computation of the model is initiated: the standardized test data is input into the trained model, and the model sequentially passes through the CNN module to extract local features, the BiLSTM module to capture bidirectional temporal dependencies, and the ASR module to optimize feature representation. Finally, the SOH estimate is output through the output layer to complete the prediction computation of the entire verification process.

[0054] A multi-metric comprehensive evaluation method was adopted to assess the model's performance. Key metrics included root mean square error (RMSE), mean absolute error (MAE), and goodness of fit (R²), with a focus on the model's prediction accuracy during the short-term rebound phase of battery capacity. After successful performance validation, the model was officially deployed for actual lithium battery SOH estimation.

[0055] Example 2 This embodiment provides a lithium battery health status assessment system, including: The data acquisition and preprocessing module is configured to acquire the charging and discharging data of the lithium battery and preprocess the charging and discharging data. The feature extraction module is configured to: extract features from the preprocessed charge and discharge data to obtain multiple original features characterizing the battery state, and filter the original features to obtain filtered features. The feature selection module is configured to: evaluate the importance of filtered features based on the model class dependency method, select a subset of key features, and divide them into training and test sets; The model building and training module is configured to: build a hybrid prediction model that includes a convolutional neural network module, a temporal network module and an adaptive state optimization module; train the hybrid prediction model using the training set to obtain the trained lithium battery health state prediction model. The health status assessment module is configured to: validate the trained lithium battery health status prediction model using a test set, and use the validated model to assess the health status of the lithium battery under test.

[0056] It should be noted that the above modules correspond to the steps in Embodiment 1, and the examples and application scenarios implemented by the above modules and their corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules can be executed in a computer system as part of the system.

[0057] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0058] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0059] A computer-readable storage medium for storing computer instructions that, when executed by a processor, perform the method of Embodiment 1.

[0060] The method in Example 1 can be directly executed by a hardware processor, or it can be executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0061] A computer program product includes a computer program that, when executed by a processor, implements the method in Embodiment 1.

[0062] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0063] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0064] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0065] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0066] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for assessing the health status of a lithium battery, characterized in that, Includes the following steps: Acquire the charge and discharge data of the lithium battery and preprocess the charge and discharge data; Feature extraction is performed on the preprocessed charge and discharge data to obtain multiple original features characterizing the battery state. The original features are then filtered to obtain filtered features. The importance of filtered features is evaluated based on the model class dependency method, a subset of key features is selected, and then divided into training and test sets. A hybrid prediction model comprising a convolutional neural network module, a temporal network module, and an adaptive state optimization module is constructed. The hybrid prediction model is trained using a training set to obtain a trained lithium battery health state prediction model. The trained lithium battery health status prediction model was validated using a test set, and the validated model was used to evaluate the health status of the lithium battery under test.

2. The method for assessing the health status of a lithium battery as described in claim 1, characterized in that, Acquire lithium battery charge and discharge data, and preprocess the data, specifically including: Acquire charging and discharging data of lithium batteries during the constant current charging phase, when the voltage is within a preset range; The charging and discharging data are interpolated or resampled to a fixed number of sampling points; Based on the unified data, multiple features were calculated, including voltage sequence, charging time sequence, temperature sequence, incremental capacity curve, and differential voltage analysis curve.

3. The method for assessing the health status of a lithium battery as described in claim 2, characterized in that, The original features are filtered, specifically by using the Savitzky-Golay filtering method to smooth and reduce noise on the incremental capacity curve and the differential voltage analysis curve.

4. The method for assessing the health status of a lithium battery as described in claim 1, characterized in that, The importance of filtered features is evaluated based on a model-dependent approach, and a subset of key features is selected, including: The time-series-divided training set is input into the hybrid prediction model for training; The baseline prediction error was calculated using the test set without feature replacement. Construct a replacement feature matrix, and replace each feature in the test set individually in turn, calculating the prediction error after each replacement; Based on the change in prediction error after feature replacement relative to the baseline prediction error, the importance index of each feature is calculated and ranked accordingly. Based on the ranking results, the features with the highest importance ranking are selected to form a subset of key features.

5. The method for assessing the health status of a lithium battery as described in claim 1, characterized in that, In the hybrid prediction model, the temporal network module is a bidirectional long short-term memory network, and the adaptive state optimization module is an adaptive state extractor. The adaptive state extractor processes historical and future information separately and uses a gating mechanism to adaptively fuse information in order to optimize feature representation and strengthen core features related to health status.

6. The method for assessing the health status of a lithium battery as described in claim 1, characterized in that, The training process of the hybrid prediction model adopts a dynamically adjusted learning rate scheduling strategy and an early stopping mechanism, and uses mean squared error as the loss function.

7. A lithium battery health status assessment system, characterized in that, include: The data acquisition and preprocessing module is configured to acquire the charging and discharging data of the lithium battery and preprocess the charging and discharging data. The feature extraction module is configured to: extract features from the preprocessed charge and discharge data to obtain multiple original features characterizing the battery state, and filter the original features to obtain filtered features. The feature selection module is configured to: evaluate the importance of filtered features based on the model class dependency method, select a subset of key features, and divide them into training and test sets; The model building and training module is configured to: build a hybrid prediction model that includes a convolutional neural network module, a temporal network module and an adaptive state optimization module; train the hybrid prediction model using the training set to obtain the trained lithium battery health state prediction model. The health status assessment module is configured to: validate the trained lithium battery health status prediction model using a test set, and use the validated model to assess the health status of the lithium battery under test.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images