Battery health state adaptive evaluation method and system based on transfer learning

Abstract

The invention provides a battery health state adaptive evaluation method and system based on transfer learning in the technical field of lithium battery management and state monitoring. The method comprises the following steps: S1, collecting charging voltage data of a target lithium battery to generate an incremental capacity curve; s2, taking a voltage value corresponding to the maximum increment capacity as a center reference, and adaptively selecting voltage fragment data from the charging voltage data based on a preset voltage range window; s3, fixing shared parameters used for extracting universal features in the pre-trained battery health state evaluation model, and selecting a preset number of fragment sub-data from the voltage fragment data to finely adjust the battery health state evaluation model so as to adapt to the target lithium battery; and S4, enabling the voltage fragment data to belong to the fine-tuned battery health state evaluation model, and obtaining a battery health state evaluation result. The method has the advantages that the precision, the robustness and the generalization ability of battery health state evaluation are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of lithium battery management and state monitoring technology, and in particular to an adaptive assessment method and system for battery health state based on transfer learning. Background Technology

[0002] With the widespread application of new energy vehicles, portable electronic devices, and large-scale energy storage systems, the accurate assessment and safe monitoring of the State of Health (SOH) of lithium batteries, as a key power source, is particularly important. SOH is a core indicator reflecting the degree of battery performance degradation, directly affecting the safety of device use, range, and operational reliability of energy systems. Therefore, developing efficient and accurate battery health assessment methods has significant practical implications and broad application value.

[0003] However, traditional battery SOH evaluation methods have the following technical limitations: (1) Rigid Models and Dependence on Specific Parameters: Existing methods are mostly based on pre-set, fixed electrochemical models or empirical decay formulas. Once established, these models are difficult to adjust, and their evaluation results heavily depend on specific parameters such as internal resistance and capacity obtained under specific operating conditions. When applied to lithium batteries of different batches, models, or materials, due to the performance differences between batteries, such fixed models often fail to provide accurate evaluations and lack the necessary adaptability. Essentially, they cannot utilize existing knowledge (such as a large amount of data from other battery models) to quickly adapt to new target batteries.

[0004] (2) Low adaptation efficiency for new types of batteries: In the context of rapid iteration of lithium battery technology, traditional methods often need to build a completely new evaluation model from scratch when facing new or niche battery models due to the lack of sufficient historical data. This process involves re-engineering the entire feature, designing the model structure, and training the parameters, resulting in long development cycles, high computational resource consumption, and high adaptation costs, making it difficult to meet the application requirements for rapid mass deployment.

[0005] (3) Insufficient cross-domain generalization ability: The feature extraction rules of the evaluation model trained on a specific dataset are deeply coupled with the inherent characteristics of the battery, making it difficult to transfer the "learned knowledge" to other different types of batteries. This high dependence on the training data domain makes the model's generalization ability severely insufficient when facing new domains (new battery types) with scarce data, which restricts its universality for large-scale engineering applications.

[0006] Therefore, how to provide an adaptive battery health status assessment method and system based on transfer learning to improve the accuracy, robustness and generalization ability of battery health status assessment has become an urgent technical problem to be solved. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an adaptive battery health status assessment method and system based on transfer learning, so as to improve the accuracy, robustness and generalization ability of battery health status assessment.

[0008] In a first aspect, the present invention provides an adaptive evaluation method for battery health status based on transfer learning, comprising the following steps: Step S1: Collect charging voltage data of the target lithium battery during the constant current charging stage, calculate and generate an incremental capacity curve based on the charging voltage data, and smooth the incremental capacity curve to suppress noise. Step S2: On the smoothed incremental capacity curve, find the maximum incremental capacity, and take the voltage value corresponding to the maximum incremental capacity as the center reference, and adaptively select voltage segment data from the charging voltage data based on the preset voltage range window. Step S3: Fix the shared parameters used to extract general features in the pre-trained battery health status assessment model, and select a preset number of segment sub-data from the voltage segment data to fine-tune the battery health status assessment model to adapt it to the target lithium battery. Step S4: Assign the voltage segment data to the fine-tuned battery health status assessment model to obtain the battery health status assessment result.

[0009] Furthermore, in step S1, the incremental capacity curve is smoothed using a Gaussian filtering algorithm.

[0010] Furthermore, in step S2, the width of the voltage range window is preset according to the incremental capacity curve characteristics of different types of lithium batteries to ensure that the voltage range window can cover the electrochemical characteristic range that is strongly correlated with the battery health status.

[0011] Furthermore, in step S3, the battery health status assessment model is a convolutional neural network model, and the network structure includes a convolutional layer, a pooling layer, and a fully connected layer connected in sequence. The convolutional layer is used to scan and process the input voltage segment data through the convolutional kernel, and superimpose the bias to extract battery state features; the pooling layer is used to scale the dimensions of the battery state features; the fully connected layer is used to output the battery health status assessment result based on the scaled battery state features. The shared parameters include at least the weight parameters of the convolutional layers.

[0012] Furthermore, in step S3, the battery health status assessment model is pre-trained by using a dataset constructed from historical voltage segment data annotated with real SOH as input and minimizing the mean square error between the predicted SOH and the real SOH; the predicted SOH is the battery health status assessment result.

[0013] Secondly, the present invention provides a battery health state adaptive assessment system based on transfer learning, comprising the following modules: The incremental capacity curve generation module is used to collect charging voltage data of the target lithium battery during the constant current charging stage, calculate and generate an incremental capacity curve based on the charging voltage data, and smooth the incremental capacity curve to suppress noise. The voltage segment adaptive selection module is used to find the maximum incremental capacity on the smoothed incremental capacity curve, and adaptively select voltage segment data from the charging voltage data based on the voltage value corresponding to the maximum incremental capacity as the central reference and a preset voltage range window. The transfer learning module is used to fix the shared parameters used to extract general features in the pre-trained battery health status assessment model, and to fine-tune the battery health status assessment model by selecting a preset number of sub-data segments from the voltage segment data to adapt it to the target lithium battery. The health status assessment module is used to classify the voltage segment data into a fine-tuned battery health status assessment model to obtain the battery health status assessment result.

[0014] Furthermore, in the incremental capacity curve generation module, the incremental capacity curve is smoothed using a Gaussian filtering algorithm.

[0015] Furthermore, in the voltage segment adaptive selection module, the width of the voltage range window is preset according to the incremental capacity curve characteristics of different types of lithium batteries to ensure that the voltage range window can cover the electrochemical characteristic range that is strongly correlated with the battery health status.

[0016] Furthermore, in the transfer learning module, the battery health status assessment model is a convolutional neural network model, and the network structure includes a convolutional layer, a pooling layer, and a fully connected layer connected in sequence. The convolutional layer is used to scan and process the input voltage segment data through the convolutional kernel, and superimpose the bias to extract battery state features; the pooling layer is used to scale the dimensions of the battery state features; the fully connected layer is used to output the battery health status assessment result based on the scaled battery state features. The shared parameters include at least the weight parameters of the convolutional layers.

[0017] Furthermore, in the transfer learning module, the battery health status assessment model uses a dataset constructed from historical voltage segment data annotated with real SOH as input, and is pre-trained by minimizing the mean square error between the predicted SOH and the real SOH; the predicted SOH is the battery health status assessment result.

[0018] The advantages of this invention are: 1. By collecting charging voltage data of the target lithium battery during the constant current charging stage, an incremental capacity curve is calculated and generated based on the charging voltage data. The incremental capacity curve is smoothed to suppress noise. On the smoothed incremental capacity curve, the maximum incremental capacity is found, and the voltage value corresponding to the maximum incremental capacity is used as the central benchmark. Voltage segment data is adaptively selected from the charging voltage data based on a preset voltage range window. Then, the shared parameters used to extract general features in the pre-trained battery health status assessment model are fixed. A preset number of sub-data segments are selected from the voltage segment data to fine-tune the battery health status assessment model to adapt it to the target lithium battery. Finally, the voltage segment data belongs to the fine-tuned battery health status assessment model. The model is evaluated to obtain battery health status assessment results. Specifically, a pre-trained convolutional neural network model is used to solidify the general feature extraction capabilities learned from a large amount of historical data, serving as a high-precision assessment basis. The incremental capacity curve is then smoothed, and key voltage segments strongly correlated with electrochemical characteristics are adaptively selected as input, effectively suppressing noise interference and focusing on data with the highest signal-to-noise ratio, thereby improving the accuracy of the assessment and its robustness to individual differences. Finally, by fixing the shared parameters of the model and fine-tuning it using only a small amount of data from the target lithium battery, the model can quickly adapt to different models or batches of lithium batteries, greatly enhancing its cross-domain generalization ability. Ultimately, this synergistically improves the accuracy, robustness, and generalization ability of battery health status assessment.

[0019] 2. By collecting charging voltage data during the constant current charging phase and generating incremental capacity curves, and then smoothing them (e.g., using Gaussian filtering algorithms), noise interference can be effectively suppressed, thereby improving data quality. The smoothed incremental capacity curves more clearly reflect the electrochemical characteristics of the battery, making subsequent feature extraction more accurate. In addition, adaptively selecting voltage segment data based on the maximum incremental capacity ensures that the selected data segments cover the key voltage range that is strongly correlated with the battery's health state. This helps the model focus on the most discriminative features, ultimately improving the accuracy of the evaluation results.

[0020] 3. By employing a transfer learning framework, the shared parameters of the pre-trained model (such as the weights of the convolutional neural network) are fixed, and fine-tuning is performed using only a small amount of voltage fragment data from the target lithium battery. This allows the model to quickly adapt to lithium batteries of different models or under different usage conditions. This approach reduces the reliance on a large amount of labeled data for new batteries, thus lowering data collection costs and time. By learning general battery characteristics through the pre-trained model and then fine-tuning it to adapt to specific targets, the model not only maintains the stability of the original knowledge but also flexibly responds to individual differences in batteries, thereby improving generalization performance and making it suitable for diverse application scenarios, such as electric vehicles or energy storage systems.

[0021] 4. By fixing shared parameters and performing local fine-tuning, the high computational overhead of training the model from scratch is avoided, significantly reducing processing time and hardware requirements. At the same time, the adaptive voltage segment selection mechanism narrows the data input range, focusing only on the key voltage window, reducing the amount of data processed by the model and further improving computational efficiency. This design makes the method suitable for embedded systems or real-time monitoring applications, enabling efficient battery health status assessment in resource-constrained environments, thereby reducing overall operating costs.

[0022] 5. Utilizing a Convolutional Neural Network (CNN) as the evaluation model, its convolutional layers can automatically scan and extract deep features from voltage segment data, such as electrochemical patterns related to battery aging, without the need for manual feature design. The combination of pooling layers and fully connected layers enables feature dimensionality scaling and final evaluation output, making the process more automated. Adaptive data selection based on a preset voltage range window further reduces human intervention and ensures that feature extraction matches the characteristics of the battery model. This intelligent design improves the automation level of the evaluation process and is suitable for the deployment of large-scale battery management systems.

[0023] 6. By smoothing to suppress noise and optimizing the pre-trained model to minimize the mean square error, it exhibits strong robustness to data fluctuations and noise, and can stably handle uncertainties in practical applications. In addition, the width of the voltage range window can be preset according to different battery models, making it easy to extend to various lithium battery types. The transfer learning framework allows the model to quickly integrate new data. This scalability enables the technical solution to adapt to the future development and diversified needs of battery technology, and has long-term application value.

[0024] 7. By smoothing incremental capacity curves and adaptively selecting voltage segment data, the accuracy and reliability of the assessment are significantly improved. Simultaneously, the transfer learning framework is used to fix shared parameters and fine-tune them, enhancing the model's adaptability and generalization ability and reducing reliance on large amounts of labeled data. Furthermore, computational efficiency and resource utilization are optimized. Focusing on voltage segments and local fine-tuning reduces processing overhead, and intelligent feature extraction is achieved through convolutional neural networks, improving automation. In addition, its robust design ensures the method's stability under different battery types and noise conditions, exhibiting good scalability. Overall, it achieves an efficient, accurate, and easy-to-deploy battery health monitoring solution. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Figure 1 This is a flowchart of an adaptive battery health status assessment method based on transfer learning according to the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of an adaptive battery health status assessment system based on transfer learning according to the present invention. Detailed Implementation

[0028] The overall approach of the technical solution in this application is as follows: A pre-trained convolutional neural network model is used to solidify the general feature extraction capabilities learned from a large amount of historical data, serving as a high-precision evaluation basis. The incremental capacity curve is then smoothed, and key voltage segments strongly correlated with electrochemical characteristics are adaptively selected as input. Noise interference is suppressed, and the data with the highest signal-to-noise ratio is focused, thereby improving the evaluation accuracy and robustness to individual differences. Finally, by fixing the shared parameters of the model and fine-tuning it using only a small amount of data from the target lithium battery, the model can quickly adapt to different models or batches of lithium batteries, greatly enhancing its cross-domain generalization ability, and thus improving the accuracy, robustness, and generalization ability of battery health status evaluation.

[0029] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the battery health state adaptive evaluation method based on transfer learning according to the present invention includes the following steps: Step S1: Collect charging voltage data of the target lithium battery during the constant current charging stage, calculate and generate an incremental capacity curve based on the charging voltage data, and smooth the incremental capacity curve to suppress noise. Step S2: On the smoothed incremental capacity curve, find the maximum incremental capacity, and take the voltage value corresponding to the maximum incremental capacity as the center reference, and adaptively select voltage segment data from the charging voltage data based on the preset voltage range window. Step S3: Fix the shared parameters used to extract general features in the pre-trained battery health status assessment model, and select a preset number of segment sub-data from the voltage segment data to fine-tune the battery health status assessment model to adapt it to the target lithium battery. Step S4: Assign the voltage segment data to the fine-tuned battery health status assessment model to obtain the battery health status assessment result. The battery health status assessment model automatically extracts and calculates features, directly outputting the battery health status assessment result (predicted SOH) for the current cycle of the target lithium battery. This result can be used by the battery management system for life prediction, fault warning, and equalization management.

[0030] In step S1, the incremental capacity curve is smoothed using a Gaussian filtering algorithm.

[0031] In practice, the charging voltage data of the target lithium battery during the constant current charging (CC) phase can be collected through the battery management system (BMS). The constant current charging phase can eliminate the interference of current changes on the voltage curve, which is beneficial for extracting stable characteristics. The charging voltage data is usually a voltage-time series or a voltage-capacity series.

[0032] Incremental Capacity (IC) Curve Calculation: Based on the charging voltage data, an incremental capacity curve is generated by calculating the capacity change corresponding to a unit voltage change. Specifically, the differential method can be used for calculation: IC = ΔQ / ΔV, where ΔQ is the capacity increment and ΔV is the voltage increment. In actual operation, voltage data can be sampled at equal intervals, and the ratio of the capacity difference to the voltage difference between adjacent sampling points can be calculated.

[0033] Smoothing: Due to measurement and electrochemical noise, the original IC curve often contains many spikes. To suppress noise, a Gaussian filtering algorithm is used for smoothing. In practice, the standard deviation of the Gaussian filter can be set (e.g., σ = 0.5-2mV), and convolution operations can be performed on the IC curve to effectively filter out high-frequency noise while preserving the main peak and valley characteristics of the curve, laying the foundation for subsequent feature point localization. Smoothing significantly improves the robustness of subsequent steps.

[0034] In step S2, the width of the voltage range window is preset according to the incremental capacity curve characteristics of different types of lithium batteries to ensure that the voltage range window can cover the electrochemical characteristic range that is strongly correlated with the battery health status.

[0035] The core of this step is to intelligently select the key data segments from the entire charging data that best reflect the degradation of battery health and have the highest signal-to-noise ratio.

[0036] Positioning the reference center: On the smoothed IC curve, find its maximum incremental capacity value (ICmax). This peak value usually corresponds to the main phase change reaction inside the battery and is sensitive to the aging process. Record the voltage value corresponding to this ICmax value, denoted as V_peak.

[0037] Determining the voltage window: Using V_peak as the central reference, voltage segments are extracted according to a preset voltage range window. The width of the window (e.g., V_peak ± 50mV or V_peak ± 100mV) needs to be preset based on the IC curve characteristics of different lithium battery models. For example, for lithium iron phosphate (LFP) batteries, their IC peaks may be wider, so the window can be appropriately widened; while for ternary lithium (NCM) batteries, their peaks are sharper, so the window can be slightly narrower. The principle of setting the window is to ensure that it covers the core electrochemical characteristic range that is strongly correlated with the battery's state of health (SOH), thereby avoiding the introduction of redundant or interfering information. This adaptive selection mechanism is key to improving the accuracy and generalization ability of this method.

[0038] In step S3, the battery health status assessment model is a convolutional neural network model, and the network structure includes a convolutional layer, a pooling layer and a fully connected layer connected in sequence. The convolutional layer is used to scan and process the input voltage segment data through the convolutional kernel, and superimpose the bias to extract battery state features; the pooling layer is used to scale the dimensions of the battery state features; the fully connected layer is used to output the battery health status assessment result based on the scaled battery state features. The shared parameters include at least the weight parameters of the convolutional layers.

[0039] In step S3, the battery health status assessment model is pre-trained by using a dataset constructed from historical voltage segment data annotated with real SOH as input and minimizing the mean square error between the predicted SOH and the real SOH; the predicted SOH is the battery health status assessment result.

[0040] This step utilizes the knowledge of the pre-trained model and performs rapid adaptation using a small amount of data from the target lithium battery.

[0041] Pre-trained Model: The battery health status assessment model is preferably a convolutional neural network (CNN) model. During the pre-training phase, a large-scale dataset is constructed using historical battery data with real SOH annotations, encompassing various battery models and aging paths. The real SOH is typically calibrated through capacity decay experiments, i.e., SOH = current maximum capacity / rated initial capacity. The model undergoes supervised training by minimizing the mean squared error (MSE) between the predicted and real SOH values, enabling it to learn a universal mapping relationship from voltage segment data to SOH. The model structure includes the following sequentially connected components: Convolutional layers: These layers use multiple convolutional kernels (e.g., 3x1 or 5x1) to scan the input voltage data segments, extracting local and global battery state features. The weights of these convolutional layers are "general feature extractors" learned by the model and are shared parameters.

[0042] Pooling layers (such as max pooling layers): Dimensionally scale the feature maps output by convolutional layers, preserving significant features while reducing the number of parameters and preventing overfitting.

[0043] Fully connected layer: Flatten the pooled feature map and perform nonlinear combination through a multilayer perceptron, finally outputting a scalar value, namely the predicted SOH (battery health state assessment result).

[0044] Transfer learning and fine-tuning: Fixed shared parameters: The shared parameters (at least the weight parameters of the convolutional layers) used to extract common features in the pre-trained CNN model are fixed (frozen) so that they are not updated during fine-tuning. This ensures that the model's core feature extraction capability is not compromised by the small amount of data from the target lithium battery.

[0045] Fine-tuning and adaptation: From the voltage segment data selected for the target lithium battery, a predetermined number of sub-data segments (e.g., data from the first 5-10 charging cycles of the target lithium battery) are randomly selected as the fine-tuning dataset. Only some parameters of the CNN model (such as the weights of fully connected layers) are fine-tuned to minimize the prediction error on the target lithium battery. This process allows the model to quickly "adapt" to the individual characteristics of the target lithium battery, achieving knowledge transfer.

[0046] A preferred embodiment of the battery health state adaptive assessment system based on transfer learning of the present invention includes the following modules: The incremental capacity curve generation module is used to collect charging voltage data of the target lithium battery during the constant current charging stage, calculate and generate an incremental capacity curve based on the charging voltage data, and smooth the incremental capacity curve to suppress noise. The voltage segment adaptive selection module is used to find the maximum incremental capacity on the smoothed incremental capacity curve, and adaptively select voltage segment data from the charging voltage data based on the voltage value corresponding to the maximum incremental capacity as the central reference and a preset voltage range window. The transfer learning module is used to fix the shared parameters used to extract general features in the pre-trained battery health status assessment model, and to fine-tune the battery health status assessment model by selecting a preset number of sub-data segments from the voltage segment data to adapt it to the target lithium battery. The health status assessment module is used to classify the voltage segment data into a fine-tuned battery health status assessment model to obtain the battery health status assessment result. The battery health status assessment model automatically performs feature extraction and calculation, and directly outputs the battery health status assessment result (predicted SOH) for the current cycle of the target lithium battery. This result can be used by the battery management system for life prediction, fault warning, and equalization management.

[0047] In the incremental capacity curve generation module, the incremental capacity curve is smoothed using a Gaussian filtering algorithm.

[0048] In practice, the charging voltage data of the target lithium battery during the constant current charging (CC) phase can be collected through the battery management system (BMS). The constant current charging phase can eliminate the interference of current changes on the voltage curve, which is beneficial for extracting stable characteristics. The charging voltage data is usually a voltage-time series or a voltage-capacity series.

[0049] Incremental Capacity (IC) Curve Calculation: Based on the charging voltage data, an incremental capacity curve is generated by calculating the capacity change corresponding to a unit voltage change. Specifically, the differential method can be used for calculation: IC = ΔQ / ΔV, where ΔQ is the capacity increment and ΔV is the voltage increment. In actual operation, voltage data can be sampled at equal intervals, and the ratio of the capacity difference to the voltage difference between adjacent sampling points can be calculated.

[0050] Smoothing: Due to measurement and electrochemical noise, the original IC curve often contains many spikes. To suppress noise, a Gaussian filtering algorithm is used for smoothing. In practice, the standard deviation of the Gaussian filter can be set (e.g., σ = 0.5-2mV), and convolution operations can be performed on the IC curve to effectively filter out high-frequency noise while preserving the main peak and valley characteristics of the curve, laying the foundation for subsequent feature point localization. Smoothing significantly improves the robustness of subsequent steps.

[0051] In the voltage segment adaptive selection module, the width of the voltage range window is preset according to the incremental capacity curve characteristics of different types of lithium batteries to ensure that the voltage range window can cover the electrochemical characteristic range that is strongly correlated with the battery health status.

[0052] The core of this step is to intelligently select the key data segments from the entire charging data that best reflect the degradation of battery health and have the highest signal-to-noise ratio.

[0053] Positioning the reference center: On the smoothed IC curve, find its maximum incremental capacity value (ICmax). This peak value usually corresponds to the main phase change reaction inside the battery and is sensitive to the aging process. Record the voltage value corresponding to this ICmax value, denoted as V_peak.

[0054] Determining the voltage window: Using V_peak as the central reference, voltage segments are extracted according to a preset voltage range window. The width of the window (e.g., V_peak ± 50mV or V_peak ± 100mV) needs to be preset based on the IC curve characteristics of different lithium battery models. For example, for lithium iron phosphate (LFP) batteries, their IC peaks may be wider, so the window can be appropriately widened; while for ternary lithium (NCM) batteries, their peaks are sharper, so the window can be slightly narrower. The principle of setting the window is to ensure that it covers the core electrochemical characteristic range that is strongly correlated with the battery's state of health (SOH), thereby avoiding the introduction of redundant or interfering information. This adaptive selection mechanism is key to improving the accuracy and generalization ability of this method.

[0055] In the transfer learning module, the battery health status assessment model is a convolutional neural network model, and the network structure includes a convolutional layer, a pooling layer and a fully connected layer connected in sequence. The convolutional layer is used to scan and process the input voltage segment data through the convolutional kernel, and superimpose the bias to extract battery state features; the pooling layer is used to scale the dimensions of the battery state features; the fully connected layer is used to output the battery health status assessment result based on the scaled battery state features. The shared parameters include at least the weight parameters of the convolutional layers.

[0056] In the transfer learning module, the battery health status assessment model is pre-trained by using a dataset constructed from historical voltage segment data annotated with real SOH as input and minimizing the mean square error between the predicted SOH and the real SOH; the predicted SOH is the battery health status assessment result.

[0057] This step utilizes the knowledge of the pre-trained model and performs rapid adaptation using a small amount of data from the target lithium battery.

[0058] Pre-trained Model: The battery health status assessment model is preferably a convolutional neural network (CNN) model. During the pre-training phase, a large-scale dataset is constructed using historical battery data with real SOH annotations, encompassing various battery models and aging paths. The real SOH is typically calibrated through capacity decay experiments, i.e., SOH = current maximum capacity / rated initial capacity. The model undergoes supervised training by minimizing the mean squared error (MSE) between the predicted and real SOH values, enabling it to learn a universal mapping relationship from voltage segment data to SOH. The model structure includes the following sequentially connected components: Convolutional layers: These layers use multiple convolutional kernels (e.g., 3x1 or 5x1) to scan the input voltage data segments, extracting local and global battery state features. The weights of these convolutional layers are "general feature extractors" learned by the model and are shared parameters.

[0059] Pooling layers (such as max pooling layers): Dimensionally scale the feature maps output by convolutional layers, preserving significant features while reducing the number of parameters and preventing overfitting.

[0060] Fully connected layer: Flatten the pooled feature map and perform nonlinear combination through a multilayer perceptron, finally outputting a scalar value, namely the predicted SOH (battery health state assessment result).

[0061] Transfer learning and fine-tuning: Fixed shared parameters: The shared parameters (at least the weight parameters of the convolutional layers) used to extract common features in the pre-trained CNN model are fixed (frozen) so that they are not updated during fine-tuning. This ensures that the model's core feature extraction capability is not compromised by the small amount of data from the target lithium battery.

[0062] Fine-tuning and adaptation: From the voltage segment data selected for the target lithium battery, a predetermined number of sub-data segments (e.g., data from the first 5-10 charging cycles of the target lithium battery) are randomly selected as the fine-tuning dataset. Only some parameters of the CNN model (such as the weights of fully connected layers) are fine-tuned to minimize the prediction error on the target lithium battery. This process allows the model to quickly "adapt" to the individual characteristics of the target lithium battery, achieving knowledge transfer.

[0063] In summary, the advantages of this invention are: 1. By collecting charging voltage data of the target lithium battery during the constant current charging stage, an incremental capacity curve is calculated and generated based on the charging voltage data. The incremental capacity curve is smoothed to suppress noise. On the smoothed incremental capacity curve, the maximum incremental capacity is found, and the voltage value corresponding to the maximum incremental capacity is used as the central benchmark. Voltage segment data is adaptively selected from the charging voltage data based on a preset voltage range window. Then, the shared parameters used to extract general features in the pre-trained battery health status assessment model are fixed. A preset number of sub-data segments are selected from the voltage segment data to fine-tune the battery health status assessment model to adapt it to the target lithium battery. Finally, the voltage segment data belongs to the fine-tuned battery health status assessment model. The model is evaluated to obtain battery health status assessment results. Specifically, a pre-trained convolutional neural network model is used to solidify the general feature extraction capabilities learned from a large amount of historical data, serving as a high-precision assessment basis. The incremental capacity curve is then smoothed, and key voltage segments strongly correlated with electrochemical characteristics are adaptively selected as input, effectively suppressing noise interference and focusing on data with the highest signal-to-noise ratio, thereby improving the accuracy of the assessment and its robustness to individual differences. Finally, by fixing the shared parameters of the model and fine-tuning it using only a small amount of data from the target lithium battery, the model can quickly adapt to different models or batches of lithium batteries, greatly enhancing its cross-domain generalization ability. Ultimately, this synergistically improves the accuracy, robustness, and generalization ability of battery health status assessment.

[0064] 2. By collecting charging voltage data during the constant current charging phase and generating incremental capacity curves, and then smoothing them (e.g., using Gaussian filtering algorithms), noise interference can be effectively suppressed, thereby improving data quality. The smoothed incremental capacity curves more clearly reflect the electrochemical characteristics of the battery, making subsequent feature extraction more accurate. In addition, adaptively selecting voltage segment data based on the maximum incremental capacity ensures that the selected data segments cover the key voltage range that is strongly correlated with the battery's health state. This helps the model focus on the most discriminative features, ultimately improving the accuracy of the evaluation results.

[0065] 3. By employing a transfer learning framework, the shared parameters of the pre-trained model (such as the weights of the convolutional neural network) are fixed, and fine-tuning is performed using only a small amount of voltage fragment data from the target lithium battery. This allows the model to quickly adapt to lithium batteries of different models or under different usage conditions. This approach reduces the reliance on a large amount of labeled data for new batteries, thus lowering data collection costs and time. By learning general battery characteristics through the pre-trained model and then fine-tuning it to adapt to specific targets, the model not only maintains the stability of the original knowledge but also flexibly responds to individual differences in batteries, thereby improving generalization performance and making it suitable for diverse application scenarios, such as electric vehicles or energy storage systems.

[0066] 4. By fixing shared parameters and performing local fine-tuning, the high computational overhead of training the model from scratch is avoided, significantly reducing processing time and hardware requirements. At the same time, the adaptive voltage segment selection mechanism narrows the data input range, focusing only on the key voltage window, reducing the amount of data processed by the model and further improving computational efficiency. This design makes the method suitable for embedded systems or real-time monitoring applications, enabling efficient battery health status assessment in resource-constrained environments, thereby reducing overall operating costs.

[0067] 5. Utilizing a Convolutional Neural Network (CNN) as the evaluation model, its convolutional layers can automatically scan and extract deep features from voltage segment data, such as electrochemical patterns related to battery aging, without the need for manual feature design. The combination of pooling layers and fully connected layers enables feature dimensionality scaling and final evaluation output, making the process more automated. Adaptive data selection based on a preset voltage range window further reduces human intervention and ensures that feature extraction matches the characteristics of the battery model. This intelligent design improves the automation level of the evaluation process and is suitable for the deployment of large-scale battery management systems.

[0068] 6. By smoothing to suppress noise and optimizing the pre-trained model to minimize the mean square error, it exhibits strong robustness to data fluctuations and noise, and can stably handle uncertainties in practical applications. In addition, the width of the voltage range window can be preset according to different battery models, making it easy to extend to various lithium battery types. The transfer learning framework allows the model to quickly integrate new data. This scalability enables the technical solution to adapt to the future development and diversified needs of battery technology, and has long-term application value.

[0069] 7. By smoothing incremental capacity curves and adaptively selecting voltage segment data, the accuracy and reliability of the assessment are significantly improved. Simultaneously, the transfer learning framework is used to fix shared parameters and fine-tune them, enhancing the model's adaptability and generalization ability and reducing reliance on large amounts of labeled data. Furthermore, computational efficiency and resource utilization are optimized. Focusing on voltage segments and local fine-tuning reduces processing overhead, and intelligent feature extraction is achieved through convolutional neural networks, improving automation. In addition, its robust design ensures the method's stability under different battery types and noise conditions, exhibiting good scalability. Overall, it achieves an efficient, accurate, and easy-to-deploy battery health monitoring solution.

[0070] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An adaptive battery health status assessment method based on transfer learning, characterized in that: Includes the following steps: Step S1: Collect charging voltage data of the target lithium battery during the constant current charging stage, calculate and generate an incremental capacity curve based on the charging voltage data, and smooth the incremental capacity curve to suppress noise. Step S2: On the smoothed incremental capacity curve, find the maximum incremental capacity, and take the voltage value corresponding to the maximum incremental capacity as the center reference, and adaptively select voltage segment data from the charging voltage data based on the preset voltage range window. Step S3: Fix the shared parameters used to extract general features in the pre-trained battery health status assessment model, and select a preset number of segment sub-data from the voltage segment data to fine-tune the battery health status assessment model to adapt it to the target lithium battery. Step S4: Assign the voltage segment data to the fine-tuned battery health status assessment model to obtain the battery health status assessment result.

2. The adaptive battery health status assessment method based on transfer learning as described in claim 1, characterized in that: In step S1, the incremental capacity curve is smoothed using a Gaussian filtering algorithm.

3. The adaptive battery health status assessment method based on transfer learning as described in claim 1, characterized in that: In step S2, the width of the voltage range window is preset according to the incremental capacity curve characteristics of different types of lithium batteries to ensure that the voltage range window can cover the electrochemical characteristic range that is strongly correlated with the battery health status.

4. The adaptive battery health status assessment method based on transfer learning as described in claim 1, characterized in that: In step S3, the battery health status assessment model is a convolutional neural network model, and the network structure includes a convolutional layer, a pooling layer and a fully connected layer connected in sequence. The convolutional layer is used to scan and process the input voltage segment data through the convolutional kernel, and superimpose the bias to extract battery state features; the pooling layer is used to scale the dimensions of the battery state features; the fully connected layer is used to output the battery health status assessment result based on the scaled battery state features. The shared parameters include at least the weight parameters of the convolutional layers.

5. The adaptive battery health status assessment method based on transfer learning as described in claim 1, characterized in that: In step S3, the battery health status assessment model is pre-trained by using a dataset constructed from historical voltage segment data annotated with real SOH as input and minimizing the mean square error between the predicted SOH and the real SOH; the predicted SOH is the battery health status assessment result.

6. A battery health state adaptive assessment system based on transfer learning, characterized in that: Includes the following modules: The incremental capacity curve generation module is used to collect charging voltage data of the target lithium battery during the constant current charging stage, calculate and generate an incremental capacity curve based on the charging voltage data, and smooth the incremental capacity curve to suppress noise. The voltage segment adaptive selection module is used to find the maximum incremental capacity on the smoothed incremental capacity curve, and adaptively select voltage segment data from the charging voltage data based on the voltage value corresponding to the maximum incremental capacity as the central reference and a preset voltage range window. The transfer learning module is used to fix the shared parameters used to extract general features in the pre-trained battery health status assessment model, and to fine-tune the battery health status assessment model by selecting a preset number of sub-data segments from the voltage segment data to adapt it to the target lithium battery. The health status assessment module is used to classify the voltage segment data into a fine-tuned battery health status assessment model to obtain the battery health status assessment result.

7. The battery health state adaptive assessment system based on transfer learning as described in claim 6, characterized in that: In the incremental capacity curve generation module, the incremental capacity curve is smoothed using a Gaussian filtering algorithm.

8. The battery health state adaptive assessment system based on transfer learning as described in claim 6, characterized in that: In the voltage segment adaptive selection module, the width of the voltage range window is preset according to the incremental capacity curve characteristics of different types of lithium batteries to ensure that the voltage range window can cover the electrochemical characteristic range that is strongly correlated with the battery health status.

9. The battery health state adaptive assessment system based on transfer learning as described in claim 6, characterized in that: In the transfer learning module, the battery health status assessment model is a convolutional neural network model, and the network structure includes a convolutional layer, a pooling layer and a fully connected layer connected in sequence. The convolutional layer is used to scan and process the input voltage segment data through the convolutional kernel, and superimpose the bias to extract battery state features; the pooling layer is used to scale the dimensions of the battery state features; the fully connected layer is used to output the battery health status assessment result based on the scaled battery state features. The shared parameters include at least the weight parameters of the convolutional layers.

10. The battery health state adaptive assessment system based on transfer learning as described in claim 6, characterized in that: In the transfer learning module, the battery health status assessment model is pre-trained by using a dataset constructed from historical voltage segment data annotated with real SOH as input and minimizing the mean square error between the predicted SOH and the real SOH; the predicted SOH is the battery health status assessment result.

Images