Lithium battery health state estimation method and system based on resting stage characteristics for underwater vehicle

Abstract

The invention discloses a lithium battery health state estimation method and system based on resting stage characteristics for an underwater vehicle, and the method comprises the steps: firstly recognizing a resting stage of the underwater vehicle, and collecting terminal voltage and time sequence data to construct a resting characteristic vector; constructing an initial health state SOH sequence, and performing anomaly detection and correction; inputting the corrected sequence into an adaptive Kalman filter, and dynamically adjusting and measuring noise covariance to obtain a health state reference sequence; then taking the resting feature vector as an input and the health state reference sequence as an output label, constructing a plurality of health state prediction models, and carrying out training and parameter optimization; and during prediction, inputting the new resting feature vector into each health state prediction model to obtain a plurality of health state prediction results, carrying out weighted fusion, and outputting a final health state estimation value of the lithium battery of the underwater vehicle. The method is high in applicability, and the lithium battery health state estimation precision and stability under the low-frequency sampling condition can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of energy management and lithium battery status monitoring technology for underwater vehicles, specifically to a method and system for estimating the health status of lithium batteries based on resting phase characteristics for underwater vehicles. Background Technology

[0002] With the rapid development of deep-sea exploration, seabed resource development, and underwater intelligent systems, underwater vehicles, as important marine operation platforms, are widely used in scientific research, military reconnaissance, and engineering operations. Lithium-ion batteries, with their high energy density, long cycle life, and excellent discharge performance, have become the primary power source for underwater vehicles. However, the underwater environment is characterized by high pressure, low temperature, high humidity, and difficult maintenance. In practical missions, limitations such as communication bandwidth, strict energy consumption budgets, and low data sampling frequencies often exist, making traditional battery state-of-health (SOH) estimation methods, which rely on high-frequency data or complex models, difficult to directly apply to underwater applications.

[0003] Existing State of Health (SOH) estimation techniques include those based on equivalent circuit models, high-frequency dynamic data feature analysis, and machine learning methods driven by large amounts of sensor data. These methods generally rely on high-resolution voltage, current, or impedance measurements, requiring continuous data uploads or high-frequency electrochemical parameter acquisition, and are relatively mature in land-based equipment and new energy vehicles. However, in underwater vehicles, due to long mission cycles, limited energy consumption, and short communication windows, complete operational information is often unavailable. Furthermore, noise in the seawater environment, system switching jumps, and sampling sparsity can cause significant fluctuations in the health status curve, further reducing the estimation stability and reliability of traditional methods. In addition, underwater platforms generally lack the hardware configuration for online measurement of internal resistance or frequency characteristics, making it impossible to deploy many models that rely on such features.

[0004] Therefore, how to accurately, stably, and embeddably estimate the health status of lithium batteries using only resting phase data that naturally occurs during the mission, under conditions of low-frequency sampling, strong noise environment, and limited communication, has become a key technical problem that urgently needs to be solved in the field of underwater energy management. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention aims to provide a lithium battery health status estimation method and system based on resting phase characteristics for underwater vehicles. By utilizing the voltage change information that naturally occurs during the resting phase of the underwater vehicle's mission, and combining anomaly detection, adaptive filtering, and multi-model fusion strategies, a high-precision estimation of the lithium battery health status is achieved. This allows the model to obtain stable and accurate health status prediction results even under low-frequency data conditions, and maintains the continuity and robustness of the output when noise disturbances, data jumps, or sampling anomalies occur. Thus, the health status assessment remains reliable in long-term missions and restricted communication environments, providing effective protection for the safe operation and mission planning of underwater vehicles.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for estimating the health status of lithium batteries based on resting phase characteristics for underwater vehicles includes the following steps: Step 1: After the underwater vehicle completes the discharge process, the resting phase is automatically identified based on the condition that the current drops to zero and remains stable; during the resting phase, terminal voltage and time series data are collected, and a resting feature vector reflecting the battery's resting recovery behavior is constructed based on the data; Step 2: Construct an initial health state SOH sequence based on historical capacity data or calibration data; perform anomaly detection on the initial health state SOH sequence, and identify any point that deviates from a preset threshold as an anomaly; perform substitution correction on the detected anomalies to improve the continuity and effectiveness of the initial health state SOH sequence; Step 3: Input the corrected SOH sequence from Step 2 into an adaptive Kalman filter. Dynamically adjust the measurement noise covariance based on the magnitude or trend of the filter residual. This increases the filter gain to enhance correction when the noise is high and decreases the filter gain to maintain smoothness when the sequence is stable. Finally, a continuous, noise-suppressing, and noise-suppressing healthy state reference sequence is obtained, which can be used as a training target for the model. Step 4: Using the resting feature vector from Step 1 as input and the health status reference sequence from Step 3 as output label, construct multiple complementary health status prediction models and train, optimize parameters, and validate them respectively; during actual prediction, input the new resting feature vector into each health status prediction model to obtain multiple independent health status prediction results. Step 5: Perform weighted fusion on the multiple health status prediction results obtained in Step 4, search for the minimum residual combination by weight enumeration, and output the final health status estimate of the underwater vehicle's lithium battery.

[0007] Step 1: After data identification during the resting phase is completed, a resting feature vector is constructed to characterize the voltage recovery behavior by statistically calculating the starting and ending voltages of the resting phase. The resting feature vector is obtained by averaging several sampling points at the start of the resting phase as the resting starting voltage. U start The average value of several sampling points at the end of the resting period is obtained as the resting termination voltage. U end The voltage recovery amplitude ΔU is obtained based on the difference between the two values; then the voltage recovery rate is calculated based on the duration of the resting phase. re_rate This is used to reflect the voltage recovery characteristics per unit time; simultaneously, the average of the resting start voltage and the resting end voltage is used as the resting average voltage. U mean This forms a resting feature vector used to predict battery degradation trends.

[0008] Step 2 involves anomaly detection of the initial health state sequence. The variation amplitude of the health state difference between adjacent cycles is analyzed to identify abrupt changes caused by instrument fluctuations, communication noise, or instantaneous disturbances. When the difference value exceeds a preset threshold, the health state corresponding to that cycle is determined to be an anomaly. The average of the health states on both sides is used to correct the anomaly value, thereby obtaining a continuous and smooth health state sequence.

[0009] After anomaly correction is completed, step 3 applies an adaptive Kalman filter to the health state sequence. By constructing a state update model and a measurement model, the observed noise of the health state is dynamically estimated, and the filter gain is adjusted in real time according to the noise changes. This ensures that the filtering results maintain the consistency of the decay trend while effectively suppressing local fluctuations, thereby obtaining a smooth health state reference curve as the training target of the model. The adaptive noise update corrects the measurement noise covariance based on the amplitude and changes of the residual at the previous time step, realizing the adaptive adjustment of the noise level by the adaptive Kalman filter at different aging stages.

[0010] In step 4, the prediction stage of the multi-health state prediction model trains multiple structurally complementary regression models using resting feature vectors to capture the nonlinear laws of health state changes with cycles. The models used include a random forest regression model based on random subsampling and multi-tree averaging, an extreme gradient boosting model based on second-order gradient optimization and regularization, and a gradient boosting regression model based on stepwise residual fitting. These three models provide complementary prediction information in terms of stability, fine-fitting ability, and ability to characterize complex decay patterns, respectively. In the prediction stage, the same resting feature vector is input into the above models to obtain multiple health state prediction results, which are then used as inputs for subsequent fusion.

[0011] After obtaining multiple health status predictions from the multi-model prediction, step 5 constructs a weighted least squares regression fusion model to weight and combine the prediction results of different health status prediction models. The fusion weights are obtained by searching for each weight combination on the validation set to minimize the sum of the absolute values ​​of the residuals, thereby achieving a comprehensive balance of performance for the random forest regression model based on random subsampling and multi-tree averaging mechanism, the extreme gradient boosting model based on second-order gradient optimization and regularization structure, and the gradient boosting regression model based on stepwise residual fitting. The weights satisfy non-negativity and normalization constraints, so that the final prediction results have higher stability and accuracy in different aging stages.

[0012] A lithium battery health state estimation system based on resting phase characteristics for underwater vehicles, implementing the aforementioned lithium battery health state estimation method based on resting phase characteristics for underwater vehicles, includes: The feature extraction module is used to acquire data on the resting phase of the lithium battery of the underwater vehicle after discharge, and to construct a resting feature vector to characterize the health status of the battery based on the data. The anomaly detection and filtering module is used to perform differential anomaly detection on the healthy state sequence and correct anomaly points, and to perform adaptive filtering on the corrected sequence to obtain a smooth healthy state reference curve. The sub-model prediction module is used to call multiple constructed regression prediction models based on the resting feature vector as health status prediction models to obtain multiple health status prediction results. The fusion estimation module is used to perform weighted fusion of multiple health status prediction results to obtain the final health status estimate of the lithium battery.

[0013] The lithium battery health state estimation system comprises a feature extraction module, an anomaly detection and filtering module, a sub-model prediction module, and a fusion estimation module connected in sequence. The feature extraction module identifies the resting phase after the battery has completed discharge and acquires voltage, current, and time data. It constructs a resting feature vector characterizing the battery's polarization relaxation dynamics and long-term degradation trend by calculating the resting phase characteristics such as the resting period voltage recovery amplitude, resting period voltage recovery rate, resting period average voltage, and cycle number. The anomaly detection and filtering module detects jump anomalies in the SOH sequence based on the first-order difference method and corrects them using the mean of adjacent points. Furthermore, it utilizes adaptive Kalman filtering to suppress noise and dynamically smooth the corrected SOH sequence to obtain a continuous and consistent target output. The sub-model prediction module constructs and trains a random forest regression model, an extreme gradient boosting regression model, and a gradient boosting regression model respectively to predict the health state of the input resting feature vector from multiple angles and through multiple mechanisms. The fusion estimation module performs weighted fusion of the prediction results from each sub-model and outputs the final SOH estimate, achieving multi-model complementarity and prediction result optimization.

[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements all the steps of the described method for estimating the state of health of lithium batteries based on resting phase characteristics for underwater vehicles.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention fully utilizes data from the post-discharge resting phase during underwater vehicle missions, without relying on high-frequency operating conditions, electrochemical impedance spectroscopy, or additional sensor measurements. Therefore, it can operate stably in underwater environments with low communication bandwidth, low sampling frequency, limited energy consumption, and significant noise interference. By employing differential anomaly detection and neighborhood correction, it effectively eliminates anomalous jumps in the health state sequence and dynamically suppresses sequence noise using adaptive Kalman filtering, resulting in a more continuous, smoother health state reference curve that conforms to the true degradation trend, thus significantly improving data quality. By constructing multiple complementary regression models and applying a weighted fusion strategy based on validation set residual optimization to their outputs, the models exhibit better fitting ability, robustness, and generalization performance at different degradation stages, achieving superior health state estimation accuracy and stability compared to a single model. The overall method of this invention has low computational cost and a clear structure, allowing deployment in embedded battery management systems to achieve online health assessment and lifespan management of underwater vehicles. This improves the reliability of mission planning, energy scheduling, and fault early warning, demonstrating significant engineering application value and promotional significance. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the resting phase features extracted in the embodiments of the present invention, wherein (a) is a graph showing the trend of voltage change with the cycle; (b) is a graph showing the trend of voltage recovery rate with the cycle; and (c) is a graph showing the trend of voltage mean with the cycle.

[0018] Figure 3 This is a schematic diagram of the results of anomaly detection and correction and adaptive Kalman filtering in an embodiment of the present invention, wherein (a) is a comparison diagram of the SOH curves of battery pack 1 before and after anomaly detection and Kalman filtering; and (b) is a comparison diagram of the SOH curves of battery pack 2 before and after anomaly detection and Kalman filtering.

[0019] Figure 4 The above is a graph showing the SOH prediction results of one set of battery data in an embodiment of the present invention. (a) is a comparison curve and a magnified view of the SOH predicted by the four models and the actual SOH; (b) is a graph of the prediction error of the four models; (c) is a box plot of the prediction residuals of the four models; and (d) is a scatter plot of the predicted value and the actual value of the method in the present invention.

[0020] Figure 5 This is a structural block diagram of the system according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0022] like Figure 1 As shown, a method for estimating the health status of lithium batteries based on resting phase characteristics for underwater vehicles includes the following steps: Step 1: This method first identifies the resting phase of the lithium battery after the discharge process ends. Underwater vehicles usually have multiple natural resting phases during the mission. When the current drops to zero and remains stable, it can be determined as the resting phase. After identifying the resting phase, the terminal voltage data at the start and end of the resting phase are collected to construct a resting feature vector that reflects the battery's resting recovery behavior. In this embodiment, during the construction of the resting feature vector, the terminal voltage curve after the discharge ends is sampled, and the start and end times of the resting phase are used as reference positions for feature extraction. The resting start voltage can be obtained by averaging the voltage within a short time window after the start of resting, and the resting end voltage can be obtained by averaging the voltage data before the end of resting. To characterize the voltage recovery behavior during the resting phase, the resting voltage recovery amplitude is obtained by calculating the difference between the resting end voltage and the resting start voltage. Simultaneously, the recovery amplitude is normalized according to the duration of the resting phase to obtain the resting voltage recovery rate, which is used to characterize the speed of the polarization relaxation process. Furthermore, the resting average voltage is obtained by averaging the average voltages at the start and end of resting to reflect the overall voltage level of the battery in a resting state. These multiple features together constitute the resting feature vector, which serves as the input to the health state prediction model, including: in U start The voltage at the start of rest. U end The voltage at the end of the resting period. t rest This is the rest period.

[0023] The resting voltage recovery amplitude characterizes the battery's recovery capability during the polarization relaxation process after discharge, the resting voltage recovery rate reflects the speed of the recovery process, and the resting average voltage reflects the overall voltage level of the battery during the resting phase. These features are collectively used to characterize the battery's degradation trend. The features extracted in this embodiment are as follows: Figure 2 As shown, all selected features exhibit a clear evolution trend during battery aging, demonstrating strong observability and high correlation with battery degradation. They can provide sufficient and valuable degradation information at different aging stages, providing a reliable basis for subsequent SOH modeling. Step 2: Construct an initial healthy state SOH sequence based on historical capacity data or calibration data; To improve the stability of the healthy state sequence, anomaly detection and correction processing is implemented on the initial healthy state sequence. By analyzing the difference in healthy state values ​​between two adjacent periods, it is determined whether there are abrupt changes or sharp peaks. After anomalies are detected, the mean or neighborhood smoothing strategy of adjacent healthy states is used to correct the anomalies. To improve the stability of model training and prediction, it is necessary to identify and correct any abnormal changes that may exist in the initial health state sequence. This embodiment adopts an anomaly detection method based on health state difference, which determines whether the difference between the health states of adjacent cycles exceeds a preset threshold. When the difference exceeds the threshold, the current health state can be identified as an anomaly. For a single anomaly, the average value of the preceding and following health states is used to replace it. When anomaly regions appear consecutively, piecewise linear interpolation can be performed based on the normal health state values ​​at the boundaries of the anomaly intervals to achieve continuous correction of the health state sequence. This method can effectively remove local abrupt changes caused by measurement jumps, data gaps, or communication jitter, making the corrected health state sequence more consistent with the physical laws of natural battery degradation, including: in: SOH n For the first n The SOH value of the next cycle. SOH n-1 and SOH n+1 These are the SOH values ​​for the previous and next cycles, respectively. δ A threshold set for experience; Step 3: Input the corrected SOH sequence from Step 2 into an adaptive Kalman filter. Dynamically adjust the measurement noise covariance based on the magnitude or trend of the filter residual. This increases the filter gain to enhance correction when the noise is high and decreases the filter gain to maintain smoothness when the sequence is stable. Finally, a continuous, noise-suppressing, and noise-suppressing healthy state reference sequence is obtained, which can be used as a training target for the model. After anomaly detection and correction, an adaptive Kalman filter is applied to further improve the smoothness and predictability of the health state sequence. The adaptive Kalman filter automatically adjusts the measurement noise parameters based on the residual between the observed health state values ​​and the filter's predicted values. This increases the filter's sensitivity to observations when they fluctuate significantly, while reducing its response to noise when observations are stable. This achieves dynamic noise suppression at different aging stages, including: in: express The predicted state value at time; This is the updated estimate; , These are the state covariances for the prediction and update phases, respectively. F This is the state transition matrix; H The observation matrix; QFor process noise covariance; K k Kalman gain; z k These are the observed values; I The identity matrix is ​​used; the filtered health state sequence is more stable and continuous, and can serve as a reliable target for multi-model training and validation, thereby improving the overall prediction performance. By constructing a state update model and a measurement model, the observed noise of the health state is dynamically estimated, and the filter gain is adjusted in real time according to the noise changes. This ensures that the filtering results can maintain the consistency of the decay trend and effectively suppress local fluctuations, thereby obtaining a smooth health state reference curve as the training target of the model. The adaptive noise update corrects the measurement noise covariance based on the amplitude and changes of the residual at the previous time step, realizing the adaptive adjustment of the noise level by the adaptive Kalman filter at different aging stages. The results of anomaly detection, correction, and adaptive Kalman filtering in this embodiment are as follows: Figure 3 As shown, after outlier detection and repair, and adaptive Kalman filtering, the SOH curves of each battery sample are significantly smoother and more continuous, and the abrupt changes and abnormal fluctuations in the original data are effectively suppressed. Under different battery samples and rate conditions, the evolution trend of the filtered SOH remains consistent and conforms to the actual aging law, significantly reducing the impact of noise and operating condition disturbances. Local magnification results demonstrate that this method can maintain the continuity and stability of SOH changes during key degradation stages, providing a reliable data foundation for subsequent SOH modeling and prediction. Step 4: Using the resting feature vector from Step 1 as input and the health status reference sequence from Step 3 as output label, construct multiple complementary health status prediction models and train, optimize parameters, and validate them respectively; during actual prediction, input the new resting feature vector into each health status prediction model to obtain multiple independent health status prediction results. In the prediction phase of the multi-health state prediction model, multiple structurally complementary regression models are trained using resting feature vectors to capture the nonlinear patterns of health state changes over time. The models employed include a random forest regression model based on random subsampling and multi-tree averaging, an extreme gradient boosting model based on second-order gradient optimization and regularization, and a gradient boosting regression model based on stepwise residual fitting. These three models provide complementary predictive information in terms of stability, fine-fitting ability, and ability to characterize complex decay patterns, respectively. During the prediction phase, the same resting feature vector is input into these models to obtain multiple health state prediction results, which are then used as input for subsequent fusion. This embodiment uses random forest, extreme gradient boosting model and gradient boosting regression model as base learners. Since the above models are complementary in terms of training method, structure and fitting ability, they can learn the relationship between resting phase features and health status from different perspectives; multiple prediction outputs can be obtained by inputting the same feature vector into each base learner respectively. Step 5: Perform weighted fusion on the multiple health status prediction results obtained in Step 4, search for the minimum residual combination by weight enumeration, and output the final health status estimate of the underwater vehicle's lithium battery.

[0024] To improve the overall stability and accuracy of the prediction results, after obtaining multiple health status predictions from multiple models, a weighted least squares regression fusion model is constructed to weight and combine the prediction results of different health status prediction models. The fusion weights are obtained by searching for each weight combination on the validation set to minimize the sum of the absolute values ​​of the residuals, thereby achieving a comprehensive balance of performance for the random forest regression model based on random subsampling and multi-tree averaging, the extreme gradient boosting model based on second-order gradient optimization and regularization structure, and the gradient boosting regression model based on stepwise residual fitting. The weights satisfy non-negativity and normalization constraints, ensuring higher stability and accuracy of the final prediction results at different aging stages. The fusion strategy can fully leverage the advantages of each model at different decay stages, reducing the potential error accumulation or performance fluctuations of a single model, and improving the overall consistency and accuracy of the prediction results. The prediction results and errors in this embodiment are as follows: Figure 4 As shown.

[0025] Before performing health status prediction, the obtained resting feature vectors are standardized to avoid the impact of dimensional differences between features of different dimensions on model training. Standardization can be achieved by calculating the mean and standard deviation of the features in the training set, and scaling each feature dimension by zeroing the mean and scaling it by the standard deviation. Standardized features are kept on the same scale, and the same transformation parameters are used in both model training and prediction phases to ensure the stability of the training process.

[0026] During the model training phase, multiple regression models with complementary structures are trained based on the features of the resting phase, including a random forest model, an extreme gradient boosting model, and a gradient boosting regression model. In this embodiment, the random forest model improves robustness to noise by constructing multiple decision trees and randomly sampling features and samples; the extreme gradient boosting model utilizes second-order gradient information and regularization mechanisms to suppress overfitting while maintaining high fitting ability; and the gradient boosting regression model improves the model's fine-fit ability by progressively approximating the residuals, and can better capture the changing trends of health status in different decline stages. The above three models differ in structure, training strategy, and fitting ability, thus providing complementary prediction results when facing complex degradation curves.

[0027] In the prediction phase, for any resting feature vector to be predicted, the feature is input in parallel into each trained base learner to obtain multiple health state prediction values. For example... Figure 4 As shown, the output curves of the multi-model prediction results may exhibit certain differences in some areas, reflecting the model's fitting preference and generalization characteristics for different decay stages. To obtain a stable and accurate final prediction result, this embodiment performs weighted fusion of the multiple prediction values. By pre-evaluating the prediction errors corresponding to different weight combinations on the validation set, the weight combination with the smallest error is selected as the final fusion weight. During the fusion process, the weights of each base learner remain non-negative and satisfy the condition that the sum of the weights is 1, ensuring that the fusion result is within a reasonable range, including: in , , The predicted outputs of the three models are shown below. w 1, w 2, w 3 represents the fusion weights of the three base learners.

[0028] The optimal weight combination in the fusion process is solved by the following optimization problem: in: y i For genuine SOH labels, For the prediction results of each base model, w j For the first j Weights of each model Based on the fused prediction results, the prediction error can be calculated to evaluate the overall performance of the model. The prediction error can be obtained by calculating the difference between the predicted value and the health status reference value. The concentration of the error distribution can reflect the stability of the model, while the fit between the error curve and the actual health status curve can reflect the model's predictive ability for different aging stages. The weighted fusion strategy described in this embodiment can effectively reduce the bias of the single model in certain intervals, making the fused prediction curve smoother and the error distribution more compact.

[0029] Through the above training and prediction process, this embodiment achieves high-precision prediction of the health status of lithium batteries under low-frequency sampling and noise disturbance conditions, and further improves the overall robustness and stability of the prediction through multi-model fusion.

[0030] like Figure 5 As shown, a lithium battery health status estimation system based on resting phase characteristics for underwater vehicles is implemented. The system includes a feature extraction module, an anomaly detection and filtering module, a sub-model prediction module, and a fusion estimation module. These modules are interconnected and work together to achieve a stable estimation of the lithium battery health status of the underwater vehicle. The feature extraction module is used to automatically extract key features from the resting phase data after the underwater vehicle has finished discharging, including resting start voltage, resting end voltage, resting voltage recovery amplitude, resting voltage recovery rate, and resting average voltage. This module constructs a feature vector that can characterize the battery degradation state from the natural polarization relaxation behavior of the resting phase by reading basic measurements such as terminal voltage and resting time that depend on low-frequency sampling, and provides input data for subsequent prediction tasks. The anomaly detection and filtering module is used to identify and correct abnormal data points in the health status reference curve, and perform adaptive Kalman filtering on the corrected sequence. This module first detects abrupt changes based on the changes in health status differences, and corrects the abnormal points using neighborhood mean or piecewise linear interpolation. Then, it dynamically adjusts the noise parameters through adaptive Kalman filtering to obtain a smooth and continuous health status reference curve, which serves as the training target for the sub-model. The sub-model prediction module is used to construct multiple regression models based on resting feature vectors and perform health status predictions respectively. In this embodiment, the sub-model prediction module includes three sub-models: a random forest model, an extreme gradient boosting model, and a gradient boosting regression model. Each sub-model is independently trained based on standardized feature vectors and outputs the corresponding health status estimate during the prediction stage. Multiple sub-models adopt different structures and training strategies, thereby possessing complementary fitting capabilities and generalization performance. The fusion estimation module receives the prediction results from each sub-model and performs weighted fusion based on preset weights to obtain the final health status estimate. By evaluating the prediction performance of each model on the validation set, the module determines the optimal weight combination as the fusion rule, resulting in a more stable and accurate final output health status prediction curve. This module can also generate the final prediction error result for model performance analysis and system validation.

[0031] Through the above structural design, the system described in this embodiment can achieve highly accurate estimation of the health status of lithium batteries and strong environmental robustness under typical operating conditions of underwater vehicles, including low-frequency sampling, limited communication, and noise interference. This embodiment describes the system structure; those skilled in the art can replace or extend the implementation of modules based on actual deployment requirements without affecting the overall technical solution of this invention.

[0032] A computer-readable storage medium, wherein the computer-readable storage medium may be a non-transitory storage medium, such as a read-only memory (ROM), a disk storage device, an optical disk storage device, a flash memory, a solid-state drive, or other data storage device capable of storing computer program code.

[0033] The computer-readable storage medium stores a computer program. When the computer program is called and executed by the processor, it implements all the steps of the lithium battery health state estimation method based on resting phase characteristics as described in any embodiment of the present invention. During execution, the processor sequentially completes the identification of resting phase data, feature extraction, anomaly detection and correction of health state sequence, adaptive Kalman filter smoothing, multi-model prediction, and fusion estimation by running the program, thereby outputting the lithium battery health state estimation result.

[0034] When executing the computer program, the processor can acquire resting phase voltage and time data from the acquisition device or external system via a communication interface or data bus, and construct a feature vector in memory. Subsequently, the processor calls the anomaly detection logic defined in the program to identify abrupt changes in the health state sequence, and performs corrections and filtering to obtain a processed reference curve. The processor further runs multiple trained prediction models to generate health state prediction results, and calculates the final health state estimate according to the weighted fusion strategy in the program. All of the above program instructions can be implemented in software, or some algorithms can be executed in conjunction with hardware acceleration units to meet the real-time requirements of underwater vehicles.

[0035] Through the above methods, the computer-readable storage medium described in this embodiment enables the lithium battery health status estimation method of the present invention to be deployed in a programmed form in the battery management system, energy system controller of an underwater vehicle, or a shore-based data analysis platform, thereby achieving a reliable estimation of the lithium battery health status. Those skilled in the art can select different types of storage media or adjust the program structure according to the actual application scenario without affecting the technical essence of the present invention.

Claims

1. A method for estimating the health status of lithium batteries based on resting phase characteristics for underwater vehicles, characterized in that, Includes the following steps: Step 1: After the underwater vehicle completes the discharge process, the resting phase is automatically identified based on the condition that the current drops to zero and remains stable; during the resting phase, terminal voltage and time series data are collected, and a resting feature vector reflecting the battery's resting recovery behavior is constructed based on the data; Step 2: Construct an initial health state SOH sequence based on historical capacity data or calibration data; perform anomaly detection on the initial health state SOH sequence, and identify any point that deviates from a preset threshold as an anomaly; perform substitution correction on the detected anomalies to improve the continuity and effectiveness of the initial health state SOH sequence; Step 3: Input the corrected SOH sequence from Step 2 into an adaptive Kalman filter. Dynamically adjust the measurement noise covariance based on the magnitude or trend of the filter residual. This increases the filter gain to enhance correction when the noise is high and decreases the filter gain to maintain smoothness when the sequence is stable. Finally, a continuous, noise-suppressing, and noise-suppressing healthy state reference sequence is obtained, which can be used as a training target for the model. Step 4: Using the resting feature vector from Step 1 as input and the health status reference sequence from Step 3 as output label, construct multiple complementary health status prediction models and train, optimize parameters, and validate them respectively; during actual prediction, input the new resting feature vector into each health status prediction model to obtain multiple independent health status prediction results. Step 5: Perform weighted fusion on the multiple health status prediction results obtained in Step 4, search for the minimum residual combination by weight enumeration, and output the final health status estimate of the underwater vehicle's lithium battery.

2. The lithium battery health status estimation method according to claim 1, characterized in that: Step 1: After the data identification in the resting phase is completed, a resting feature vector is constructed to characterize the voltage recovery behavior by statistically calculating the starting and ending voltages in the resting phase. The resting eigenvector is obtained by averaging several sampling points at the start of the resting state as the resting initiation voltage. U start The average value of several sampling points at the end of the resting period is obtained as the resting termination voltage. U end The voltage recovery amplitude ΔU is obtained based on the difference between the two values; then the voltage recovery rate is calculated based on the duration of the resting phase. re_rate This is used to reflect the voltage recovery characteristics per unit time; simultaneously, the average of the resting start voltage and the resting end voltage is used as the resting average voltage. U mean This constitutes the resting phase characteristics used to predict battery degradation trends.

3. The lithium battery health status estimation method according to claim 1, characterized in that: Step 2 involves anomaly detection of the original health state sequence. The variation amplitude of the health state difference between adjacent cycles is analyzed to identify abrupt changes caused by instrument fluctuations, communication noise, or instantaneous disturbances. When the difference value exceeds a preset threshold, the health state corresponding to that cycle is determined to be an anomaly. The average of the health states on both sides is used to correct the anomaly value, thereby obtaining a continuous and smooth health state sequence.

4. The lithium battery health status estimation method according to claim 1, characterized in that: After anomaly correction is completed, step 3 applies an adaptive Kalman filter to the health state sequence. By constructing a state update model and a measurement model, the observation noise of the health state is dynamically estimated, and the filter gain is adjusted in real time according to the noise change. This ensures that the filtering result can maintain the consistency of the decline trend and effectively suppress local fluctuations, thereby obtaining a smooth health state reference curve as the model training target. The adaptive noise update corrects the measurement noise covariance based on the magnitude and change of the residual at the previous time step, thereby enabling the adaptive Kalman filter to adaptively adjust the noise level at different aging stages.

5. The lithium battery health status estimation method according to claim 1, characterized in that: In step 4, the prediction stage of the multi-health state prediction model trains multiple structurally complementary regression models using resting feature vectors to capture the nonlinear laws of health state changes with cycles. The models used include a random forest regression model based on random subsampling and multi-tree averaging mechanism, an extreme gradient boosting model based on second-order gradient optimization and regularization structure, and a gradient boosting regression model based on stepwise residual fitting. The three models provide complementary prediction information in terms of stability, fine fitting ability, and ability to characterize complex decay patterns, respectively. During the prediction phase, the same resting feature vector is input into the above model to obtain multiple health status prediction results, which are then used as inputs for subsequent fusion.

6. The lithium battery health status estimation method according to claim 1, characterized in that: After obtaining multiple health status predictions from multiple models, step 5 constructs a weighted least squares regression fusion model to weight and combine the prediction results of different health status prediction models. The fusion weights are obtained by searching for each weight combination on the validation set to minimize the sum of the absolute values ​​of the residuals, thereby achieving a comprehensive balance of performance for the random forest regression model based on random subsampling and multi-tree averaging mechanism, the extreme gradient boosting model based on second-order gradient optimization and regularization structure, and the gradient boosting regression model based on stepwise residual fitting. The weights satisfy nonnegativity and normalization constraints, which makes the final prediction results more stable and accurate at different aging stages.

7. A lithium battery health status estimation system based on resting phase characteristics for underwater vehicles, implementing the method described in any one of claims 1-6, characterized in that, include: The feature extraction module is used to acquire data on the resting phase of the lithium battery of the underwater vehicle after discharge, and to construct a resting feature vector to characterize the health status of the battery based on the data. The anomaly detection and filtering module is used to perform differential anomaly detection on the healthy state sequence and correct anomaly points, and to perform adaptive filtering on the corrected sequence to obtain a smooth healthy state reference curve. The sub-model prediction module is used to call multiple constructed regression prediction models based on the resting feature vector as health status prediction models to obtain multiple health status prediction results. The fusion estimation module is used to perform weighted fusion of multiple health status prediction results to obtain the final health status estimate of the lithium battery.

8. The lithium battery health status estimation system according to claim 7, characterized in that: The lithium battery health status estimation system includes a feature extraction module, an anomaly detection and filtering module, a sub-model prediction module, and a fusion estimation module connected in sequence. The feature extraction module identifies the resting phase after the battery has finished discharging and acquires voltage, current, and time data. It constructs a resting feature vector characterizing the battery's polarization relaxation dynamics and long-term degradation trend by calculating the resting phase characteristics such as the resting period voltage recovery amplitude, resting period voltage recovery rate, resting period average voltage, and cycle number. The anomaly detection and filtering module detects jump anomalies in the SOH sequence based on the first-order difference method and corrects them using the mean of adjacent points. It further utilizes adaptive Kalman filtering to suppress noise and dynamically smooth the corrected SOH sequence to obtain a continuous and consistent target output. The sub-model prediction module constructs and trains a random forest regression model, an extreme gradient boosting regression model, and a gradient boosting regression model respectively to predict the health status of the input resting feature vector from multiple angles and through multiple mechanisms. The fusion estimation module performs weighted fusion of the prediction results from each sub-model and outputs the final SOH estimate, achieving multi-model complementarity and prediction result optimization.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements all the steps of the lithium battery health state estimation method based on resting phase characteristics for underwater vehicles as described in any one of claims 1-6.

Images