Battery life prediction and health assessment system based on BMS expansion parameter analysis

By collecting and analyzing battery expansion characteristic data, combined with the deep regression network model, the accuracy of battery life evaluation and health management in the BMS system is solved, and accurate prediction and dynamic regulation of the battery health status are achieved, improving the adaptability and efficiency of the battery management system.

CN120314796AActive Publication Date: 2025-07-15DONGGUAN LITHIUM VALLEY ENERGY CO LTD

Patent Information

Application Number
CN202510806509.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-17
Publication Date
2025-07-15
Estimated Expiration
2045-06-17

AI Technical Summary

Technical Problem

In battery life assessment and health management, existing BMS systems have failed to effectively introduce physical-level characterization indicators, ignoring the changes in battery structure, resulting in lack of deformation dimension support for health assessment and difficulty in adapting to state migration during battery aging. Regulatory strategies lack intelligence and dynamic feedback mechanisms, making it difficult to deploy in different BMS architectures.

Method used

By collecting battery expansion characteristic data, combining electrochemical parameters, a multi-dimensional timing operation data set is constructed, a deep regression network model is used to predict health status, and an expansion characteristic sensitivity adjustment mechanism is introduced to dynamically adjust the BMS control parameters to achieve accurate life prediction and health assessment of the battery.

Benefits of technology

It significantly improves the accuracy of battery health status recognition and life expectancy, enhances the generalization ability of the model, realizes high-precision prediction of different battery types and aging modes, optimizes charging strategies, and extends battery life.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120314796A_ABST
    Figure CN120314796A_ABST
Patent Text Reader

Abstract

The invention relates to a battery life prediction and health assessment system based on BMS expansion parameter analysis, and the system fuses deformation quantity and electrochemical parameters, and remarkably improves the modeling precision and life prediction capability of a battery health state. The acquisition module obtains battery expansion characteristics and operation parameters through a sensor or a modeling mode, and constructs a multi-dimensional time sequence data set; the processing module performs normalization and feature reconstruction on the data to form a health track sample for learning; the prediction module outputs residual life and degradation trend scores based on a deep regression model, and introduces an expansion feature sensitivity adjustment mechanism; the evaluation module evaluates the health level of the battery according to the prediction result and a threshold rule and generates a report; the regulation and control module dynamically adjusts BMS control parameters according to an evaluation result, and optimizes energy efficiency and a safety strategy; the method breaks through the dependence of a traditional BMS system on electrochemical characteristics, realizes more comprehensive and accurate intelligent sensing and life management of the battery health state, and is suitable for various BMS architecture scenes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of batteries, and particularly to a battery life prediction and health assessment system based on BMS expansion parameter analysis. Background Art

[0002] In the prior art, the battery management system (BMS), as a core component of the energy storage system, is widely used in the state monitoring and safety control of lithium-ion batteries. The current mainstream BMS mainly relies on electrochemical parameters such as voltage, current, and temperature for battery state estimation and health management, and realizes remaining life prediction and charge-discharge strategy adjustment through static models or empirical threshold methods. In such solutions, the prediction accuracy and control response are usually limited by the data dimension and model adaptability, and it is difficult to fully reflect the dynamic evolution process of battery performance under complex working conditions.

[0003] With the diversification of battery application environments and the improvement of high-reliability requirements, there are still many problems in the prior art in terms of life assessment and health state perception. First, traditional BMS fails to effectively introduce physical-level characterization indicators and ignores the structural change characteristics during battery operation, resulting in the lack of deformation dimension support for health assessment; second, relying on fixed control logic is difficult to adapt to the state migration during battery aging, and the regulation strategy lacks intelligence and dynamic feedback mechanism; in addition, most solutions lack scalable data fusion methods and are difficult to be generally deployed in different BMS architectures, which restricts the fusion efficiency of algorithms and hardware.

[0004] In view of the above background, there is an urgent need for a new battery health management solution with high integration and adaptability in terms of data dimension, prediction method, and regulation mechanism. Summary of the Invention

[0005] The present application provides a battery life prediction and health assessment system based on BMS expansion parameter analysis to improve the accuracy of battery health state identification and life prediction.

[0006] The present application provides a battery life prediction and health assessment system based on BMS expansion parameter analysis, including: An acquisition module, configured to collect expansion characteristic data of the battery under different charge-discharge cycles based on an expansion sensor, modeling estimation, or deformation quantity inference mechanism connected to the BMS system, and synchronously obtain conventional electrochemical operation parameters including voltage, current, and temperature, and construct a multi-dimensional time series operation data set; A processing module, configured to perform normalization, filtering, and time window reconstruction on the expansion characteristic data and the electrochemical operation parameters, extract a composite feature vector representing the evolution trend of the battery health state, and construct a health trajectory sample for learning in combination with cycle index, load state, and environmental temperature tags; A prediction module, which is used to build and train a deep regression network model based on the health trajectory samples, output the remaining service life and degradation trend score of the battery under the current state, and introduce an inflation feature sensitivity adjustment mechanism during the model update process to enhance the adaptability to different types of inflation behaviors; An evaluation module, which is used to perform hierarchical analysis on the degradation trend score and the remaining service life, combine the set health state threshold and historical performance, determine the life stage and evaluate the health level of the battery cell and the battery cluster, and generate a visual health assessment map and a strategy recommendation report; A regulation module, which is used to receive the remaining service life and the degradation trend score, dynamically adjust the BMS control parameters, including the charging cut-off voltage, the current limit value and the power distribution strategy, and optimize the energy efficiency and protection balance of the control strategy based on the feedback learning mechanism.

[0007] The beneficial effects of this application mainly include: (1) By collecting the inflation feature data of the battery under different charge and discharge cycles, breaking through the limitation of traditional methods that only rely on electrochemical parameters, enabling the system to comprehensively reflect the structural changes of the battery during use, thus significantly improving the accuracy of health state recognition and life prediction. (2) Through normalization, filtering and time window reconstruction processing, the inflation features, electrochemical parameters and working condition labels are unifiedly modeled to construct a composite health trajectory that can be used for machine learning, effectively capturing the dynamic evolution law during the battery degradation process, and enhancing the generalization ability and stability of the model. (3) The prediction module introduces an inflation feature sensitivity adjustment mechanism, which can adapt to different battery types and aging modes, realize high-precision prediction of the remaining service life and degradation trend, and meet the health management requirements under complex usage scenarios. (4) Through the regulation module, the prediction results are fed back to the BMS control logic to realize the real-time optimization of the charging cut-off voltage, the current limit value and the power strategy, improving the energy efficiency utilization rate while effectively extending the battery life. Description of the Drawings

[0008] Figure 1 It is a schematic diagram of a battery life prediction and health assessment system based on BMS inflation parameter analysis provided by the first embodiment of this application. Detailed Embodiment

[0009] Many specific details are set forth in the following description in order to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the connotation of this application. Therefore, this application is not limited by the specific embodiments disclosed below.

[0010] The first embodiment of this application provides a battery life prediction and health assessment system based on BMS inflation parameter analysis. Please refer to Figure 1, this figure is a schematic diagram of the first embodiment of the present application. The following will be combined with Figure 1 A battery life prediction and health assessment system based on BMS expansion parameter analysis provided by the first embodiment of the present application will be described in detail.

[0011] The battery life prediction and health assessment system based on BMS expansion parameter analysis includes a collection module 101, a processing module 102, a prediction module 103, an evaluation module 104, and a regulation module 105.

[0012] The collection module 101 is used to collect the expansion characteristic data of the battery under different charge and discharge cycles based on the expansion sensor, modeling estimation or deformation variable inference mechanism connected in the BMS system, and simultaneously obtain conventional electrochemical operation parameters such as voltage, current, and temperature, and construct a multi-dimensional time series operation data set.

[0013] The collection module 101 is used to collect multi-source operation data required for subsequent health status analysis and life prediction on the basis of the battery management system (BMS), especially physical deformation information including expansion characteristics. Its main functions are as follows: First, the module 101 is connected to the data interface in the BMS in real time, and retrieves the basic electrochemical parameter data related to the operation of the battery cell and the battery pack, including but not limited to voltage, current, temperature, charge and discharge time, SOC (State of Charge), and SOH (State of Health). These parameters can be obtained through conventional sensing devices such as existing current sensors, voltage detection circuits, and thermosensitive elements.

[0014] On this basis, the collection module 101 further introduces the ability to collect the expansion characteristics of the battery. Specifically, the module can obtain the expansion characteristic data through two types of paths: one is the expansion sensor based on the hardware level. The expansion sensor can be a linear or planar deformation detector based on the principles of capacitance, resistance, or fiber Bragg grating (FBG), which is installed on the battery case or module gap and is used to collect the minute deformation amounts caused by factors such as internal gas generation, SEI film growth, or temperature expansion during the charge and discharge process in real time. The other is based on the modeling estimation or inference mechanism, that is, in the scenario where the physical expansion sensor cannot be installed, the indirect expansion amount estimation is carried out by combining the temperature, current rate, and cell type and other parameters in the existing operation data with the pre-built expansion response curve and material model in the laboratory. This type of model can be realized through multiple linear regression, support vector regression (SVR), or neural network models. Its input is the operation state data, and the output is the expansion increment or absolute value per unit time.

[0015] To ensure the consistency and alignment of data in the time dimension, the acquisition module 101 is equipped with a data synchronization control mechanism, which can uniformly schedule the acquisition cycles of various data sources through a sampling clock, so that the expansion feature data and electrochemical parameters are exactly corresponding in the sampling timestamp, facilitating the time window reconstruction of the subsequent processing module. After the data acquisition is completed, the module 101 will structure all the acquired data and store it in the buffer or directly transmit it to the processing module 102, forming a running data set containing multi-dimensional indicators and continuous time series. This data set can be represented in a standard time series format, such as JSON, CSV or binary sequence, and contains fields: timestamp, voltage, current, temperature, expansion amount, battery number and sampling condition label, etc.

[0016] In addition, the acquisition module 101 supports the adaptation and configuration of different types of battery structures and BMS architectures, allowing users to preset the acquisition frequency, expansion sensor sensitivity, redundant sensor fusion method, and whether to enable the modeling acquisition path in the inference mode. Through the decoupling of software and hardware, this module can be deployed in a centralized or distributed BMS environment, with strong versatility and scalability, and can meet the comprehensive perception requirements of the battery operating state in a variety of application scenarios.

[0017] In summary, the acquisition module 101 not only breaks through the limitation of relying only on electrochemical data in traditional BMS, but also constructs a more comprehensive and physically meaningful battery operating state data foundation by introducing battery expansion characteristics as a new monitoring dimension, providing high-quality and high-dimensional data support for subsequent health state assessment and life prediction.

[0018] The processing module 102 is used to normalize, filter and reconstruct the time window of the expansion feature data and electrochemical operation parameters, extract the composite feature vector representing the evolution trend of the battery health state, and construct a health trajectory sample for learning in combination with the cycle index, load state and environmental temperature label.

[0019] The processing module 102 is used to systematically process the expansion feature data and electrochemical operation parameters obtained by the acquisition module 101, aiming to extract a composite feature vector from the original multi-source time-series data that can accurately reflect the evolution trend of the battery health state, and further construct health trajectory samples for training and inference. This module first receives the data set passed in by the acquisition module, and the data structure usually includes multiple dimensions such as timestamp, voltage, current, temperature, expansion feature value, battery identifier, and operating conditions. To eliminate the differences in dimension and scale of each parameter and improve the convergence efficiency and accuracy of subsequent modeling, the processing module first performs normalization processing on all input parameters. The normalization method can adopt Min-Max Scaling or Z-score Normalization, and is differentially set according to the battery type, sensor accuracy, and sampling resolution. For example, the voltage range is usually linearly scaled to the interval [0,1], while the expansion feature may be normalized based on the absolute value and historical peak value.

[0020] After normalization, the module filters the data to remove instantaneous outliers caused by sensor jitter, electromagnetic interference, or abnormal operating conditions. The filtering method can adopt moving average filtering, median filtering, or low-pass filter, and the specific parameters (such as window size or cut-off frequency) are determined according to the sampling frequency and the target time scale. For example, at a sampling frequency of 1Hz, a sliding window with a length of 5 can effectively suppress single-point spike interference while still retaining the degradation trend.

[0021] After completing normalization and filtering, the processing module performs a reconstruction operation based on a fixed-length time window, dividing the continuous time-series data into several overlapping or non-overlapping sample segments. Each segment contains a complete charge-discharge cycle or several key charging stages, which are used to form an input feature matrix. The data within each time window is organized as a two-dimensional tensor in chronological order, where one dimension is the time step and the other dimension is different types of processed features, including expansion value, voltage, current, temperature, and derivative features such as their first-order difference and second-order difference, to enhance the model's perception ability of trend changes.

[0022] To improve the effectiveness of the samples in representing the evolution path of the battery state, the processing module also introduces cycle index, load status, and environmental temperature label as additional features. Among them, the cycle index represents the cumulative number of cycles that the current sample is in, reflecting the battery aging process; the load status distinguishes special operating conditions such as high-power and high-rate charge and discharge by calculating the average current and power volatility; the environmental temperature label represents the external environmental temperature distribution during sampling to help the model identify the impact of the thermal environment on expansion and degradation behaviors. The above features are aligned with the main time-series features during the reconstruction process and encoded into the same input tensor to form a complete composite feature structure.

[0023] Finally, the processing module saves the processed samples above as a standardized health trajectory sample set for the prediction module 103 to perform supervised training or inference analysis. The health trajectory samples are dynamically updated in a sliding window manner, enabling the model to continuously learn the evolution law of the battery state and adapt to the time-varying environment and usage strategy changes in actual operation. The entire processing flow is implemented through software, which can be embedded in the BMS main control firmware or deployed on an external diagnostic platform, featuring low computational complexity, high real-time performance, and high scalability to meet the online prediction data requirements of various battery management scenarios.

[0024] Furthermore, the processing module is specifically configured to: Perform short-time Fourier transform on different frequency sub-bands based on the swelling feature data and electrochemical operating parameters to obtain a frequency band signal tensor, and calculate the cross-correlation coefficient between each frequency band and the swelling feature data in combination with the degree of co-fluctuation of the swelling feature data with voltage and current, generate a corresponding frequency band channel weighting factor, and perform channel-by-channel weighted fusion on the frequency band signal tensor and the frequency band channel weighting factor to obtain a frequency band channel weighted tensor; Based on the frequency band channel weighted tensor, extract the difference between the maximum value and the minimum value of each swelling feature dimension within adjacent time windows, construct a deformation residual vector, and splice and fuse the deformation residual vector with the electrochemical operating parameters within the time window to generate a composite residual coding sequence for characterizing the battery structure degradation trend; Use the composite residual coding sequence as a basic sample, perform enhanced sampling using a sliding window mechanism, and adaptively adjust the sampling window width and step size based on the battery operation cycle density estimation strategy to improve the sample coverage rate in the low-frequency degradation stage, and generate a structure degradation sample set covering all stages of the life cycle; Input the structure degradation sample set into the contrast coding unit, use the structure degradation samples of the same battery cell at different discharge rates or different ambient temperatures as positive sample pairs to construct a variational contrast loss function, and add an invariance constraint during the coding stage to make the finally constructed health trajectory samples maintain consistency and generalization ability in the structure degradation dimension under multi-condition conditions.

[0025] In the battery life prediction and health assessment system based on BMS swelling parameter analysis according to the present invention, the processing module is not only responsible for normalizing, filtering, and time window reconstruction of the collected swelling feature data and electrochemical operating parameters, but also further undertakes the tasks of deep construction of structure degradation trend features and enhancement of expression stability. For this purpose, the processing module introduces a composite processing flow that synergistically combines frequency domain decomposition, deformation coding, sample enhancement, and contrast learning, aiming to fully exploit the dynamic coupling characteristics between swelling and electrochemical signals, and construct high-quality, uniformly distributed health trajectory samples for subsequent model training.

[0026] Specifically, in the initial stage, the system first performs short-time Fourier transform processing on the swelling feature data and the voltage and current in the electrochemical operating parameters, decomposing them into time-frequency domain signals of multiple frequency sub-bands to form a three-dimensional frequency band signal tensor. The dimensions of this tensor include the time window, the frequency band index, and the feature channel. To further identify the sensitivity of each frequency band to structural deformation, the system calculates the cross-correlation coefficient as the channel weighting factor based on the degree of co-fluctuation between the signal of each frequency band and the swelling feature data within the same time window. This weighting factor reflects the adaptability of the frequency band response to physical deformation and is dynamic, changing with different operating states. Subsequently, the system performs channel-by-channel fusion processing on the frequency band signal tensor and the frequency band channel weighting factor, that is, multiplying the signal of each frequency band channel by its corresponding weighting factor to obtain a frequency band channel weighted tensor, thereby enhancing the response ability to key deformation information while maintaining the time-frequency structure.

[0027] Next, the processing module extracts the dynamic change features of structural degradation from the frequency band channel weighted tensor. For this purpose, the system calculates the difference between the maximum value and the minimum value for each swelling feature dimension within adjacent time windows to obtain a set of deformation residual vectors, which reflect the degree of strain mutation within the current period. To fuse multi-modal information, this deformation residual vector is concatenated and combined with the electrochemical operating parameters (including but not limited to the voltage, current, and temperature after normalization) within the same time window, thereby generating a composite residual coding sequence with structural-electrochemical fusion attributes to characterize the complex evolution state of the battery aging trend.

[0028] To address the problem of uneven distribution of the number of structural degradation samples and scarcity of low-frequency samples in different life stages, the system further introduces a sliding window enhanced sampling mechanism. Based on the composite residual coding sequence as the basic sample source, while maintaining the coherence of the time series context, the samples are resampled by means of a sliding window, and the width of the sampling window and the sliding step size are adaptively adjusted according to the battery operation cycle density (such as the swelling change rate per unit time, the charge-discharge cycle frequency, etc.). For time intervals with a lower degradation rate, the sampling window is appropriately narrowed and the sliding step size is slowed down to increase the number of samples in this stage, thereby obtaining a set of structural degradation samples covering the entire life stage of the battery, and performing distribution balance processing on them to improve the representational integrity and class balance of the training samples.

[0029] Finally, to further enhance the consistency of the model's structural degradation characterization under various complex working conditions, the system inputs the above structural degradation sample set into the contrastive encoding unit to construct a variational contrastive training mechanism. In this mechanism, the system selects samples with the same life stage from the structural degradation samples collected from the same battery cell under different discharge rates or environmental temperature conditions as positive sample pairs, and randomly selects samples from other stages or working conditions as negative samples. By introducing a contrastive loss function with invariance constraints, the encoder is guided to learn a stable and highly generalized healthy trajectory representation for the structural degradation dimension in the latent space. This process not only improves the model's robust modeling ability for the structural evolution trend under real working condition changes, but also provides training inputs with consistent semantics and high feature integration for the subsequent prediction module.

[0030] In summary, through four closely connected steps of band weighting, deformation encoding, sample reconstruction, and contrast enhancement, the processing module has completely realized the process of constructing healthy trajectory samples from physical deformation signals that can be used for deep learning.

[0031] The following is the reference implementation code of the processing module: import numpy as np import torch import torch.nn as nn import torch.nn.functional as F # Set global constants EPSILON = 1e - 8 # Assumption: Each sample is a tensor of shape (T, D), where T is the time window length and D is the feature dimension # Among them, the first N_feat dimensions are electrochemical parameters (such as voltage and current), and the last N_expansion dimensions are expansion features def short_time_fourier_transform(data, window_size = 64, hop_length = 32): """ Perform short-time Fourier transform on each channel and return the frequency band tensor (number of samples, number of channels, number of time steps, number of frequency bands) """ # data: (B, T, D) B, T, D = data.shape fft_data = [] for d in range(D): channel_data = data[:, :, d] # (B, T) # Simplified processing: Use abs(fft) to simulate STFT output spec = torch.abs(torch.fft.rfft(channel_data, dim=1)) # (B,freq) fft_data.append(spec.unsqueeze(1)) # Add channel dimension fft_tensor = torch.cat(fft_data, dim=1) # (B, D, freq) return fft_tensor # Output shape (B, D, F) def compute_correlation_weights(fft_tensor, expansion_data): """ Calculate the cross - correlation coefficient between the frequency bands of each channel and the expansion features to obtain the channel weighting factors fft_tensor: (B, D, F) expansion_data: (B, D_exp) Return the weight coefficients: (B, D) """ B, D, F = fft_tensor.shape weights = [] for i in range(D): fft_feat = fft_tensor[:, i, :] # (B, F) exp_feat = expansion_data[:, i % expansion_data.shape[1]] # Assume D is aligned with the expansion feature dimension # Calculate cross - correlation: Simplified to inner product normalization corr = torch.sum(fft_feat exp_feat.unsqueeze(1), dim=1) / (torch.norm(fft_feat, dim=1) (torch.norm(exp_feat, dim = 0)+ EPSILON) weights.append(corr.unsqueeze(1)) # (B, 1) weights = torch.cat(weights, dim = 1) # (B, D) return weights def apply_frequency_weights(fft_tensor, weights): """ Apply channel weighting factors to weight frequency band data channel by channel fft_tensor: (B, D, F) weights: (B, D) Output: (B, D, F) """ return fft_tensor weights.unsqueeze(2) def compute_strain_residual(encoded_tensor): """ Compute the deformation residual within adjacent time windows, which is the difference between the maximum and minimum values Assume the input is (B, D, T) Return: (B, D) """ max_val = torch.max(encoded_tensor, dim = 2)[0] min_val = torch.min(encoded_tensor, dim = 2)[0] return max_val - min_val # Difference tensor (B, D) def sliding_window_augmentation(data, window_size = 10, stride = 5): """ Sliding window augmentation sampling mechanism: suitable for processing sequence data data: (T, D) Return: (N, window_size, D) """ T, D = data.shape samples = [] for start in range(0, T - window_size + 1, stride): end = start + window_size samples.append(data[start:end, :].unsqueeze(0)) # (1, window_size, D) return torch.cat(samples, dim=0) # (N, window_size, D) class ContrastiveEncoder(nn.Module): def __init__(self, input_dim, hidden_dim): super(ContrastiveEncoder, self).__init__() self.encoder = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.ReLU(), nn.Linear(hidden_dim, hidden_dim) ) def forward(self, x): return self.encoder(x) def variational_contrastive_loss(anchor, positive, negative, margin=1.0): """ Variational contrastive loss function: represents the structural invariance constraint through the Euclidean distance in the embedding space anchor, positive, negative: (B, D) """ pos_dist = F.pairwise_distance(anchor, positive) neg_dist = F.pairwise_distance(anchor, negative) loss = F.relu(pos_dist - neg_dist + margin) return loss.mean() # Example workflow: # 1. STFT # stft_data = short_time_fourier_transform(input_tensor) # 2. Compute correlation weights # weights = compute_correlation_weights(stft_data, expansion_features) # 3. Apply weights # fused_tensor = apply_frequency_weights(stft_data, weights) # 4. Compute strain residual encoding # residual = compute_strain_residual(fused_tensor.transpose(1, 2)) # 5. Concatenate electrochemistry parameters # fused_representation = torch.cat([residual, electrochem_features],dim=1) # 6. Augmented sample generation # augmented_samples = sliding_window_augmentation(full_sequence_tensor) # 7. Contrastive training encoding # anchor = encoder(augmented_samples[0]) # positive = encoder(augmented_samples[1]) # negative = encoder(augmented_samples[2]) # loss = variational_contrastive_loss(anchor, positive, negative) # The complete training process will combine all the above components and execute the training steps iteratively in a loop.

[0032] The core workflow of this code starts from the original dilated feature data and electrochemical operating parameters, gradually extracts and constructs a high-quality feature sequence that can reflect the battery structure degradation trend, and enhances the model's generalization ability under multiple working conditions through contrastive learning. The system first performs short-time Fourier transform on the dilated features and voltage and current signals to extract time-frequency tensor signals in multiple frequency bands. Subsequently, the code calculates the degree of co-variation between each frequency band and the dilated features to obtain a set of weighting factors reflecting the importance of frequency bands, which are used to strengthen the channels in the frequency band that are sensitive to the dilated response. After weighted fusion, the system extracts the maximum and minimum differences of each dilated dimension within adjacent time windows from the frequency band signals to form a deformation residual, and concatenates it with operating parameters such as voltage, current, and temperature within the corresponding window to generate a multi-modal fusion residual coding sequence. To address the problem of uneven sample distribution in different life stages, the code uses a sliding window method for sample enhancement sampling and estimates the operating cycle density based on the dilation rate, etc., and adaptively adjusts the sampling strategy to generate a set of structure degradation samples covering the entire life cycle of the battery. Finally, the system inputs these samples into the contrastive encoder network, constructs positive and negative sample pairs and introduces variational contrastive loss to guide the model to learn a degradation trajectory representation that maintains consistency under different working conditions during the training process. The overall process forms a closed-loop processing chain from the original signal to a semantically consistent feature representation, with structural, transferable, and engineering realizability.

[0033] Furthermore, the sliding window mechanism is implemented through the following steps: Based on the battery operating cycle density estimation strategy, divide the composite residual coding sequence into multiple life cycle stage sub-intervals, and within each life cycle stage sub-interval, perform density-aware hierarchical enhancement sampling operations according to the variance of the change rate of the dilated feature dimension in the local time period in the composite residual coding sequence to obtain a preliminarily enhanced set of structure degradation samples for the life cycle stage; Input the preliminarily enhanced set of structure degradation samples for the life cycle stage into the context consistency filtering mechanism, which calculates the consistency of the change direction of the dilated feature dimension between all time steps within each candidate sample window and the change trend consistency of the electrochemical operating parameters, and screens out candidate windows with inconsistent internal structure or conflicting trends based on the weighted results of the change direction consistency and change trend consistency of the dilated feature dimension, and outputs a context consistency sample set; Input the context consistency sample set into a sample retention scoring function. The sample retention scoring function performs multi-factor weighted fusion based on the deformation residual amplitude of the dilated feature dimension, the power spectral density energy of the voltage and current signals, and the confidence of the life cycle stage to which the sample belongs in each candidate sample, outputs a retention scoring result, and filters and retains samples according to a scoring threshold to form a sample set finally used for constructing a structural degradation sample set.

[0034] In the battery life prediction and health assessment system based on BMS dilatation parameter analysis according to the present invention, the sliding window mechanism is not only used for generating time series samples in the traditional sense, but also realizes the balanced coverage and high-quality construction of the structural degradation sample set in all stages of the life cycle by introducing cycle density, adaptive enhancement, context consistency discrimination and sample quality scoring mechanisms.

[0035] First, the system segments the composite residual coding sequence according to the battery operation cycle density estimation strategy. This strategy can estimate the degradation density at different stages based on historical behavior data such as the charge and discharge frequency, the change rate of the expansion rate, and the structural response time of the battery during use. Based on this, the system divides the composite residual coding sequence into multiple life cycle stage sub-intervals, such as the early stage, the slow degradation stage, and the accelerated degradation stage, etc. Each sub-interval represents a specific structural degradation process.

[0036] Next, within each life cycle stage sub-interval, the system performs density-aware analysis on the variance of the change rate within the local time period of the dilated feature dimension in the composite residual coding sequence. The region with a high variance of the change rate is considered to have a stronger ability to express degradation signals and is suitable for sample enhancement. Accordingly, the system performs hierarchical enhanced sampling operations to appropriately increase the sample density of key degradation stages (such as structural mutation points, load perturbation response windows) to obtain a preliminarily enhanced structural degradation sample set of the life cycle stage. This process not only ensures the time continuity of the samples but also precisely regulates the information density.

[0037] Subsequently, the system inputs the preliminarily enhanced sample set into a context consistency filtering mechanism for further screening. This mechanism uses the consistency of the change direction of the dilated feature dimension between all time steps within the sample window as the main indicator, that is, to judge whether the change direction of the deformation amount within this window remains stable in time, and at the same time combines the trend consistency index of electrochemical operating parameters such as voltage and current to evaluate the degree of co-variation between signals. The two consistency indicators are weighted and fused to form a unified evaluation value, and then abnormal windows such as the reversal of the deformation mutation direction and the inconsistency of the electrochemical trend in the sample set are removed, and only the context consistency sample set with consistent structural changes and electrochemical state evolution is retained.

[0038] Finally, the system inputs the above context-consistent sample set into the sample retention scoring function for quality assessment. This scoring function is designed based on multi-dimensional factors, including the magnitude of the maximum deformation residual in the dilation feature dimension, the power spectral density energy distribution of voltage and current signals within the current window, and the confidence index of the sample in the current life cycle stage (such as the weight of the sample in the cycle density model). The system outputs a unified retention score value by weighted fusion of the above factors. If the score exceeds the preset threshold, the system includes the sample in the final structural degradation sample set; otherwise, it is discarded. This scoring mechanism aims to ensure that the samples finally used for model training possess the triple characteristics of structural degradation significance, sufficient signal expression, and life cycle representativeness.

[0039] Through the phased enhanced sampling, context consistency verification, and multi-factor quality scoring of this sliding window mechanism, the system can significantly improve the stability and representativeness of training data while maintaining sample diversity, thereby providing a solid data foundation for subsequent health trajectory modeling and life prediction.

[0040] The prediction module 103 is used to construct and train a deep regression network model based on the health trajectory samples, output the remaining useful life and degradation trend score of the battery under the current state, and introduce a dilation feature sensitivity adjustment mechanism during the model update process to enhance the adaptability to different types of dilation behaviors.

[0041] The prediction module 103 is used to model, train, and infer the health trajectory samples output by the processing module 102. The goal is to output the remaining useful life (RUL) and degradation trend score of the battery under the current state on the basis of fully considering the dilation features, electrochemical features, and their evolution trends, so as to achieve a more accurate and targeted health state prediction. This module first receives the constructed health trajectory samples, which are usually organized as a time series tensor in a sliding window format. Each sample contains multiple time steps, and at each time step, standardized composite features including dilation amplitude, voltage, current, temperature, cycle index, ambient temperature, etc. are recorded.

[0042] The core prediction logic inside the prediction module is built based on a deep regression network model. This model can choose a structure with time series modeling capabilities, such as a bidirectional long short-term memory network (Bi-LSTM), a gated recurrent unit (GRU), a temporal convolutional network (TCN), or a Transformer structure. In a specific implementation, the input layer of the model receives each healthy trajectory sample tensor and maintains its time series structure without being disrupted. Subsequently, through several layers of recurrent or convolutional units, the evolution patterns within the time dimension are extracted, and then regression output is performed through a fully connected layer. Finally, the remaining useful life value and the degradation trend score corresponding to the sample are generated. The remaining useful life value represents the number of cycles or the time length that is expected to maintain normal performance from the current state, and the degradation trend score is a floating point number between 0 and 1, reflecting the relative position of the current health state in the entire service life interval. The closer it is to 1, the closer it is to the severely degraded state.

[0043] To ensure that the prediction results can adapt to the characteristic expression differences under different battery types, different structural designs, or different usage environments, the prediction module introduces an inflation feature sensitivity adjustment mechanism during the training stage. The core idea of this mechanism is to assign adaptive weights to the inflation type variables in the input features and dynamically adjust the contribution of the inflation features to the training objective in the loss function. Specifically, the module can set a group of weight coefficient vectors. In the initial state, the weights of each feature dimension are equal. As the training process progresses, the system automatically increases the weights of the channels that are more sensitive to inflation features, suppresses the weights of redundant channels or features with low signal-to-noise ratios, based on the gradient magnitude, feature activation frequency, and the information backpropagated from the prediction error, so that the model focuses more on the structural indicators with degradation indication significance. This mechanism can be implemented through an attention layer, a feature gating mechanism, or a feature selection network, and can significantly improve the generalization ability of the model in data heterogeneity scenarios.

[0044] During the deployment stage, the prediction module has the ability to continuously learn and can update the model based on the actual operation data. When the system monitors that the prediction results deviate from the actual performance for a long time, the module will automatically trigger an online fine-tuning process, using recent healthy trajectory samples and real life labels (such as forced offline time, manual maintenance records) as new training data to perform parameter fine-tuning or local retraining on the deployed model, thereby improving the prediction accuracy and long-term stability.

[0045] The entire prediction module can be implemented using Python, C++, or other embedded frameworks and supports deployment on cloud servers or local edge devices. In actual engineering, this module communicates with the processing module and the evaluation module through a standard data format interface (such as NumPy arrays or TensorFlow tensors), and can complete the update of health status prediction within a time period of minutes or seconds, meeting the real-time and high-precision requirements of the battery management system for life prediction. Through the above design, the prediction module 103 ensures the intelligent perception ability of the battery degradation trend, makes up for the prediction blind area caused by traditional models ignoring structural indicators, and provides a clear, quantitative, and credible reference basis for subsequent regulation.

[0046] Furthermore, the prediction module introduces a dilation feature sensitivity adjustment mechanism during the training phase. The dilation feature sensitivity adjustment mechanism includes the following steps: When constructing the input tensor of the health trajectory sample, mark the dimension indices containing dilation features and initialize a set of trainable weight coefficient vectors consistent with the number of feature dimensions. The weight coefficient vectors are initially set to a unified constant value; Introduce the weight coefficient vectors into the feature input layer of the prediction model, perform element-wise weighting on each input sample's feature dimensions, and form a weight-corrected input feature tensor; During the model training process, continuously track the absolute value of the gradient and the activation frequency corresponding to the dilation feature dimensions. Combining the backpropagation results of the loss function, calculate the dilation feature importance score, and the dilation feature importance score is updated based on exponential moving average; Dynamically update the weight coefficient vectors according to the dilation feature importance score. Among them, higher weights are assigned to dimensions with higher dilation feature importance scores, and the weights of dimensions with lower dilation feature importance scores are gradually attenuated. The weight update process is achieved through gradient descent or attention normalization mechanism; After each training cycle, use L2 regularization to constrain the weight coefficient vectors to prevent the abnormal amplification of individual dimension weights from causing model instability; Enable the prediction module to gradually enhance its sensitivity to dilation features during multiple rounds of training, improve the modeling accuracy of the structural degradation trend, and enhance the prediction generalization ability of the prediction module under different battery types and operating environments.

[0047] In the battery life prediction and health assessment system based on BMS expansion parameter analysis described in the present invention, a swelling feature sensitivity adjustment mechanism is introduced in the training phase to address the issues that the relative contributions of swelling features are not clear in multi-dimensional feature inputs and are submerged or misjudged as redundant information during the model learning process. Through a set of clear steps, from sample construction to model training, and then to weight update and regularization control, this mechanism systematically enhances the sensitivity of the model to features related to structural degradation, ensuring the stability, adaptability, and generalizability of the battery health prediction process.

[0048] In the training sample construction phase, the system first clearly marks the dimensional indices of all swelling-related features in the input tensor. For example, for a sample with input dimensions, if the -th dimension corresponds to swelling parameters (such as battery housing strain, module gap deformation, estimated swelling amount, etc.), these indices will be recorded in the internal mapping table and used as the target dimensions for subsequent adjustment. The input tensor dimension represents the total number of features of each sample, that is, the number of dimensions of each input vector row.

[0049] For example, the features collected by the system include voltage, current, temperature, battery housing strain, module gap change, and swelling amount estimation, a total of 6 features, then . The swelling feature dimension index indicates which of these features are dimensions related to battery swelling. For example, the 4th, 5th, and 6th features are housing strain, module gap, and swelling amount estimation, then , that is, the swelling feature dimension index is [4, 5, 6].

[0050] Subsequently, the system initializes a weight coefficient vector of length , and the initial values are uniformly set to a constant , such as , to ensure that all features have equal contributions in the initial stage.

[0051] For example, for d = 6, the system initializes = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0], that is, the initial weights of each feature are equal, all 1.0. This vector will be automatically adjusted according to the importance of each feature later.

[0052] This weight coefficient vector will be introduced into the input processing flow of the prediction model as a feature weighting factor. In each input tensor (where In a time window of length T (where T is the length of the time window), according to Equation 1 below, the model will perform an element-wise weighting operation: ; where denotes the Hadamard product broadcast by column, is the weight vector for the current training epoch. This operation ensures that in each training epoch, each feature dimension received by the model has been dynamically weighted, laying the foundation for subsequent backpropagation adjustments.

[0053] is the input tensor, representing the input feature data for a batch or a single sample, where each row is a feature vector at a certain moment, containing a total of T time steps and d features. It is constructed from multi-dimensional time series data (such as voltage, current, temperature, swelling amount, etc.) collected from the BMS system after preprocessing (normalization, filtering, window reconstruction). For example, for T = 10 (representing the past 10 time points) and d = 6 (including voltage, current, temperature, housing strain, gap change, estimated swelling amount), then X is a 10×6 matrix.

[0054] is the weight vector for training epoch t, representing the weighting factor assigned by the model to each feature dimension in the current training epoch, used to emphasize or weaken the role of specific dimensions in the prediction task. This vector is initialized to all 1s (or a unified constant) and is adaptively updated during training based on gradients, activation frequencies, and feature importance scores. For example, in the t-th training epoch, might be: = [0.8, 0.9, 1.0, 1.5, 1.3, 1.6]; indicating that the model considers the 4th, 5th, and 6th dimensions (i.e., swelling features) more important at this time and thus assigns higher weights.

[0055] is the weighted input tensor, representing the new input formed after the features of each time step in a row have been weight-adjusted, used for subsequent model training. It is input to the next layer of the model to enhance the sensitivity of the prediction to key features (such as swelling behavior).

[0056] During the model training process, the system continuously monitors the gradient magnitude and activation frequency of each swelling feature dimension. The absolute value of the gradient reflects the direct influence degree of this feature in the loss function, and the activation frequency reflects its participation in non-linear transformations (such as ReLU, Sigmoid). The system fuses these two types of information to calculate the swelling feature importance score , and its update adopts an exponential moving average strategy: ; is the expansion feature dimension in the training round of the importance score, indicating the current round in, the i th feature (such as the dimension related to the expansion amount) in the relative importance score in model training.

[0057] Indicates the current round in the th dimension in the average gradient value of the loss function, reflecting the direct influence of this feature on the loss function. It is automatically obtained during the backpropagation stage of the model during training. For example, if the gradient of the 5th feature dimension is [0.01, -0.02, 0.015] and the average value is about 0.015, then ≈ 0.015.

[0058] Indicates its activation frequency, which represents the frequency that the value of this feature dimension is not suppressed to zero after passing through the non-linear transformation layer (such as ReLU, Sigmoid) of the neural network, reflecting its "participation degree". It comes from the forward propagation, recording the proportion of the effective output times of the feature in the non-linear transformation. For example, if after the ReLU activation layer, 80 out of 100 values of the 5th feature are non-zero, the activation frequency = 80 / 100 = 0.8.

[0059] is the smoothing factor, such as taking 0.9 to control the smoothness of the score update. This score comprehensively considers the participation degree and the influence degree, and has clear physical meaning and feasibility. This scoring mechanism not only integrates the "historical performance" with strong statistics, but also introduces the "current contribution" with high sensitivity, and is a dynamic weighting strategy with physical rationality, statistical stability and feasibility during the model training process.

[0060] In each training round, the system dynamically adjusts the weights of the corresponding dimensions according to the importance scores of the expansion features of all dimensions. The weights corresponding to the feature dimensions with high scores will be proportionally increased, while the weights of the dimensions with low scores will slowly decay. For example, the following update formula is used for weighted correction: ; is the weight of the th feature dimension in the round , indicating the iThe weighting coefficient of a feature. It is initialized to a uniform constant (such as 1.0) and is dynamically updated after each round of training.

[0061] For the weight of the updated

[0062] dimension, representing the new weight value to be used in the next training round, which is adaptively adjusted based on the current importance of this dimension.

[0063] represents the average value of the importance scores of all dilated feature dimensions, which is obtained by taking the arithmetic mean of the importance scores among the dilated feature dimensions. For example, if the 4th, 5th, and 6th dimensions are dilated features with importance scores [0.10, 0.18, 0.12], then = (0.10 + 0.18 + 0.12) / 3 = 0.1333; is a very small constant to prevent division by zero. This update mechanism is adaptive, that is, if the importance score of a certain dilated feature is significantly higher than the average level, its corresponding weight will be significantly enhanced, and vice versa.

[0064] The system will perform attention normalization on the weight coefficient vector after each update to ensure that its sum is constant and avoid affecting the numerical stability of the overall model.

[0065] To prevent the weights of individual feature dimensions from being amplified out of control due to local activation or short-term perturbations, the system performs L2 regularization on the weight coefficient vector at the end of each training cycle to constrain its norm not to exceed a preset upper limit. The specific normalization formula can be as follows: ; is the weight coefficient vector updated in the current training cycle, representing the feature weight value vector after this round of weight update (such as according to the feature importance score).

[0066] where is a very small constant term to avoid division by zero errors. This process can be performed uniformly once after each round of training, or it can be completed along with the gradient update to control the computational cost.

[0067] Through the above mechanism, the system can gradually focus on the dilation features that play a key role in structural degradation prediction during multiple rounds of training, shifting the model's learning focus from simple electrochemical features to physical deformation features with more fault precursor properties, thus significantly improving the adaptability, accuracy, and stability of the prediction model in complex scenarios. This mechanism is highly realizable and structurally different from traditional feature weighting models. It does not rely on fixed channel structures or manual hyperparameter configurations and can be deployed in various model architectures, with good generality and technological advancement.

[0068] The following is the implementation code of the dilation feature sensitivity adjustment mechanism based on PyTorch: import torch import torch.nn as nn # Set parameters feature_dim = 64 expansion_indices = [5, 12, 27, 41] lambda_ema = 0.9 eta = 0.05 epsilon = 1e - 6 delta = 1e - 5 c_init = 1.0 # Simulate input data: batch_size × time_steps × feature_dim batch_size, time_steps = 32, 10 X = torch.rand((batch_size, time_steps, feature_dim), requires_grad=True) target = torch.rand((batch_size, 1)) # Initialize the weight coefficient vector to the constant c_init w = nn.Parameter(torch.full((feature_dim,), c_init), requires_grad=True) # Initialize the dilation feature importance scoring vector S = torch.zeros(feature_dim) # Define the prediction model class SimpleRegressor(nn.Module): def __init__(self, input_dim): super().__init__() self.fc = nn.Linear(input_dim, 1) def forward(self, x): x = x.mean(dim=1) # Temporal average pooling return self.fc(x) model = SimpleRegressor(feature_dim) criterion = nn.MSELoss() # Weighting function: Hadamard product (per-dimensional feature weighting) def weighted_input(X, w): return X w # Training forward pass X_weighted = weighted_input(X, w) output = model(X_weighted) loss = criterion(output, target) loss.backward() # Inflated feature gradient and activation tracking, and update importance scores and weights with torch.no_grad(): grad = w.grad.abs() activation = X_weighted.abs().mean(dim=(0, 1)) # Update S_i^(t) = λ S_i^(t-1) + (1 - λ) ( A_i) for i in expansion_indices: S[i] = lambda_ema S[i]+ (1 - lambda_ema) (grad[i] activation[i]) # Mean of all dilated feature scores for the current round

[0069] S_mean = S[expansion_indices].mean() + epsilon # Weight update: w_i = w_i (1 + η (S_i - ) / ( + ε)) for i in expansion_indices: correction = 1.0 + eta ((S[i] - S_mean) / S_mean) w[i] = w[i] correction # Attention normalization (sum is constant, mean is 1) w_sum = w.sum() + epsilon w[:] = w / w_sum feature_dim # L2 regularization: w = w / ( _2 + ) norm = torch.norm(w) + delta w[:] = w / norm # Clear gradients model.zero_grad() w.grad.zero_() The above code implements a dilation feature sensitivity adjustment mechanism for the training phase, and the overall process includes the following steps. First, the system extracts time series data from the input feature tensor and performs element-wise weighting on each feature dimension through a set of trainable weight vectors to highlight the influence of dilation-like features. Then, the weighted features are input into a simple regression model to perform forward propagation and loss calculation. In the backpropagation phase, the system updates the importance score according to the gradient magnitude and activation degree of each dilation feature dimension, and dynamically adjusts the weight value corresponding to this dimension accordingly, so as to enhance the response ability to key degradation signals. After that, the system performs normalization and regularization processing on the updated weight vectors to maintain the numerical stability of the model and prevent local overfitting. The whole process is automatically executed in each round of training, enabling the model to gradually enhance its adaptability to dilation parameters and prediction accuracy.

[0070] Furthermore, the prediction module includes a distribution-invariant self-calibration mechanism, and the distribution-invariant self-calibration mechanism includes the following steps: Based on the healthy trajectory samples, the distribution mean and distribution variance of the dilation feature dimensions on different battery cell samples in each training batch are statistically calculated, and the corresponding batch-level distribution deviation index is obtained. The batch-level distribution deviation index is used to quantify the statistical distribution difference of the dilation feature dimensions among different battery batches; Based on the batch-level distribution deviation index, a feature calibration vector corresponding to the dilation feature dimension is generated, and the feature calibration vector and the healthy trajectory samples are processed by element-wise weighted compensation to form a calibrated feature tensor corrected by batch distribution; The calibrated feature tensor is input into a multi-head attention channel selection network, and the multi-head attention channel selection network assigns corresponding attention weights to each dilation feature dimension in the calibrated feature tensor and outputs a channel-weighted embedding representation with channel selection ability; The channel-weighted embedding representation is respectively input into a recurrent neural network branch and a Transformer network branch. The recurrent neural network branch is used to model the historical state dependence relationship of the channel-weighted embedding representation, and the Transformer network branch is used to model the non-local interaction relationship between the dilation feature dimension and the charging rate in the channel-weighted embedding representation. A structure prior guidance term is introduced into the loss function, and the known theoretical function relationship between the dilation feature dimension and the charging rate is used as the regularization target to improve the modeling stability and physical consistency of the remaining useful life and degradation trend score.

[0071] In the battery life prediction and health assessment system based on BMS expansion parameter analysis described in the present invention, the distribution invariance self-calibration mechanism in the prediction module is a key component for enhancing the cross-batch generalization ability and physical consistency of the model. This mechanism is mainly used to address the issue of statistical distribution shift of expansion characteristics caused by different battery batches, models, or production processes in practical applications, ensuring that the model can still maintain prediction accuracy and stability under multi-source data input.

[0072] First, during the training process of the model, the system continuously collects the health trajectory data of all samples within the current batch, and for all expansion feature dimensions, calculates the distribution mean and distribution variance of each dimension in this training batch dimension by dimension. This calculation is completed based on the sample set within the entire training batch. The samples are from different battery cells and have a certain degree of statistical heterogeneity. For each expansion feature dimension, the system also refers to a reference distribution or prior expectation (such as the previous batch or historical average statistic) to measure its distribution drift. On this basis, the system introduces a batch-level distribution deviation index as a standardized measure to quantify the degree of distribution change of the current expansion feature dimension. This index can be calculated using variants of KL divergence, Wasserstein distance, or Z-score difference, and is used to characterize the statistical deviation degree between the expansion behavior in the current sample and the reference distribution.

[0073] Next, the system generates a feature calibration vector corresponding one-to-one to the expansion feature dimension according to the batch-level distribution deviation index. This feature calibration vector usually exists in vector form, and its number of dimensions is the same as the number of expansion feature dimensions. Each element indicates how the corresponding dimension should be compensated. Feature calibration can be a combination of scaling, shifting, or exponential transformation to maximize the compression of batch-induced distribution differences. The system directly applies this feature calibration vector to the original health trajectory samples, and during the process of dimension-by-dimension weighted compensation, applies the corresponding correction parameters to each expansion feature dimension, thereby obtaining a calibrated feature tensor that has been adaptively corrected for batch distribution. This tensor not only retains the temporal structure of the original samples but also has cross-batch distribution consistency, and can provide a more standardized input representation for the subsequent modeling stage.

[0074] After calibration is completed, the system inputs the calibrated feature tensor into a structure designed as a multi-head attention channel selection network. The core objective of this network is to selectively enhance important parts in the inflated feature dimension. In each attention head, the network dynamically assigns attention weights based on the representation strength, stability of each inflated feature dimension in the current sample, and its correlation with the prediction target (such as the life score). The introduction of multiple attention heads can comprehensively evaluate the importance of each channel from different perspectives, making the finally output channel-weighted embedding representation highly information-focused, that is, the system only strengthens the dimensions that are indicative of the life degradation trend and weakens the influence of redundant or unstable dimensions.

[0075] To further improve the performance of the model in modeling the time series and structural physical relationship, the system parallelly inputs the above-mentioned channel-weighted embedding representation into two complementary sub-networks: the recurrent neural network branch and the Transformer network branch. Among them, the recurrent neural network branch is mainly used to capture the dependence information of the battery in historical charge and discharge cycles, such as the trend fluctuations in the time series, periodic inflation patterns, etc.; while the Transformer network branch focuses on modeling the non-local interaction relationship between the inflated feature dimension and the charge rate, especially its context modeling ability is particularly prominent when there is a potential non-linear response between different dimensions. The two branches respectively extract the deep representations of time series dependence and structural coupling, and then merge them to form a complete representation vector for outputting the remaining useful life and degradation trend score.

[0076] It should be noted that in this joint modeling process, the system introduces a structural prior guidance term as a regularization component in the loss function to improve the physical consistency of the model. This guidance term is based on the known functional relationship between the inflated feature dimension and the charge rate in experimental or theoretical knowledge (such as as the rate increases, the shell strain increases linearly or exponentially), and is embedded in the loss function in the form of squared difference, KL divergence or fitting residual, so as to impose structural constraints on the model prediction, making it conform to the battery physical mechanism while regressing the life index.

[0077] Through the above multi-step collaborative mechanism, the prediction module can still maintain a high degree of modeling stability, prediction consistency and cross-cell general ability in the case of statistical deviation in the input health trajectory sample, meeting the actual needs of battery life and health state prediction in complex industrial application environments.

[0078] The evaluation module 104 is used to hierarchically analyze the degradation trend score and the remaining useful life, and combine the set health state threshold and historical performance to determine the life stage and evaluate the health level of the battery cell and the battery cluster, generating a visual health assessment map and a strategy recommendation report.

[0079] The evaluation module 104 is used to further logically analyze and determine the status of the degradation trend score and remaining useful life information output by the prediction module 103, and then establish a clear and operable health level system to support the operation safety assessment and maintenance strategy formulation of the battery system. The core task of this module is to form a hierarchical and labeled health status evaluation mechanism based on the numerical prediction results, combined with preset rules and historical data, to ensure that the system not only has the prediction ability, but also has the health assessment ability for actual implementation.

[0080] In the specific implementation process, the evaluation module first receives two core indicators from the prediction module, namely the degradation trend score and the remaining useful life. The degradation trend score is usually a floating point number between 0 and 1, indicating the relative degradation degree of the current state in the entire life cycle. The remaining useful life is a value in units of time, cycle count, or energy output, representing the sustainable usage ability of the battery under the current state. The evaluation module performs hierarchical processing based on these two values. First, multiple health status thresholds are set, such as 0.2, 0.4, 0.6, 0.8, etc., as the demarcation points for state changes. Different score intervals will correspond to different health levels, such as labels like "excellent", "good", "medium", "poor", "critical", etc. Such thresholds can be obtained through large-scale life test data or historical operation data statistics, or can be custom-set according to specific application scenarios and safety standards.

[0081] At the same time, to avoid the abnormal values at a single moment misleading the evaluation results, the evaluation module uses methods such as moving window averaging or weighted time series analysis to smooth the input degradation score, ensuring the stability and anti-interference ability of the health level determination. On this basis, the module also introduces a comparison mechanism of historical performance. For example, the evaluation result of the current battery is compared with the health level change trend in the past several cycles to determine whether there is a rapid decline, non-linear degradation, or abnormal fluctuation in the health level, so as to assist in identifying potential safety risks.

[0082] In addition to the evaluation at the battery cell level, the module further conducts an aggregated evaluation of the battery cluster or the entire system. This process includes weighted summarization, normalization processing, or deviation analysis of the health level results of multiple battery cells, identifying the cells with abnormal performance or significantly faster degradation speed than the population mean, and generating a health level label at the battery cluster level. Such processing is particularly crucial for systems with parallel structures, series structures, or battery pack levels, and can identify internal consistency problems, thermal runaway risks, or local performance bottlenecks.

[0083] To facilitate system integrators and operators in intuitively grasping the evaluation results, the evaluation module visualizes the analysis results in the form of graphs. The graphs may include a health trend curve graph, a remaining life decay graph, a battery cluster consistency heat map, etc., all of which are updated in real time based on the input data and have interactivity and exportability. In addition, the module will also combine the current health level and historical trends of the battery to output a strategy recommendation report, including specific operation recommendations such as whether to recommend derating operation, whether early warning maintenance is required, and whether to recommend replacing specific battery cells, to support the BMS or the upper control system in making response decisions.

[0084] Overall, the evaluation module not only has the ability to analyze numerical values, but also integrates the establishment of a health level system, the introduction of an anomaly recognition mechanism, the dynamic analysis of historical trends, and the generation of a human-machine interaction graph in its functional logic, constituting the core of the evaluation in the intelligent health management of the battery system, ensuring that the system has the ability of perceivable, interpretable, and executable state diagnosis.

[0085] The regulation module 105 is used to receive the remaining service life and degradation trend scores, and dynamically adjust the BMS control parameters, including but not limited to the charging cut-off voltage, current limit, and power distribution strategy, to optimize the energy efficiency and protection balance of the control strategy based on the feedback learning mechanism.

[0086] The regulation module 105 is used to convert the degradation trend scores and remaining service life prediction results provided by the evaluation module 104 into control strategy adjustment operations that can directly act on the battery management system (BMS), so as to achieve dynamic energy efficiency optimization and safety protection based on the actual health status of the battery. This module not only has the ability to update control parameters in real time, but also can continuously optimize the control strategy through the feedback learning mechanism to form a closed-loop adaptive control system driven by prediction.

[0087] In practical applications, the regulation module first receives the result data from the evaluation module, including the remaining service life and current degradation trend scores of each battery cell or battery cluster. These results are interpreted and processed through the strategy mapping logic inside the module. Specifically, the system presets a set of charge and discharge parameter configuration ranges for different health levels. For example, when a certain battery cell is in the "excellent" or "good" state, a higher charging cut-off voltage and charging current are allowed to maintain a higher energy density output; while when the battery health level drops to "medium" or "poor", the allowed charging cut-off voltage and discharge depth are automatically reduced, and the maximum current limit is restricted, so as to slow down its further degradation and extend its available life.

[0088] To avoid sudden changes in policy adjustments, the regulation module introduces slope limits and gradual change logic to the process of changing control parameters. For example, when the life prediction result shows a rapid downward trend, the module does not immediately execute a large-scale parameter reduction, but instead makes restrictive adjustments in stages, enabling the control changes to be gradually completed over multiple sampling periods, ensuring the smooth operation of the battery and the consistency of the system user experience. In addition, to improve the adaptability and personalization of the control strategy, the regulation module constructs a control strategy optimization subsystem based on a feedback learning mechanism. This subsystem records the state responses before and after each control parameter adjustment, including the actual working voltage, battery temperature rise, SOC change rate, and the output results of the new round of evaluation, and uses these as feedback signals to construct a reinforcement learning or online fine-tuning model for continuously optimizing the policy mapping relationship, enabling the control logic to continuously approach the optimal solution during actual operation.

[0089] The control output interface of the regulation module is docked with the BMS main control unit through standard protocols, such as CAN, RS485, or SPI interfaces, ensuring that parameter update instructions can be sent and applied to the execution layer in milliseconds. At the same time, this module supports multi-cell collaborative control. When detecting significant differences in the health status of cells within the battery cluster, mechanisms such as grouped current limiting, energy balancing, or dynamic load adjustment can be adopted to reduce local risks while maintaining overall performance. In the energy management system (EMS) or vehicle control system (VCU), this module can also provide status synchronization signals for adjusting the global power scheduling strategy or operation mode selection, such as entering an energy-saving mode or standby operation state when the health status deteriorates.

[0090] The regulation module has good configurability, allowing maintenance personnel or system developers to customize various control strategy mapping rules, parameter adjustment steps, safety limit boundaries, and learning rate parameters, and also supports remote iteration and upgrade of the policy logic through OTA update methods. In actual deployment, this module can be integrated into the embedded BMS firmware or run as the upper-layer energy management control logic on a remote monitoring server or edge control platform.

[0091] In summary, the regulation module 105 not only completes the dynamic mapping from the health prediction result to the control behavior, but also enables the battery to maintain within a safe, stable, and efficient control range under various operating scenarios through the feedback learning mechanism and soft limit adjustment strategy, ensuring that the system has the ability to continuously operate and intelligent adaptive performance.

[0092] Furthermore, the regulation module is specifically used for: Construct a residual service life interval mapping table based on the remaining service life, and dynamically adjust the charging cut-off voltage according to the interval to which the current remaining service life belongs; The adjusted charging cut-off voltage and the degradation trend score are jointly input into the life cycle adaptive current limiting strategy evolution model to generate a corresponding current limiting control curve; Based on the historical maximum swelling rate of change recorded by the processing module, an envelope function is constructed to correct the current limiting control curve, and a power regulation strategy with amplitude limiting constraints is output; and when a mutation occurs in the swelling feature dimension score, the activation order of the protection subunit in the BMS control structure is adjusted according to the power regulation strategy, and the current reduction, charging inhibition and heat dissipation control programs are preferentially started.

[0093] In the battery life prediction and health assessment system based on BMS swelling parameter analysis according to the present invention, the regulation module undertakes the key function of dynamically adjusting the BMS control parameters according to the prediction results, specifically including the intelligent correction of core parameters such as the charging cut-off voltage, current limit value and power distribution strategy, so as to realize the comprehensive optimization of battery safety, efficiency and life. The implementation of this module is based on the remaining service life and degradation trend score output by the system prediction module, and combined with the swelling change data in the historical record, and the adaptive evolution and phased response of the control strategy are realized through multiple highly coupled steps.

[0094] First, after the system outputs the current remaining service life of the battery by the prediction module, this life value is used as an index to look up a pre-constructed mapping table of remaining service life intervals. This mapping table divides the remaining battery life into several interval segments, and each interval corresponds to a corresponding target charging cut-off voltage value. For example, when the prediction result shows that the remaining life of a certain battery is in the interval of 30% to 40%, the system automatically selects the charging cut-off voltage value corresponding to this interval, such as 4.05V, to replace the initial standard value of 4.2V, so as to reduce the battery working stress and delay the degradation process.

[0095] Subsequently, the system inputs the dynamically adjusted charging cut-off voltage and the degradation trend score obtained in the same round of evaluation into the life cycle adaptive current limiting strategy evolution model. This model is used to adaptively generate a current limiting control curve according to the severity of structural degradation, that is, to set the upper limit of charging or discharging current at different remaining life stages. For example, the current of a single battery with a high degradation score is reduced to 0.5C, and the current of a battery with a low degradation score is maintained above 1C, ensuring that it operates at a relatively conservative current level in the current health state.

[0096] After obtaining the preliminary current-limiting curve, the system further introduces a dynamic feedback mechanism at the structural level, that is, it calls the maximum historical inflation rate of change recorded by the processing module as the deformation fluctuation index, and constructs an envelope function reflecting the severity of structural stress. This envelope function usually adopts the weighted exponential smoothing form of the maximum rate of change sequence of the sliding window, indicating the physical expansion stability of the battery during operation. Once the rate of change in the current cycle is higher than the historical steady-state interval, this envelope function will lower the current threshold in the current-limiting curve to further enhance the protection intensity, thereby generating a power regulation strategy with amplitude-limiting constraint ability to reduce the risk of structural stress-induced failures.

[0097] Finally, while executing the power regulation strategy, the system also monitors the fluctuation behavior of the inflation feature dimension score. If it is found that this score rises sharply continuously in multiple time windows, indicating a potential abnormal trend in the battery structure, the system will automatically adjust the activation priority of each protection subunit in the BMS control structure according to the preset strategy. At this time, modules such as current reduction, charge inhibition, and heat dissipation control are preferentially activated to achieve timely proactive protection responses. For example, in the stage with a high structural degradation score and limited power fluctuation, the system will prohibit continued charging and force the heat dissipation module to be enabled to stabilize the system state and delay further deterioration.

[0098] In summary, through the logical deconstruction and continuous reaction of the prediction results, the regulation module realizes the optimization of the multi-dimensional control strategy driven by structural degradation, with the advantages of fine granularity, high responsiveness, and high adaptability, providing the closed-loop health management and life extension regulation capabilities for this system.

[0099] The life cycle adaptive current-limiting strategy evolution model mainly consists of three parts, which are responsible for strategy stage identification, current-limiting curve generation, and strategy dynamic evolution tasks respectively. The entire model takes the remaining service life and degradation trend score collected and predicted by the system as the core input, supplemented by some operation history indicators, and jointly drives the adaptive adjustment of the charge and discharge control parameters.

[0100] In the strategy stage identification part, the system first looks up the life interval mapping table according to the current remaining service life value. This mapping table is pre-divided into multiple stages according to the full life cycle of the battery, such as the initial stage, stable stage, slow degradation stage, and accelerated decline stage, etc. Each interval obtains the typical feature range through a large number of historical data statistics and corresponds to a set of reference current-limiting strategy parameters. In practical applications, the system can use linear interpolation or fuzzy matching methods to map the continuous life value to the closest interval label, so as to identify the life stage of the current battery. This stage label will be used as one of the inputs for the subsequent strategy generation module.

[0101] The current-limiting curve generation part takes the above-mentioned life stage labels and degradation trend scores as combined inputs, and combines auxiliary features such as the ambient temperature during operation, the change trend of battery internal resistance, and the voltage fluctuation amplitude, and calls the policy template library to generate a current-limiting control curve. The template library stores multiple predefined policy functions, including a power limit model based on a piecewise linear function, a current regulation model based on an exponential decay curve, or a response mechanism based on rule reasoning. In implementation, the system can adopt a conditional weighting method to weight and combine multiple policy functions according to the importance of the life stage and degradation score, so as to form a current-limiting output curve covering the current working conditions. For example, when the degradation score is relatively high, the system preferentially activates the template that inhibits fast charge and discharge; while when the temperature is on the high side or the internal resistance increases, the current-limiting amplitude is moderately increased to balance the thermal management requirements.

[0102] The policy dynamic evolution part is used for the adaptive update of model parameters during long-term operation. This module continuously tracks the deviation between the actual power output and the pre-defined current-limiting policy, and dynamically adjusts the key parameters in the policy function in combination with the change trend of the maximum change rate of the expansion feature dimension. For example, when the system detects that the change rate of the expansion amplitude suddenly increases in the short term and the current policy does not fully suppress the power output, the system will correct the upper limit of the current-limiting curve according to the envelope function principle. This envelope function can be designed as a non-linear response function based on the maximum expansion change rate in the recent several cycles, such as implemented by a Gaussian function or a piecewise cosine function, and smoothly suppresses the control output exceeding the risk threshold within the power adjustment interval. The finally corrected current-limiting curve will be used as the instruction basis for the regulation module and act on the core policies such as the charging cut-off voltage, current limit value, and power distribution in the BMS control logic.

[0103] In order to ensure the stability and physical consistency of policy evolution, the system also executes a policy drift detection mechanism in each round of update. This mechanism is used to avoid drastic fluctuations in the current-limiting policy caused by short-term noise or abnormal samples, and can be realized through standard deviation constraints within a sliding window, policy curve smoothness constraints, or structural prior knowledge matching rules, so as to achieve a balance between stability and responsiveness.

[0104] Through the continuous processing of the above three links, the life cycle adaptive current-limiting policy evolution model can effectively integrate the health prediction results and the dynamic signals of structural deformation, realize the personalized adjustment and high-frequency response of the battery charge and discharge policy, provide a control policy with actual physical constraint significance for the BMS, and enhance the controllability of battery life and the safety of system operation.

[0105] Although this application is disclosed above with preferred embodiments, it is not used to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the protection scope of this application should be subject to the scope defined by the claims of this application.

Claims

1. A battery life prediction and health assessment system based on BMS expansion parameter analysis, characterized in that, Including: A collection module, which is used to collect the expansion characteristic data of the battery under different charge and discharge cycles based on the expansion sensors, modeling estimation or deformation variable inference mechanism accessed in the BMS system, and synchronously obtain the conventional electrochemistry operation parameters including voltage, current, and temperature, and construct a multi-dimensional time-series operation data set; A processing module, which is used to normalize, filter and reconstruct the time window of the expansion characteristic data and the electrochemistry operation parameters, extract the composite feature vector representing the evolution trend of the battery health state, and combine the cycle index, load state and environmental temperature label to construct a health trajectory sample for learning; A prediction module, which is used to construct and train a deep regression network model based on the health trajectory sample, output the remaining service life and degradation trend score of the battery under the current state, and introduce an expansion feature sensitivity adjustment mechanism during the model update process to enhance the adaptability to different types of expansion behaviors; An evaluation module, which is used to perform hierarchical analysis on the degradation trend score and the remaining service life, and combine the set health state threshold and historical performance to determine the life stage and evaluate the health level of the battery monomer and the battery cluster, and generate a visual health evaluation map and a policy recommendation report; A regulation module, which is used to receive the remaining service life and the degradation trend score, dynamically adjust the BMS control parameters, including the charging cut-off voltage, current limit and power distribution strategy, and optimize the energy efficiency and protection balance of the control strategy based on the feedback learning mechanism.

2. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 1, wherein The prediction module introduces an expansion feature sensitivity adjustment mechanism in the training stage, and the expansion feature sensitivity adjustment mechanism includes the following steps: When constructing the input tensor of the health trajectory sample, mark the dimension index containing the expansion feature, and initialize a set of trainable weight coefficient vectors consistent with the number of feature dimensions, and the weight coefficient vectors are initially set to a unified constant value; Introduce the weight coefficient vector into the feature input layer of the prediction model, and perform element-wise weighting processing on the feature dimensions of each input sample to form an input feature tensor corrected by the weight; During the model training process, continuously track the absolute value of the gradient and the activation frequency corresponding to the expansion feature dimension, and combine the backpropagation result of the loss function to calculate the expansion feature importance score, and the expansion feature importance score is updated based on the exponential moving average; Dynamically update the weight coefficient vector according to the expansion feature importance score. Among them, a higher weight is assigned to the dimension with a higher expansion feature importance score, and the weight of the dimension with a lower expansion feature importance score is gradually attenuated. The weight update process is realized by the gradient descent method or the attention normalization mechanism; After each training cycle, use the L2 regularization method to constrain the weight coefficient vector to prevent the abnormal amplification of the weight of individual dimensions from causing model instability.

3. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 1, wherein The processing module is specifically used for: Perform short-time Fourier transform on different frequency sub-bands based on the inflation feature data and electrochemical operation parameters to obtain a frequency band signal tensor. Combine the inflation feature data with the co-fluctuation degree of voltage and current, calculate the cross-correlation coefficient between each frequency band and the inflation feature data, generate a corresponding frequency band channel weighting factor, and perform channel-by-channel weighted fusion on the frequency band signal tensor and the frequency band channel weighting factor to obtain a frequency band channel weighted tensor; Based on the frequency band channel weighted tensor, extract the difference between the maximum and minimum values of each inflation feature dimension within adjacent time windows, construct a deformation residual vector, and splice and fuse the deformation residual vector with the electrochemical operation parameters within the time window to generate a composite residual coding sequence for characterizing the battery structure degradation trend; Use the composite residual coding sequence as a basic sample, perform enhanced sampling using a sliding window mechanism, and adaptively adjust the sampling window width and step size based on the battery operation cycle density estimation strategy to improve the sample coverage rate in the low-frequency degradation stage, and generate a structure degradation sample set covering all stages of the life cycle; Input the structure degradation sample set into the contrast coding unit, use the structure degradation samples of the same battery cell at different discharge rates or different ambient temperatures as positive sample pairs, construct a variational contrast loss function, and add an invariance constraint during the coding stage to make the finally constructed healthy trajectory samples maintain consistency and generalization ability in the structure degradation dimension under multi-condition conditions.

4. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 3, wherein The sliding window mechanism is implemented through the following steps: Based on the battery operation cycle density estimation strategy, divide the composite residual coding sequence into multiple life cycle stage sub-intervals, and within each life cycle stage sub-interval, perform density-aware hierarchical enhanced sampling operations according to the variance of the change rate of the inflation feature dimension in the local time period of the composite residual coding sequence to obtain a preliminary enhanced life cycle stage structure degradation sample set; Input the preliminary enhanced life cycle stage structure degradation sample set into the context consistency filtering mechanism. This mechanism calculates the consistency of the change direction of the inflation feature dimension and the consistency of the change trend of the electrochemical operation parameters between all time steps within each candidate sample window, and screens out candidate windows with inconsistent internal structure or conflicting trends based on the weighted results of the inflation feature dimension change direction consistency and change trend consistency, and outputs a context consistency sample set; Input the context consistency sample set into the sample retention scoring function. The sample retention scoring function performs multi-factor weighted fusion based on the deformation residual amplitude of the inflation feature dimension, the power spectral density energy of the voltage and current signals, and the life cycle stage confidence of the sample, outputs a retention scoring result, and filters and retains samples according to the scoring threshold to form the sample set finally used to construct the structure degradation sample set.

5. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 1, characterized in that, The prediction module includes a distribution invariance self-calibration mechanism, and the distribution invariance self-calibration mechanism includes the following steps: Based on the healthy trajectory samples, the distribution mean and distribution variance of the inflated feature dimensions on different battery cell samples in each training batch are statistically calculated, and the corresponding batch-level distribution deviation index is obtained. The batch-level distribution deviation index is used to quantify the statistical distribution difference of the inflated feature dimensions among different battery batches; Based on the batch-level distribution deviation index, a feature calibration vector corresponding to the inflated feature dimension is generated, and the feature calibration vector and the healthy trajectory samples are subjected to per-dimensional weighted compensation processing to form a calibrated feature tensor after batch distribution correction; The calibrated feature tensor is input into the multi-head attention channel selection network. The multi-head attention channel selection network assigns corresponding attention weights to each inflated feature dimension in the calibrated feature tensor and outputs a channel-weighted embedding representation with channel selection ability; The channel-weighted embedding representation is respectively input into the recurrent neural network branch and the Transformer network branch. The recurrent neural network branch is used to model the historical state dependence relationship of the channel-weighted embedding representation, and the Transformer network branch is used to model the non-local interaction relationship between the inflated feature dimension and the charging rate in the channel-weighted embedding representation. A structural prior guidance term is introduced into the loss function, and the known theoretical function relationship between the inflated feature dimension and the charging rate is used as the regularization target to improve the modeling stability and physical consistency of the remaining useful life and degradation trend score.

6. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 1, characterized in that, The regulation module is specifically used for: Construct a residual service life interval mapping table based on the remaining useful life, and dynamically adjust the charging cut-off voltage according to the interval to which the current remaining useful life belongs; The adjusted charging cut-off voltage and the degradation trend score are jointly input into the life cycle adaptive current limiting strategy evolution model to generate a corresponding current limiting control curve; Based on the historical maximum inflation rate of change recorded by the processing module, a envelope function is constructed to correct the current limiting control curve, and a power regulation strategy with amplitude limiting constraints is output; and when a mutation occurs in the inflated feature dimension score, the activation order of the protection sub-unit in the BMS control structure is adjusted according to the power regulation strategy, and the current reduction, charging inhibition and heat dissipation control programs are preferentially started.

Citation Information

Patent Citations

  • Method for predicting residual life of battery based on MTCN algorithm

    CN114660495A

  • Battery life prediction using machine learning models

    CN116113961A

  • Battery bulging and life prediction method, electrochemical device and electric equipment

    CN116540096A

  • Battery state monitoring analysis method and system based on machine learning

    CN119881669A

  • Battery swelling and life prediction method, electrochemical apparatus, and electric device

    US20240329148A1

Cited By

  • Flexible control method and system for dynamically reconstructing battery module

    CN120810882A

  • Method and system for predicting service life of hydrogen fuel cell

    CN120854617A

  • A hydrogen fuel cell life prediction method and system

    CN120854617B

  • Quality management method and system for new energy power station

    CN120875852A

  • A quality management method and system for a new energy power station

    CN120875852B