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

Abstract

The present invention relates to a battery life prediction and health assessment system based on BMS expansion parameter analysis. The system integrates deformation variables and electrochemical parameters, significantly improving the modeling accuracy and life prediction capability of the battery health status; the acquisition module obtains the battery expansion characteristics and operating parameters through sensors or modeling methods, and constructs a multi-dimensional time series data set; the processing module normalizes and reconstructs the features of the data to form a health trajectory sample for learning; the prediction module outputs the remaining life and degradation trend score based on the deep regression model, and introduces an expansion feature sensitivity adjustment mechanism; the evaluation module evaluates the battery health level according to the prediction results and threshold rules and generates a report; the control module dynamically adjusts the BMS control parameters according to the evaluation results, and optimizes energy efficiency and safety strategies; the present invention breaks through the traditional BMS system's dependence on electrochemical characteristics, realizes more comprehensive and accurate intelligent perception and life management of the battery health status, and is suitable for a variety of BMS architecture scenarios.

Description

Technical Field

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

[0002] In existing technologies, battery management systems (BMS), as a core component of energy storage systems, are widely used for condition monitoring and safety control of lithium-ion batteries. Current mainstream BMSs primarily rely on electrochemical parameters such as voltage, current, and temperature for battery state estimation and health management. They also use static models or empirical threshold methods to predict remaining life and adjust charge and discharge strategies. In these approaches, prediction accuracy and control response are often limited by data dimensionality and model adaptability, making it difficult to fully reflect the dynamic evolution of battery performance under complex operating conditions.

[0003] With the increasing diversification of battery application environments and the increasing demand for high reliability, existing technologies still face numerous challenges in lifespan assessment and health status perception. First, traditional BMSs fail to effectively incorporate physical-level characterization indicators, ignoring the structural changes during battery operation, resulting in a lack of deformation-based support for health assessments. Second, reliance on fixed control logic makes it difficult to adapt to state transitions during battery aging, and control strategies lack intelligence and dynamic feedback mechanisms. Furthermore, most solutions lack scalable data fusion methods, making them difficult to deploy universally across different BMS architectures, limiting the efficiency of integrating algorithms and hardware.

[0004] Given the above background, there is an urgent need for a new battery health management solution that is highly integrated and adaptable in terms of data dimensions, prediction methods and control mechanisms. Summary of the Invention

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

[0006] This application provides a battery life prediction and health assessment system based on BMS expansion parameter analysis, including:

[0007] The acquisition module is used to collect the battery's expansion characteristic data under different charge and discharge cycles based on the expansion sensor, modeling estimation, or deformation inference mechanism connected to the BMS system, and simultaneously obtain conventional electrochemical operating parameters such as voltage, current, and temperature to construct a multi-dimensional time-series operation data set;

[0008] a processing module for normalizing, filtering, and time-windowing the expansion feature data and electrochemical operating parameters, extracting a composite feature vector representing the evolution trend of the battery health state, and constructing a health trajectory sample for learning by combining the cycle index, load state, and ambient temperature labels;

[0009] A prediction module 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 in 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;

[0010] An assessment module is used to perform hierarchical analysis of the degradation trend score and remaining service life, and to determine the life stage and health level of battery cells and battery clusters based on the set health status threshold and historical performance, and to generate a visual health assessment map and strategy recommendation report;

[0011] The control module is used to receive the remaining service life and 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 a feedback learning mechanism.

[0012] The beneficial effects of this application mainly include: (1) By collecting the expansion characteristic data of the battery under different charge and discharge cycles, the limitation of relying solely on electrochemical parameters is broken through, so that the system can fully reflect the structural changes of the battery during use, thereby significantly improving the accuracy of health status identification and life prediction. (2) Through normalization, filtering and time window reconstruction processing, the expansion characteristics, electrochemical parameters and working condition labels are unified into a model to construct a composite health trajectory that can be used for machine learning, effectively capturing the dynamic evolution law of the battery degradation process and enhancing the generalization ability and stability of the model. (3) The prediction module introduces an expansion characteristic sensitivity adjustment mechanism, which can adapt to different battery types and aging modes, achieve high-precision prediction of remaining service life and degradation trend, and meet the health management needs in complex usage scenarios. (4) The prediction results are fed back to the BMS control logic through the control module to achieve real-time optimization of the charging cut-off voltage, current limit and power strategy, improving energy efficiency while effectively extending battery life. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a schematic diagram of a battery life prediction and health assessment system based on BMS expansion parameter analysis provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0014] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0015] The first embodiment of the present application provides a battery life prediction and health assessment system based on BMS expansion parameter analysis. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 The first embodiment of the present application provides a battery life prediction and health assessment system based on BMS expansion parameter analysis, which is described in detail.

[0016] 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 assessment module 104 and a control module 105 .

[0017] The acquisition module 101 is used to collect the battery expansion characteristic data under different charge and discharge cycles based on the expansion sensor, modeling estimation or deformation variable inference mechanism connected to the BMS system, and simultaneously obtain conventional electrochemical operating parameters such as voltage, current, and temperature to construct a multi-dimensional time series operation data set.

[0018] Acquisition module 101 is used to collect multi-source operational data, particularly physical deformation information including expansion characteristics, from the battery management system (BMS) for subsequent health status analysis and lifespan prediction. Its primary functions are as follows: First, module 101 establishes a real-time connection with the BMS's data interface to retrieve basic electrochemical parameter data related to battery cell and battery pack operation, 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 acquired using conventional sensing devices such as current sensors, voltage detection circuits, and thermistors.

[0019] Building on this foundation, acquisition module 101 further incorporates the capability to collect battery expansion characteristics. Specifically, this module can acquire expansion characteristic data through two approaches: one is hardware-based expansion sensors. These sensors can be linear or planar deformation detectors based on capacitance, resistance, or fiber Bragg grating (FBG) principles. These sensors are installed in the battery housing or module gaps, and are used to collect real-time micro-deformations caused by factors such as internal gas generation, SEI film growth, or thermal expansion during charging and discharging. The other is based on modeling estimation or inference mechanisms. In scenarios where physical expansion sensors are not available, indirect expansion estimation is performed using existing operating data, such as temperature, current rate, and cell type, combined with laboratory-prepared expansion response curves and material models. These models can be implemented using multivariate linear regression, support vector regression (SVR), or neural network models. Their input is operating status data, and their output is the incremental or absolute value of expansion per unit time.

[0020] To ensure data consistency and alignment across time, acquisition module 101 incorporates a data synchronization control mechanism. This mechanism uses a sampling clock to uniformly schedule the acquisition cycles of various data sources, ensuring that expansion signature data and electrochemical parameters fully correspond at the sampling timestamp, facilitating time window reconstruction in subsequent processing modules. After data acquisition is complete, module 101 structures all collected data into a cache or directly transmits it to processing module 102, forming a running dataset containing multi-dimensional metrics and continuous time series. This dataset can be represented in a standard time series format, such as JSON, CSV, or a binary sequence, and includes fields such as timestamp, voltage, current, temperature, expansion, cell number, and sampling condition tags.

[0021] Furthermore, the acquisition module 101 supports configurations tailored to different battery configurations and BMS architectures, allowing users to preset acquisition frequency, expansion sensor sensitivity, redundant sensor fusion methods, and whether to enable the modeling acquisition path in inference mode. By decoupling hardware and software, the module can be deployed in centralized or distributed BMS environments, offering strong versatility and scalability, meeting the comprehensive battery health awareness requirements in a variety of application scenarios.

[0022] In summary, the acquisition module 101 not only breaks through the limitation of relying solely on electrochemical data in traditional BMS, but also builds a more comprehensive and physically meaningful battery operating status data foundation by introducing battery expansion characteristics as a new monitoring dimension, providing high-quality, high-dimensional data support for subsequent health status assessment and life prediction.

[0023] Processing module 102 is used to normalize, filter, and time-window the expansion feature data and electrochemical operating parameters, extract a composite feature vector representing the evolution trend of the battery health state, and combine the cycle index, load state, and ambient temperature labels to construct a health trajectory sample for learning.

[0024] Processing module 102 systematically processes the expansion characteristic data and electrochemical operating parameters acquired by acquisition module 101. The goal is to extract composite feature vectors from the raw multi-source time series data that accurately reflect the evolution of the battery's health status and further construct healthy trajectory samples for training and inference. This module first receives the dataset passed in by the acquisition module. The data structure typically includes multiple dimensions such as timestamp, voltage, current, temperature, expansion characteristic value, battery identification, and operating condition. To eliminate dimensional and scale differences among parameters and improve the convergence efficiency and accuracy of subsequent modeling, the processing module first normalizes all input parameters. Normalization can use min-max scaling or z-score normalization, and is customized based on battery type, sensor accuracy, and sampling resolution. For example, the voltage range is typically linearly scaled to the [0, 1] range, while expansion characteristics may be normalized based on absolute value and historical peak values.

[0025] After normalization, the module filters the data to remove transient outliers caused by sensor jitter, electromagnetic interference, or abnormal operating conditions. This filtering method can use a sliding mean filter, a median filter, or a low-pass filter. Specific parameters (such as the window size or cutoff frequency) are determined based on the sampling frequency and the target time scale. For example, at a 1Hz sampling frequency, a sliding window of length 5 effectively suppresses single-point glitches while still preserving degradation trends.

[0026] After normalization and filtering, the processing module reconstructs the data based on fixed-length time windows, dividing the continuous time series data into a number of overlapping or non-overlapping sample segments. Each segment contains a complete charge / discharge cycle or several key charging stages, which are used to form the input feature matrix. The data within each time window is organized into a two-dimensional tensor in chronological order, with one dimension representing the time step and the other representing different types of processed features, including expansion value, voltage, current, temperature, and their first-order and second-order differential derivatives, to enhance the model's ability to detect trend changes.

[0027] To improve the effectiveness of the sample in characterizing the battery state evolution path, the processing module also introduces cycle index, load state, and ambient temperature label as additional features. Among them, the cycle index indicates the cumulative number of cycles in the current sample, reflecting the battery aging process; the load state distinguishes special operating conditions such as high power and high rate charge and discharge by calculating the average current and power fluctuation rate; the ambient temperature label indicates the external ambient temperature distribution at the time of sampling, helping the model identify the impact of the thermal environment on expansion and degradation behavior. During the reconstruction process, these features are aligned with the main time series features and encoded into the same input tensor, forming a complete composite feature structure.

[0028] Finally, the processing module saves the processed samples as a standardized health trajectory sample set for supervised training or inference analysis by the prediction module 103. The health trajectory samples are dynamically updated using a sliding window, enabling the model to continuously learn the evolution of battery status and adapt to the time-varying environment and usage strategy changes in actual operation. The entire processing flow is implemented in software and can be embedded in the BMS main control firmware or deployed on an external diagnostic platform. It offers low computational complexity, high real-time performance, and high scalability, meeting the online prediction data requirements of various battery management scenarios.

[0029] Furthermore, the processing module is specifically configured to:

[0030] Based on the expansion characteristic data and electrochemical operating parameters, a short-time Fourier transform is performed on different frequency sub-bands to obtain a frequency band signal tensor. The cross-correlation coefficient between each frequency band and the expansion characteristic data is calculated in combination with the degree of cooperative fluctuation of the expansion characteristic data with the voltage and current to generate a corresponding frequency band channel weighting factor. The frequency band signal tensor and the frequency band channel weighting factor are weightedly fused channel by channel to obtain a frequency band channel weighted tensor.

[0031] Based on the frequency band channel weighted tensor, the difference between the maximum and minimum values ​​of each expansion feature dimension in adjacent time windows is extracted to construct a deformation residual vector, and the deformation residual vector is spliced ​​and fused with the electrochemical operating parameters in the time window to generate a composite residual code sequence for characterizing the degradation trend of the battery structure;

[0032] The composite residual coding sequence is used as the basic sample, and a sliding window mechanism is used for enhanced sampling. Based on the battery operation cycle density estimation strategy, the sampling window width and step size are adaptively adjusted to improve the sample coverage of the low-frequency degradation stage and generate a structural degradation sample set covering all stages of the life cycle;

[0033] The structural degradation sample set is input into the comparative encoding unit, and the structural degradation samples of the same battery cell at different discharge rates or different ambient temperatures are used as positive sample pairs to construct a variational contrast loss function. In addition, an invariance constraint is added during the encoding stage, so that the finally constructed healthy trajectory samples maintain the consistency and generalization ability of the structural degradation dimension under multiple working conditions.

[0034] In the battery life prediction and health assessment system based on BMS expansion parameter analysis described in this paper, the processing module not only normalizes, filters, and reconstructs the time window of the collected expansion feature data and electrochemical operating parameters, but also further undertakes the task of deeply constructing structural degradation trend characteristics and enhancing expression stability. To this end, the processing module incorporates a composite processing flow based on frequency domain decomposition, deformation coding, sample enhancement, and contrastive learning. The aim is to fully exploit the dynamic coupling characteristics between expansion and electrochemical signals and construct high-quality, uniformly distributed health trajectory samples for subsequent model training.

[0035] Specifically, in the initial stage, the system first performs a short-time Fourier transform on the expansion signature data and the voltage and current components of the electrochemical operating parameters. This decomposition results in time-frequency signals in multiple frequency subbands, forming a three-dimensional band signal tensor. This tensor's dimensions include a time window, a band index, and a characteristic channel. To further identify the sensitivity of each band to structural deformation, the system calculates the cross-correlation coefficient between each band signal and the expansion signature data within the same time window, using the cross-correlation coefficient as a channel weighting factor. This weighting factor reflects the adaptability of the band response to physical deformation and is dynamic, varying with different operating states. The system then performs a channel-by-channel fusion of the band signal tensor and the band-channel weighting factors. Specifically, the signal of each band channel is multiplied by its corresponding weighting factor to produce a band-channel weighted tensor. This enhances the response to critical deformation information while preserving the time-frequency structure.

[0036] Next, the processing module extracts the dynamic change characteristics of structural degradation from the weighted tensor of the frequency band channel. To this end, the system calculates the difference between the maximum and minimum values ​​of each expansion feature dimension in adjacent time windows to obtain a set of deformation residual vectors, which reflects the degree of strain mutation in the current cycle. In order to fuse multimodal information, the deformation residual vector is spliced ​​and combined with the electrochemical operating parameters in the same time window (including but not limited to normalized voltage, current and temperature), thereby generating a composite residual coding sequence with structural-electrochemical fusion properties to characterize the complex evolution state of the battery aging trend.

[0037] To address the uneven distribution of structural degradation samples and the scarcity of low-frequency samples across different lifecycle stages, the system further introduces a sliding window enhanced sampling mechanism. This mechanism uses the composite residual coding sequence as the underlying sample source. While maintaining the contextual coherence of the time series, it resamples the samples using a sliding window approach. The sampling window width and sliding step size are adaptively adjusted based on the battery's operating cycle density (e.g., the rate of expansion change per unit time, the frequency of charge and discharge cycles, etc.). For time intervals with lower degradation rates, the sampling window is appropriately narrowed and the sliding step size is slowed to increase the number of samples in that period. This results in a set of structural degradation samples covering the entire battery lifecycle. This sample set is then balanced to improve the representational integrity and class balance of the training samples.

[0038] Finally, to further enhance the model's consistency in representing structural degradation under a variety of complex operating conditions, the system inputs the aforementioned structural degradation sample set into a contrastive encoding unit to construct a variational contrastive training mechanism. In this mechanism, the system selects samples with the same lifespan stage from structural degradation samples collected from the same battery cell at different discharge rates or ambient temperatures as positive sample pairs, while randomly selecting samples from other stages or operating conditions as negative samples. By introducing a contrastive loss function with an invariance constraint, the encoder is guided to learn a healthy trajectory representation in the latent space that is stable and has strong generalization capabilities for the structural degradation dimension. This process not only improves the model's ability to robustly model structural evolution trends under real-world operating conditions but also provides semantically consistent and highly feature-integrated training input for subsequent prediction modules.

[0039] In summary, the processing module completes the process of constructing healthy trajectory samples that can be used for deep learning through four closely connected steps: frequency band weighting, deformation encoding, sample reconstruction, and contrast enhancement.

[0040] The following is the reference implementation code of the processing module:

[0041] import numpy as np

[0042] import torch

[0043] import torch.nn as nn

[0044] import torch.nn.functional as F

[0045] # Set global constants

[0046] EPSILON = 1e-8

[0047] # Assumption: Each sample is a tensor of shape (T, D), T is the time window length, D is the feature dimension

[0048] # The first N_feat dimensions are electrochemical parameters (such as voltage and current), and the last N_expansion dimensions are expansion features

[0049] def short_time_fourier_transform(data, window_size=64, hop_length=32):

[0050] """

[0051] Performs a short-time Fourier transform on each channel, returning a band tensor (number of samples, number of channels, time steps, number of bands)

[0052] """

[0053] # data: (B, T, D)

[0054] B, T, D = data.shape

[0055] fft_data = []

[0056] for d in range(D):

[0057] channel_data = data[:, :, d] # (B, T)

[0058] # Simplify processing: use abs(fft) to simulate STFT output

[0059] spec = torch.abs(torch.fft.rfft(channel_data, dim=1)) # (B,freq)

[0060] fft_data.append(spec.unsqueeze(1)) # Add channel dimension

[0061] fft_tensor = torch.cat(fft_data, dim=1) # (B, D, freq)

[0062] return fft_tensor # output shape (B, D, F)

[0063] def compute_correlation_weights(fft_tensor, expansion_data):

[0064] """

[0065] Calculate the cross-correlation coefficient between the frequency band of each channel and the expansion feature to obtain the channel weighting factor

[0066] fft_tensor: (B, D, F)

[0067] expansion_data: (B, D_exp)

[0068] Return weight coefficient: (B, D)

[0069] """

[0070] B, D, F = fft_tensor.shape

[0071] weights = []

[0072] for i in range(D):

[0073] fft_feat = fft_tensor[:, i, :] # (B, F)

[0074] exp_feat = expansion_data[:, i % expansion_data.shape[1]] # Assume D is aligned with the expanded feature dimension

[0075] # Calculate cross-correlation: Simplify to inner product normalization

[0076] corr = torch.sum(fft_feat exp_feat.unsqueeze(1), dim=1) / (torch.norm(fft_feat, dim=1) torch.norm(exp_feat, dim=0) + EPSILON)

[0077] weights.append(corr.unsqueeze(1)) # (B, 1)

[0078] weights = torch.cat(weights, dim=1) # (B, D)

[0079] return weights

[0080] def apply_frequency_weights(fft_tensor, weights):

[0081] """

[0082] Apply channel weighting factors to weight the frequency band data channel by channel

[0083] fft_tensor: (B, D, F)

[0084] weights: (B, D)

[0085] Output: (B, D, F)

[0086] """

[0087] return fft_tensor weights.unsqueeze(2)

[0088] def compute_strain_residual(encoded_tensor):

[0089] """

[0090] Calculate the deformation residual in adjacent time windows, that is, the difference between the maximum and minimum values

[0091] Assume the input is (B, D, T)

[0092] Returns: (B, D)

[0093] """

[0094] max_val = torch.max(encoded_tensor, dim=2)[0]

[0095] min_val = torch.min(encoded_tensor, dim=2)[0]

[0096] return max_val - min_val # difference tensor (B, D)

[0097] def sliding_window_augmentation(data, window_size=10, stride=5):

[0098] """

[0099] Sliding window enhanced sampling mechanism: suitable for processing sequence data

[0100] data: (T, D)

[0101] Returns: (N, window_size, D)

[0102] """

[0103] T, D = data.shape

[0104] samples = []

[0105] for start in range(0, T - window_size + 1, stride):

[0106] end = start + window_size

[0107] samples.append(data[start:end, :].unsqueeze(0)) # (1, window_size, D)

[0108] return torch.cat(samples, dim=0) # (N, window_size, D)

[0109] class ContrastiveEncoder(nn.Module):

[0110] def __init__(self, input_dim, hidden_dim):

[0111] super(ContrastiveEncoder, self).__init__()

[0112] self.encoder = nn.Sequential(

[0113] nn.Linear(input_dim, hidden_dim),

[0114] nn.ReLU(),

[0115] nn.Linear(hidden_dim, hidden_dim) )

[0117] def forward(self, x):

[0118] return self.encoder(x)

[0119] def variational_contrastive_loss(anchor, positive, negative, margin=1.0):

[0120] """

[0121] Variational contrast loss function: expresses structural invariance constraints through Euclidean distance in embedding space

[0122] anchor, positive, negative: (B, D)

[0123] """

[0124] pos_dist = F.pairwise_distance(anchor, positive)

[0125] neg_dist = F.pairwise_distance(anchor, negative)

[0126] loss = F.relu(pos_dist - neg_dist + margin)

[0127] return loss.mean()

[0128] # Example workflow:

[0129] # 1. STFT

[0130] # stft_data = short_time_fourier_transform(input_tensor)

[0131] # 2. Calculate collaborative weight

[0132] # weights = compute_correlation_weights(stft_data, expansion_features)

[0133] # 3. Apply weights

[0134] # fused_tensor = apply_frequency_weights(stft_data, weights)

[0135] # 4. Deformation Residual Coding

[0136] # residual = compute_strain_residual(fused_tensor.transpose(1, 2))

[0137] # 5. Electrochemical parameters of the splicing

[0138] # fused_representation = torch.cat([residual, electrochem_features],dim=1)

[0139] # 6. Enhanced sample generation

[0140] # augmented_samples = sliding_window_augmentation(full_sequence_tensor)

[0141] # 7. Comparison of training encoding

[0142] # anchor = encoder(augmented_samples[0])

[0143] # positive = encoder(augmented_samples[1])

[0144] # negative = encoder(augmented_samples[2])

[0145] # loss = variational_contrastive_loss(anchor, positive, negative)

[0146] # The complete training process combines all the above components and iterates the training steps.

[0147] The core workflow of this code is to gradually extract and construct a high-quality feature sequence that reflects the structural degradation trend of the battery from raw expansion feature data and electrochemical operating parameters. This feature sequence is then enhanced through comparative learning to generalize the model across multiple operating conditions. The system first performs a short-time Fourier transform on the expansion feature along with the voltage and current signals to extract multi-band time-frequency tensor signals. The code then calculates the degree of synergy between each frequency band and the expansion feature, deriving a set of weighting factors that reflect the importance of the frequency band. These factors are then used to emphasize expansion-sensitive channels within the frequency band. After weighted fusion, the system extracts the maximum and minimum differences between each expansion dimension within adjacent time windows from the frequency band signals to form a deformation residual. This residual is then concatenated with operating parameters such as voltage, current, and temperature within the corresponding window to generate a multimodal fused residual code sequence. To address the uneven distribution of samples across different lifecycle stages, the code uses a sliding window approach for sample augmentation. The sampling strategy is adaptively adjusted based on an estimate of the operating cycle density based on factors such as the expansion rate, thereby generating a sample set that covers the entire battery lifecycle. Finally, the system feeds these samples into a contrastive encoder network. By constructing positive and negative sample pairs and introducing a variational contrastive loss, the model learns, during training, a consistent representation of degraded trajectories across different operating conditions. This overall process forms a closed-loop processing chain from raw signals to semantically consistent feature representations, demonstrating structural, transferable, and engineering feasibility.

[0148] Furthermore, the sliding window mechanism is implemented by the following steps:

[0149] Based on the battery operation cycle density estimation strategy, the composite residual code sequence is divided into multiple life cycle stage subintervals. Within each life cycle stage subinterval, a density-aware hierarchical enhancement sampling operation is performed based on the variance of the change rate of the expanded feature dimension in the composite residual code sequence within a local time period to obtain a preliminary enhanced life cycle stage structure degradation sample set;

[0150] The initially enhanced life cycle stage structural degradation sample set is input into a context consistency filtering mechanism. The mechanism calculates, for each candidate sample window, the consistency of the direction of change of the expansion feature dimension between all time steps within the window and the consistency of the change trend of the electrochemical operating parameters. Based on the weighted results of the consistency of the direction of change of the expansion feature dimension and the consistency of the change trend, candidate windows with inconsistent internal structures or conflicting trends are screened out, and a context consistency sample set is output.

[0151] The context-consistent sample set is input into the sample retention scoring function, which performs multi-factor weighted fusion based on the deformation residual amplitude of the expanded feature dimension of each candidate sample, the power spectral density energy of the voltage and current signals, and the confidence of the life cycle stage to which the sample belongs. The retained samples are then screened based on the scoring threshold to form the final sample set used to construct the structural degradation sample set.

[0152] In the battery life prediction and health assessment system based on BMS expansion parameter analysis described in the present invention, the sliding window mechanism is not only used for time series sample generation in the traditional sense, but also achieves balanced coverage and high-quality construction of structural degradation sample sets in all stages of the life cycle by introducing period density, adaptive enhancement, context consistency judgment and sample quality scoring mechanisms.

[0153] First, the system segments the composite residual code sequence using a battery cycle density estimation strategy. This strategy estimates degradation densities at different stages based on historical behavioral data such as the battery's charge and discharge frequency, expansion rate changes, and structural response time. Based on this data, the system divides the composite residual code sequence into multiple lifecycle stage subintervals, such as the early stage, slow degradation stage, and accelerated degradation stage. Each subinterval represents a specific structural degradation process.

[0154] Next, within each lifecycle stage subinterval, the system performs a density-aware analysis of the variance of the rate of change of the inflated feature dimension within the composite residual coding sequence within a local time period. Regions with high rate-of-change variance are considered to have stronger degradation signal representation capabilities and are therefore suitable for sample augmentation. Based on this information, the system performs layered augmentation sampling, appropriately increasing the sample density at key degradation stages (such as structural mutation points and load perturbation response windows), resulting in a preliminary enhanced set of structural degradation samples for each lifecycle stage. This process ensures temporal continuity of the samples while precisely controlling information density.

[0155] Subsequently, the system inputs this preliminary enhanced sample set into the contextual consistency filtering mechanism for further screening. This mechanism uses the consistency of the direction of change of the expansion feature dimension between all time steps in the sample window as the main indicator, that is, to determine whether the direction of change of the deformation variable in the window remains stable in time. At the same time, it combines the trend consistency indicators of electrochemical operating parameters such as voltage and current to evaluate the degree of coordinated change between signals. The two consistency indicators are weighted fused to form a unified evaluation value, and then abnormal windows such as reversal of deformation mutation direction and inconsistent electrochemical trends in the sample set are eliminated, retaining only the contextual consistency sample set whose structural changes are consistent with the evolution of the electrochemical state.

[0156] Finally, the system inputs the above context-consistent sample set into the sample retention scoring function for quality assessment. The scoring function is based on a multidimensional factor design, including the magnitude of the maximum deformation residual amplitude in the expansion feature dimension, the power spectrum density energy distribution of the voltage and current signals in the current window, and the confidence index of the sample in the current life cycle stage (for example, 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 will include the sample in the final structural degradation sample set, otherwise it will be discarded. The scoring mechanism aims to ensure that the samples finally used for model training have the triple characteristics of structural degradation significance, sufficient signal expression and life cycle representativeness.

[0157] Through the phased enhanced sampling, contextual 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 expectancy prediction.

[0158] The prediction module 103 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 in 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.

[0159] Prediction module 103 is responsible for modeling, training, and inference based on the health trajectory samples output by processing module 102. Its goal is to output the battery's remaining useful life (RUL) and degradation trend score in its current state, taking into account expansion characteristics, electrochemical characteristics, and their evolutionary trends. This allows for more accurate and targeted health status predictions. This module first receives constructed health trajectory samples, typically organized as time series tensors in a sliding window format. Each sample consists of multiple time steps, and each time step records standardized composite features such as expansion amplitude, voltage, current, temperature, cycle index, and ambient temperature.

[0160] The prediction module's core prediction logic is built on a deep regression network model. This model can utilize structures with time series modeling capabilities, such as bidirectional long short-term memory (Bi-LSTM), gated recurrent units (GRU), temporal convolutional networks (TCN), or Transformer architectures. In its implementation, the model's input layer receives each health trajectory sample tensor while preserving its time series structure. It then extracts the evolutionary pattern within the time dimension through several layers of recurrent or convolutional units. This is then regressed through a fully connected layer to generate the corresponding sample's remaining useful life value and degradation trend score. The remaining useful life value indicates the number of cycles or time from the current state that normal performance is expected to remain. The degradation trend score is a floating-point number between 0 and 1, reflecting the relative position of the current health state within the entire useful life range. A value closer to 1 indicates a more severely degraded state.

[0161] In order to ensure that the prediction results can adapt to the differences in feature expression under different battery types, different structural designs or different usage environments, the prediction module introduces an expansion feature sensitivity adjustment mechanism during the training phase. The core idea of ​​this mechanism is to assign adaptive weights to the expansion class variables in the input features, and dynamically adjust the contribution of the expansion features to the training objectives in the loss function. Specifically, the module can set a set 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 channels that are more sensitive to expansion features based on the gradient size, feature activation frequency and prediction error feedback information, suppresses the weights of redundant channels or low signal-to-noise ratio features, and makes the model more focused on structural indicators that have degradation indicative 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 scenarios with data heterogeneity.

[0162] During the deployment phase, the prediction module possesses continuous learning capabilities, allowing it to update its model based on actual operational data. If the system detects that predictions deviate from actual performance over a long period of time, the module automatically triggers an online fine-tuning process. Using recent healthy trajectory samples and real-world lifespan labels (such as forced offline times and manual maintenance records) as new training data, the module fine-tunes parameters or performs partial retraining of the deployed model, thereby improving prediction accuracy and long-term stability.

[0163] The entire prediction module can be implemented using Python, C++, or other embedded frameworks, supporting deployment on cloud servers or local edge devices. In practical applications, this module communicates with the processing and evaluation modules via standard data format interfaces (such as NumPy arrays or TensorFlow tensors), enabling complete health status prediction updates within minutes or seconds, meeting the battery management system's requirements for real-time and high-precision lifespan prediction. Through this design, prediction module 103 ensures intelligent perception of battery degradation trends, addressing the prediction blind spots caused by traditional models' neglect of structural indicators and providing a clear, quantitative, and reliable reference for subsequent regulatory adjustments.

[0164] Furthermore, the prediction module introduces an expansion feature sensitivity adjustment mechanism during the training phase, and the expansion feature sensitivity adjustment mechanism includes the following steps:

[0165] When constructing the input tensor of the health trajectory sample, the dimension index containing the expanded feature is marked, and a set of trainable weight coefficient vectors consistent with the number of feature dimensions is initialized. The weight coefficient vectors are initially set to a uniform constant value;

[0166] The weight coefficient vector is introduced into the feature input layer of the prediction model, and the feature dimension of each input sample is weighted element by element to form a weighted input feature tensor;

[0167] During model training, the absolute value of the gradient and the activation frequency corresponding to the expanded feature dimension are continuously tracked, and the expanded feature importance score is calculated based on the back-propagation result of the loss function. The expanded feature importance score is updated based on the exponential moving average.

[0168] Dynamically updating the weight coefficient vector based on the inflated feature importance score, wherein dimensions with higher inflated feature importance scores are assigned higher weights, and dimensions with lower inflated feature importance scores are gradually attenuated, wherein the weight update process is implemented by gradient descent or attention normalization mechanism;

[0169] After each training cycle, the L2 regularization method is used to constrain the weight coefficient vector to prevent the abnormal amplification of the weight of individual dimensions and cause model instability;

[0170] The prediction module is gradually made more sensitive to expansion characteristics during multiple rounds of training, the modeling accuracy of structural degradation trends is improved, and the prediction generalization capability of the prediction module under different battery types and operating environments is enhanced.

[0171] In the battery life prediction and health assessment system based on BMS expansion parameter analysis, described in this paper, the prediction module incorporates an expansion feature sensitivity adjustment mechanism during the training phase. This mechanism aims to address the problem of the relative contribution of expansion features in multi-dimensional feature inputs being unclear, and being overwhelmed or misclassified as redundant information during model learning. This mechanism systematically improves the model's sensitivity to structural degradation-related features through a set of clear steps, from sample construction to model training, and then to weight updates and regularization control, ensuring the stability, adaptability, and scalability of the battery health prediction process.

[0172] During the training sample construction phase, the system first explicitly marks the dimension indices of all dilation class features in the input tensor. For example, for a The samples of the input dimension, if Dimensions corresponding to expansion parameters (such as battery shell strain, module gap deformation, estimated expansion, etc.), these indexes will be recorded in the internal mapping table and used as the target dimensions for subsequent adjustments. Input tensor dimensions Represents the total number of features of each sample, that is, the number of dimensions of each row of the input vector.

[0173] For example, the features collected by the system include voltage, current, temperature, battery shell strain, module gap change and expansion estimation, a total of 6 features, then .Inflated class feature dimension index Indicates here Among the features, which ones are related to battery expansion? For example, the 4th, 5th and 6th features are shell strain, module gap and expansion estimation, then , that is, the expanded feature dimension index is [4, 5, 6].

[0174] Then, the system initializes a length of The weight coefficient vector , the initial value is uniformly set to a constant ,like , to ensure that all features have equal contribution in the initial stage.

[0175] For example, for d = 6, the system is initialized as = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0], which means that each feature has an equal initial weight of 1.0. This vector is automatically adjusted based on the importance of each feature.

[0176] This weight coefficient vector will be introduced into the input processing flow of the prediction model as a feature weighting factor. (in is the time window length), the model will perform element-by-element weighting operations according to the following formula 1:

[0177] ;

[0178] in represents the column-broadcasted Hadamard product, is the weight vector for the current training round. This operation ensures that in each round of training, each feature dimension received by the model has been dynamically weighted, thus laying the foundation for subsequent backpropagation adjustments.

[0179] X is an input tensor representing the input feature data for a batch or a single sample. Each row is a feature vector at a specific time, containing T time steps and d features. It is constructed from multidimensional time series data (such as voltage, current, temperature, and expansion) collected from the BMS system after preprocessing (normalization, filtering, and window reconstruction). For example, if T = 10 (representing the past 10 time points) and d = 6 (including voltage, current, temperature, housing strain, gap change, and estimated expansion), then X is a 10×6 matrix.

[0180] is the weight vector for training round t, which represents the weighting factor assigned by the model to each feature dimension in the current training round, used to emphasize or weaken the role of a specific dimension in the prediction task. This vector is initialized to all 1s (or a uniform constant) and is adaptively updated during training based on the gradient, activation frequency, and feature importance score. For example, in round t of training, It could be:

[0181] = [0.8, 0.9, 1.0, 1.5, 1.3, 1.6];

[0182] This means that the model considers the 4th, 5th, and 6th dimensions (i.e., expansion features) to be more important and therefore gives them higher weights.

[0183] This is a weighted input tensor representing the weighted features of each row of time steps, forming the new input for subsequent model training. This is fed into the next layer of the model to increase the sensitivity of predictions to key features, such as dilation behavior.

[0184] During model training, the system continuously monitors the gradient size and activation frequency of each dilated feature dimension. The absolute value of the gradient reflects the direct influence of the feature in the loss function, and the activation frequency reflects its participation in nonlinear transformations (such as ReLU, Sigmoid). The system combines these two types of information to calculate the dilated feature importance score. , which is updated using the exponential moving average strategy:

[0185] ;

[0186] To expand the feature dimension In the training round The importance score of the current round In i The relative importance score of each feature (such as the dilation-related dimension) in model training.

[0187] Indicates the current round Middle The average gradient value of the dimension in the loss function reflects the direct influence of the feature on the loss function. It is automatically obtained during the back propagation phase of the model during training. For example, 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.

[0188] Indicates its activation frequency, which indicates the frequency at which the value of this feature dimension is not suppressed to zero after passing through the nonlinear transformation layer of the neural network (such as ReLU, Sigmoid), reflecting its "participation". It comes from the forward propagation, recording the proportion of the number of effective outputs of the feature in the nonlinear transformation. For example, if after the ReLU activation layer, the 5th feature has 80 non-zero values ​​out of 100, the activation frequency is = 80 / 100 = 0.8.

[0189] is a smoothing factor, such as 0.9, to control the smoothness of score updates. This score comprehensively considers participation and influence, and has clear physical meaning and feasibility. This scoring mechanism combines the statistically robust "historical performance" with the highly sensitive "current contribution," creating a dynamic weighting strategy that is physically plausible, statistically stable, and feasible during model training.

[0190] In each training round, the system dynamically adjusts the weights of the corresponding dimensions based on the inflated feature importance scores of all dimensions. The weights of high-scoring feature dimensions will be proportionally increased, while the weights of low-scoring dimensions will slowly decay. For example, the following update formula is used for weighted correction:

[0191] ;

[0192] For the Feature dimensions in rounds The weight of i The weight coefficient of each feature is initialized to a uniform constant (such as 1.0) and dynamically updated after each round of training.

[0193] After the update The weight of each dimension represents the new weight value to be used in the next training round, which is adaptively adjusted based on the current importance of the dimension.

[0194] is the learning rate hyperparameter, with a typical value of 0.01.

[0195] Represents the average importance score of all expanded feature dimensions, which comes from the importance score The arithmetic mean of the dilation feature dimensions. For example, if the 4th, 5th, and 6th dimensions are dilation features with importances of [0.10, 0.18, 0.12], then = (0.10 + 0.18 + 0.12) / 3 = 0.1333;

[0196] A very small constant to prevent division by zero. This update mechanism is adaptive, that is, if the importance score of an inflated feature is significantly higher than the average, its corresponding weight will be significantly enhanced, and vice versa.

[0197] The system will perform attention normalization on the weight coefficient vector after each update to ensure that its sum is constant to avoid affecting the numerical stability of the overall model.

[0198] To prevent individual feature dimensions from losing control of weight amplification 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 to not exceed a preset upper limit. The following normalization formula can be used:

[0199] ;

[0200] is the weight coefficient vector after the current training cycle, which represents the feature weight value vector after the current round of weight update (such as scoring based on feature importance).

[0201] in is a very small constant term to avoid division by zero errors. It is executed once after each round of training, and can also be completed together with the gradient update to control the computational overhead.

[0202] Through this mechanism, the system gradually focuses on expansion features that play a key role in predicting structural degradation over multiple rounds of training, shifting the model's learning focus from purely electrochemical features to physical deformation features that are more predictive of failures. This significantly improves the predictive model's adaptability, accuracy, and stability in complex scenarios. This mechanism is highly feasible and structurally distinct from traditional feature-weighted models. It does not rely on fixed channel structures or manual hyperparameter configuration, and can be deployed in a variety of model architectures, demonstrating excellent versatility and technological advancement.

[0203] The following is the implementation code of the expansion feature sensitivity adjustment mechanism based on PyTorch:

[0204] import torch

[0205] import torch.nn as nn

[0206] # Set parameters

[0207] feature_dim = 64

[0208] expansion_indices = [5, 12, 27, 41]

[0209] lambda_ema = 0.9

[0210] eta = 0.05

[0211] epsilon = 1e-6

[0212] delta = 1e-5

[0213] c_init = 1.0

[0214] # Simulate input data: batch_size × time_steps × feature_dim

[0215] batch_size, time_steps = 32, 10

[0216] X = torch.rand((batch_size, time_steps, feature_dim), requires_grad=True)

[0217] target = torch.rand((batch_size, 1))

[0218] # Initialize the weight coefficient vector to a constant c_init

[0219] w = nn.Parameter(torch.full((feature_dim,), c_init), requires_grad=True)

[0220] # Initialize the inflated feature importance score vector

[0221] S = torch.zeros(feature_dim)

[0222] # Define the prediction model

[0223] class SimpleRegressor(nn.Module):

[0224] def __init__(self, input_dim):

[0225] super().__init__()

[0226] self.fc = nn.Linear(input_dim, 1)

[0227] def forward(self, x):

[0228] x = x.mean(dim=1) # Time average pooling

[0229] return self.fc(x)

[0230] model = SimpleRegressor(feature_dim)

[0231] criterion = nn.MSELoss()

[0232] # Weighting function: Hadamard product (dimensional feature weighting)

[0233] def weighted_input(X, w):

[0234] return X w

[0235] # Training forward propagation phase

[0236] X_weighted = weighted_input(X, w)

[0237] output = model(X_weighted)

[0238] loss = criterion(output, target)

[0239] loss.backward()

[0240] # Inflate feature gradients and activation traces, and update importance scores and weights

[0241] with torch.no_grad():

[0242] grad = w.grad.abs()

[0243] activation = X_weighted.abs().mean(dim=(0, 1))

[0244] # Update S_i^(t) = λ for each dilated feature S_i^(t-1) + (1 - λ) ( A_i)

[0245] for i in expansion_indices:

[0246] S[i] = lambda_ema S[i]+ (1 - lambda_ema) (grad[i] activation[i])

[0247] # Average score of all inflated features in the current round

[0248] S_mean = S[expansion_indices].mean() + epsilon

[0249] # Weight update: w_i = w_i (1 + η (S_i - ) / ( + ε))

[0250] for i in expansion_indices:

[0251] correction = 1.0 + eta ((S[i] - S_mean) / S_mean)

[0252] w[i] = w[i] correction

[0253] # Attention normalization (constant sum, mean 1)

[0254] w_sum = w.sum() + epsilon

[0255] w[:] = w / w_sum feature_dim

[0256] # L2 regularization: w = w / ( _2 + )

[0257] norm = torch.norm(w) + delta

[0258] w[:] = w / norm

[0259] # Gradient clear

[0260] model.zero_grad()

[0261] w.grad.zero_()

[0262] The above code implements a sensitivity adjustment mechanism for dilation features during training. The overall process consists of the following steps. First, the system extracts time series data from the input feature tensor and weights each feature dimension element-wise using a set of trainable weight vectors to emphasize the influence of dilation-related features. The weighted features are then fed into a simple regression model for forward propagation and loss calculation. During the backpropagation phase, the system updates the importance score of each dilated feature dimension based on its gradient magnitude and activation level. The corresponding weight is then dynamically adjusted accordingly to improve responsiveness to key degradation signals. The updated weight vectors are then normalized and regularized to maintain model numerical stability and prevent local overfitting. This entire process is automatically performed in each round of training, allowing the model to gradually improve its adaptability to the dilation parameters and predictive accuracy.

[0263] Furthermore, the prediction module includes a distribution invariance self-calibration mechanism, and the distribution invariance self-calibration mechanism includes the following steps:

[0264] Based on the health trajectory samples, the distribution mean and distribution variance of the expansion feature dimension on different battery cell samples in each training batch are calculated to obtain a corresponding batch-level distribution deviation index. The batch-level distribution deviation index is used to quantify the statistical distribution difference of the expansion feature dimension between different battery batches;

[0265] Based on the batch-level distribution deviation indicator, a feature calibration vector is generated that corresponds one-to-one to the inflated feature dimension, and the feature calibration vector and the healthy trajectory sample are subjected to dimension-by-dimension weighted compensation processing to form a calibration feature tensor that has been corrected for batch distribution;

[0266] Inputting the calibration feature tensor into a multi-head attention channel selection network, the multi-head attention channel selection network assigns corresponding attention weights to each dilated feature dimension in the calibration feature tensor and outputs a channel-weighted embedding representation with channel selection capability;

[0267] The channel weighted embedding representation is input into a recurrent neural network branch and a Transformer network branch, respectively, wherein the recurrent neural network branch is used to model the historical state dependency of the channel weighted embedding representation, and the Transformer network branch is used to model the non-local interaction relationship between the expanded feature dimension and the charging rate in the channel weighted embedding representation. A structural prior guided term is introduced into the loss function, and the known theoretical functional relationship between the expanded 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.

[0268] In the battery life prediction and health assessment system based on BMS expansion parameter analysis described in this paper, the distribution invariance self-calibration mechanism within the prediction module is a key component for improving the model's cross-batch generalization and physical consistency. This mechanism primarily addresses the issue of statistical distribution shifts in expansion characteristics caused by different battery batches, models, or production processes in practical applications, ensuring that the model maintains prediction accuracy and stability despite multi-source data input.

[0269] First, during the model training process, the system continuously collects health trajectory data for all samples in the current batch, and calculates the distribution mean and distribution variance of each dimension in the training batch for all expansion feature dimensions. This calculation is completed based on the sample set in the entire training batch. The samples come from different battery cells and have a certain degree of statistical heterogeneity. For each expansion feature dimension, the system will also refer to the reference distribution or prior expectation (such as the previous batch or historical average statistics) to measure its distribution drift. On this basis, the system introduces a batch-level distribution deviation indicator as a standardized measure to quantify the degree of change in the distribution of the current expansion feature dimension. This indicator can be calculated using a variant of KL divergence, Wasserstein distance or Z-score difference, and is used to characterize the degree of statistical deviation between the expansion behavior in the current sample and the reference distribution.

[0270] Next, the system generates a feature calibration vector that corresponds one-to-one to the inflated feature dimension based on the batch-level distribution deviation indicator. The feature calibration vector usually exists in the form of a vector, and its number of dimensions is consistent with the number of inflated feature dimensions, and each element indicates how the corresponding dimension should be compensated. Feature calibration can be a combination of scaling, shifting, or exponential transformation to minimize the distribution differences caused by batches. The system will directly apply the feature calibration vector to the original healthy trajectory sample, and in the process of performing dimension-by-dimension weighted compensation, apply the corresponding correction parameters to each inflated feature dimension to obtain a calibration feature tensor that has been adaptively corrected for the batch distribution. This tensor not only retains the temporal structure of the original sample, but also has cross-batch distribution consistency, which can provide a more standardized input representation for the subsequent modeling stage.

[0271] After the correction is completed, the system inputs the calibrated feature tensor into a structure designed as a multi-head attention channel selection network. The core goal of this network is to selectively enhance the important parts of the expanded feature dimensions. In each attention head, the network dynamically assigns attention weights based on the representation strength and stability of each expanded feature dimension in the current sample and its correlation with the prediction target (such as lifespan score). The introduction of multiple attention heads can comprehensively evaluate the importance of each channel from different perspectives, so that the final output channel weighted embedding representation has a high degree of information focus, that is, the system only strengthens the dimensions that are indicative of lifespan degradation trends, while weakening the influence of redundant or unstable dimensions.

[0272] In order to further improve the model's performance in modeling the relationship between time series and structural physics, the system inputs the above-mentioned channel weighted embedding representation in parallel into two complementary sub-networks: a recurrent neural network branch and a transformer network branch. Among them, the recurrent neural network branch is mainly used to capture the dependency information of the battery in the historical charge and discharge cycles, such as trend fluctuations in time series, periodic expansion patterns, etc.; while the transformer network branch focuses on modeling the non-local interaction between the expansion feature dimension and the charging rate, especially when there are potential nonlinear responses between different dimensions. The context modeling ability is particularly outstanding. The two branches extract the deep representation of temporal dependency and structural coupling respectively, and then merge them to form a complete representation vector to output the remaining service life and degradation trend score.

[0273] Notably, during this joint modeling process, the system introduces a structural prior guidance term as a regularizer in the loss function to improve the physical consistency of the model. This guidance term is based on the functional relationship between the expansion characteristic dimension and the charge rate known from experimental or theoretical knowledge (for example, the shell strain increases linearly or exponentially with increasing charge rate). It is embedded in the loss function in the form of squared differences, KL divergence, or fitting residuals, thereby imposing structural constraints on the model predictions, ensuring that they conform to the battery's physical mechanisms while regressing the lifespan indicators.

[0274] Through this multi-step collaborative mechanism, the prediction module can maintain high modeling stability, prediction consistency, and cross-cell universality even when there is statistical deviation in the input health trajectory samples, meeting the actual needs of battery life and health status prediction in complex industrial application environments.

[0275] The evaluation module 104 is used to perform a hierarchical analysis of the degradation trend score and the remaining service life, and to determine the life stage and health level of the battery cells and battery clusters in combination with the set health status threshold and historical performance, and to generate a visual health assessment map and a strategy recommendation report.

[0276] Evaluation module 104 is used to further analyze the degradation trend score and remaining useful life information output by prediction module 103, and conduct further logical analysis and status determination. This allows the establishment of a clear and operational health rating system to support operational safety assessments and maintenance strategy formulation for the battery system. This module's core task is to develop a hierarchical and labeled health status evaluation mechanism based on numerical prediction results, combined with pre-set rules and historical data. This ensures that the system possesses both predictive and operational health assessment capabilities.

[0277] During the specific implementation process, the evaluation module first receives two core indicators from the prediction module, namely the degradation trend score and the remaining service life. The degradation trend score is usually a floating point number between 0 and 1, which indicates the relative degree of degradation of the current state in the entire life cycle. The remaining service life is a numerical value in units of time, number of cycles or energy output, which represents the sustainable use capacity of the battery in its current state. The evaluation module performs a grading process 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 dividing points for state changes. Different scoring intervals will correspond to different health levels, such as "excellent", "good", "medium", "poor", "critical" and other labels. Such thresholds can be obtained through large-scale life test data or historical operation data statistics, or they can be customized according to specific application scenarios and safety standards.

[0278] To prevent outliers at a single moment from misleading the assessment results, the assessment module uses sliding window averaging or weighted time series analysis to smooth the input degradation score, ensuring the stability and anti-interference of the health level determination. Furthermore, the module also introduces a historical performance comparison mechanism. For example, it compares the current battery assessment results with the health level change trend over several cycles to determine whether there is a rapid decline in health level, nonlinear degradation, or abnormal fluctuations, thereby assisting in identifying potential safety risks.

[0279] In addition to the battery cell-level assessment, the module further performs an aggregated assessment of the battery cluster or the entire system. This process includes weighted aggregation, standardization, or deviation analysis of the health level results of multiple battery cells, identifying cells with abnormal performance or degradation rates significantly faster than the group average, and generating a battery cluster-level health level label. This processing is particularly critical for systems with parallel structures, series structures, or battery pack levels, and can identify internal consistency issues, thermal runaway risks, or local performance bottlenecks.

[0280] In order to facilitate system integration and operation and maintenance personnel to intuitively grasp the evaluation results, the evaluation module will visualize the analysis results in the form of a graph. The graph may include a health trend curve, 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 are interactive and exportable. In addition, the module will also output a strategy recommendation report based on the current health level and historical trends of the battery, including specific operational suggestions such as whether to recommend derating operation, whether early warning maintenance is required, and whether to recommend replacing specific battery cells, to support BMS or upper control systems to make response decisions.

[0281] Overall, the evaluation module not only has numerical analysis capabilities, but also integrates the establishment of a health level system, the introduction of anomaly identification mechanisms, dynamic analysis of historical trends, and the generation of human-computer interaction maps into its functional logic. It constitutes the evaluation core of the intelligent health management of the battery system and ensures that the system has perceptible, explainable, and executable status diagnosis capabilities.

[0282] The control module 105 is configured to receive the remaining service life and degradation trend score, dynamically adjust BMS control parameters, including but not limited to charging cut-off voltage, current limit, and power distribution strategy, and optimize the energy efficiency and protection balance of the control strategy based on a feedback learning mechanism.

[0283] The control module 105 converts the degradation trend score and remaining useful life prediction results provided by the evaluation module 104 into control strategy adjustments that can directly affect the battery management system (BMS), thereby achieving 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 continuously optimizes the control strategy through a feedback learning mechanism, forming a closed-loop adaptive control system driven by prediction.

[0284] In actual applications, the control module first receives the result data from the evaluation module, including the remaining service life and current degradation trend score of each battery cell or battery cluster. These results are interpreted and processed by the strategy mapping logic within the module. Specifically, the system presets a set of charge and discharge parameter configuration ranges for different health levels. For example, when a battery cell is in an "excellent" or "good" state, a higher charge cut-off voltage and charge current are allowed to maintain a higher energy density output; when the battery health level drops to "medium" or "poor", the allowable charge cut-off voltage and discharge depth are automatically reduced, and the maximum current limit is limited, thereby slowing its further degradation and extending its usable life.

[0285] To avoid sudden changes in strategy adjustments, the control module introduces slope limits and slow-change logic to the change process of the control parameters. For example, when the life prediction results show a rapid downward trend, the module will not immediately implement a large-scale parameter reduction, but will make restrictive adjustments in stages so that the control changes are gradually completed within multiple sampling cycles, ensuring the stability of battery operation and the consistency of the system user experience. In addition, to improve the adaptability and personalization of the control strategy, the control module has constructed a control strategy optimization subsystem based on a feedback learning mechanism. This subsystem records the state response before and after each control parameter adjustment, including the actual operating voltage, battery temperature rise, SOC change rate and the output results of the new round of evaluation, and uses this as a feedback signal to construct a reinforcement learning or online fine-tuning model to continuously optimize the strategy mapping relationship so that the control logic continues to approach the optimal solution in actual operation.

[0286] The control output interface of the regulation module is connected to the BMS main control unit through standard protocols, such as CAN, RS485 or SPI interfaces, to ensure that parameter update instructions can be issued in milliseconds and applied to the execution layer. At the same time, the module supports collaborative control of multiple cells. When cells with large differences in health status are detected in the battery cluster, group current limiting, energy balancing or dynamic load adjustment mechanisms can be used to maintain overall performance while reducing local risks. In the energy management system (EMS) or vehicle control system (VCU), the module can also provide status synchronization signals for adjusting the global power scheduling strategy or operating mode selection, such as entering energy-saving mode or standby operation when the health status decreases.

[0287] The control module offers excellent configurability, allowing operators or system developers to customize various control strategy mapping rules, parameter adjustment steps, safety limit boundaries, and learning rate parameters. It also supports remote iteration and upgrades of strategy logic via OTA updates. In actual deployments, this module can be integrated into embedded BMS firmware or run as upper-layer energy management control logic on a remote monitoring server or edge control platform.

[0288] In summary, the control module 105 not only completes the dynamic mapping from health prediction results to control behaviors, but also, through the feedback learning mechanism and soft limit adjustment strategy, enables the battery to maintain a safe, stable and efficient control range in various operating scenarios, ensuring that the system has continuous operation capabilities and intelligent adaptive performance.

[0289] Furthermore, the control module is specifically used to:

[0290] Constructing a residual service life interval mapping table based on the remaining service life, and dynamically adjusting the charging cut-off voltage according to the interval to which the current remaining service life belongs;

[0291] Inputting the adjusted charge cut-off voltage and the degradation trend score into a lifecycle adaptive current limiting strategy evolution model to generate a corresponding current limiting control curve;

[0292] Based on the historical maximum expansion amplitude change rate recorded by the processing module, an envelope function is constructed to correct the current limiting control curve and output a power regulation strategy with a limiting constraint. When a sudden change occurs in the expansion characteristic dimension score, the activation order of the protection subunits in the BMS control structure is adjusted according to the power regulation strategy, and the current reduction, charging suppression and heat dissipation control programs are preferentially started.

[0293] In the battery life prediction and health assessment system based on BMS expansion parameter analysis described in this paper, the control module performs the key function of dynamically adjusting BMS control parameters based on prediction results. This includes intelligently modifying core parameters such as charge cutoff voltage, current limit, and power allocation strategy to achieve comprehensive optimization of battery safety, efficiency, and lifespan. This module is based on the remaining useful life and degradation trend scores output by the system prediction module, combined with historical expansion change data. Through multiple highly coupled steps, this module achieves adaptive evolution of the control strategy and phased response.

[0294] First, after the prediction module outputs the battery's current remaining service life, the system uses this life value as an index to consult a pre-built residual service life interval mapping table. This mapping table divides the battery's remaining service life into several intervals, each corresponding to a corresponding target charge cutoff voltage value. For example, if the prediction results indicate that a battery's remaining service life is between 30% and 40%, the system automatically selects the charge cutoff voltage value corresponding to this interval, such as 4.05V, to replace the initial standard value of 4.2V, thereby reducing the battery's operating stress and delaying degradation.

[0295] The system then inputs this dynamically adjusted charge cutoff voltage, along with the degradation trend score from the same assessment cycle, into a lifecycle adaptive current-limiting strategy evolution model. This model adaptively generates a current-limiting control curve based on the severity of structural degradation. This curve sets upper limits for charge or discharge current at different stages of the remaining lifespan, such as reducing it to 0.5C for cells with high degradation scores and maintaining it above 1C for cells with low degradation scores, ensuring they operate at a relatively conservative current level given their current health.

[0296] After obtaining the preliminary current-limiting curve, the system further introduces a dynamic feedback mechanism at the structural level. This mechanism uses the historical maximum rate of change of expansion amplitude recorded by the processing module as a deformation fluctuation indicator to construct an envelope function reflecting the severity of structural stress. This envelope function typically uses a weighted exponential smoothing form of a sliding window maximum rate of change sequence to represent the physical expansion stability of the battery during operation. If the rate of change in the current cycle exceeds the historical steady-state range, the envelope function will lower the current threshold in the current-limiting curve, further enhancing protection, thereby generating a power control strategy with amplitude limiting constraints to reduce the risk of structural stress-induced failures.

[0297] Finally, while executing the power regulation strategy, the system also monitors the fluctuation behavior of the expansion characteristic dimension score. If this score is found to rise 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 sub-unit in the BMS control structure according to the preset strategy. At this time, modules such as current reduction, charge suppression, and heat dissipation control are prioritized for activation, thereby achieving a timely and proactive protection response. For example, in the stage where the structural degradation score is high and power fluctuations are limited, the system will prohibit further charging and force the activation of the heat dissipation module to stabilize the system state and delay further degradation.

[0298] In summary, the control module realizes the optimization of multi-dimensional control strategies driven by structural degradation through logical deconstruction and continuous reaction of the prediction results. It has the advantages of fine granularity, high responsiveness and high adaptability, and provides the system with closed-loop health management and life extension control capabilities.

[0299] The lifecycle adaptive current-limiting strategy evolution model consists of three main components: identifying strategy phases, generating current-limiting curves, and dynamically evolving the strategy. The model uses the remaining useful life and degradation trend scores collected and predicted by the system as its core inputs, supplemented by selected historical operating indicators, to drive the adaptive adjustment of charge and discharge control parameters.

[0300] During the strategy phase identification, the system first consults a lifespan interval mapping table based on the current remaining service life value. This mapping table is pre-divided into multiple stages based on the battery's entire life cycle, such as the initial stage, stable stage, slow degradation stage, and accelerated decay stage. Each interval is statistically analyzed using a large amount of historical data to derive a typical characteristic range and corresponds to a set of baseline current limiting strategy parameters. In actual applications, the system can use linear interpolation or fuzzy matching to map continuous lifespan values ​​to the closest interval label, thereby identifying the current battery lifespan stage. This stage label will serve as one of the inputs to the subsequent strategy generation module.

[0301] The current limiting curve generation part uses the above-mentioned life stage label and degradation trend score as joint inputs, combined with auxiliary features such as the operating ambient temperature, battery internal resistance change trend, and voltage fluctuation amplitude, and calls the strategy template library to generate the current limiting control curve. The template library stores multiple predefined strategy functions, including a power limiting model based on a piecewise linear function, a current control model based on an exponential decay curve, or a response mechanism based on rule reasoning. In implementation, the system can use a conditional weighting method to weight and combine multiple strategy functions according to the importance of the life stage and degradation score, thereby forming a current limiting output curve covering the current operating conditions. For example, when the degradation score is high, the system prioritizes activating the template that suppresses rapid charging and discharging; when the temperature is high or the internal resistance increases, the current limiting amplitude is appropriately increased to balance the thermal management requirements.

[0302] The dynamic evolution of the strategy is used to adaptively update model parameters during long-term operation. This module continuously tracks the deviation between actual power output and the expected current-limiting strategy, and dynamically adjusts key parameters in the strategy function based on the trend of the maximum rate of change of the expansion feature dimension. For example, when the system detects a sharp increase in the rate of change of the expansion amplitude in a short period of time, and the current strategy is not sufficiently suppressing power output, the system will adjust the upper limit of the current-limiting curve based on the principle of the envelope function. This envelope function can be designed as a nonlinear response function based on the maximum rate of change of expansion over the most recent several cycles, for example, using a Gaussian function or a piecewise cosine function, to smoothly suppress control output that exceeds the risk threshold within the power regulation range. The final corrected current-limiting curve will serve as the instruction basis for the control module, affecting core strategies such as charging cut-off voltage, current limit, and power distribution in the BMS control logic.

[0303] To ensure the stability and physical consistency of policy evolution, the system also implements a policy drift detection mechanism in each update round. This mechanism is used to prevent short-term noise or abnormal samples from causing drastic fluctuations in the current limiting policy. This mechanism can be implemented through standard deviation constraints within the sliding window, policy curve smoothness constraints, or structural prior knowledge matching rules, thus achieving a balance between stability and responsiveness.

[0304] Through the continuous processing of the above three links, the life cycle adaptive current limiting strategy 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 charging and discharging strategy, and provide the BMS with a control strategy with actual physical constraint significance, thereby enhancing the controllability of battery life and the safety of system operation.

[0305] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A battery life prediction and health assessment system based on BMS expansion parameter analysis, characterized in that: include: The acquisition module is used to collect the battery's expansion characteristic data under different charge and discharge cycles based on the expansion sensor, modeling estimation, or deformation inference mechanism connected to the BMS system, and simultaneously obtain conventional electrochemical operating parameters such as voltage, current, and temperature to construct a multi-dimensional time-series operation data set; a processing module for normalizing, filtering, and time-windowing the expansion feature data and electrochemical operating parameters, extracting a composite feature vector representing the evolution trend of the battery health state, and constructing a health trajectory sample for learning by combining the cycle index, load state, and ambient temperature labels; A prediction module 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 in 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 assessment module is used to perform hierarchical analysis of the degradation trend score and remaining service life, and to determine the life stage and health level of battery cells and battery clusters based on the set health status threshold and historical performance, and to generate a visual health assessment map and strategy recommendation report; The control module is used to receive the remaining service life and 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 a feedback learning mechanism.

2. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 1 is characterized in that: The prediction module introduces an expansion feature sensitivity adjustment mechanism during the training phase. The expansion feature sensitivity adjustment mechanism includes the following steps: When constructing the input tensor of the health trajectory sample, the dimension index containing the expanded feature is marked, and a set of trainable weight coefficient vectors consistent with the number of feature dimensions is initialized. The weight coefficient vectors are initially set to a uniform constant value; The weight coefficient vector is introduced into the feature input layer of the prediction model, and the feature dimension of each input sample is weighted element by element to form a weighted input feature tensor; During model training, the absolute value of the gradient and activation frequency corresponding to the expanded feature dimension are continuously tracked, and the backpropagation result of the loss function is combined to calculate the expanded feature importance score, which is updated based on the exponential moving average. Dynamically updating the weight coefficient vector based on the inflated feature importance score, wherein dimensions with higher inflated feature importance scores are assigned higher weights, and dimensions with lower inflated feature importance scores are gradually attenuated, wherein the weight update process is implemented by gradient descent or attention normalization mechanism; After each training cycle, the L2 regularization method is used to constrain the weight coefficient vector to prevent the abnormal amplification of the weight of individual dimensions and cause model instability.

3. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 1, characterized in that: The processing module is specifically used for: Based on the expansion characteristic data and electrochemical operating parameters, a short-time Fourier transform is performed on different frequency sub-bands to obtain a frequency band signal tensor. The cross-correlation coefficient between each frequency band and the expansion characteristic data is calculated in combination with the degree of cooperative fluctuation of the expansion characteristic data with the voltage and current to generate a corresponding frequency band channel weighting factor. The frequency band signal tensor and the frequency band channel weighting factor are weightedly fused channel by channel to obtain a frequency band channel weighted tensor. Based on the frequency band channel weighted tensor, the difference between the maximum and minimum values ​​of each expansion feature dimension in adjacent time windows is extracted to construct a deformation residual vector, and the deformation residual vector is spliced ​​and fused with the electrochemical operating parameters in the time window to generate a composite residual code sequence for characterizing the degradation trend of the battery structure; The composite residual coding sequence is used as the basic sample, and a sliding window mechanism is used for enhanced sampling. Based on the battery operation cycle density estimation strategy, the sampling window width and step size are adaptively adjusted to improve the sample coverage of the low-frequency degradation stage and generate a structural degradation sample set covering all stages of the life cycle; The structural degradation sample set is input into the comparative encoding unit, and the structural degradation samples of the same battery cell at different discharge rates or different ambient temperatures are used as positive sample pairs to construct a variational contrast loss function. In addition, an invariance constraint is added during the encoding stage, so that the finally constructed healthy trajectory samples maintain the consistency and generalization ability of the structural degradation dimension under multiple working conditions.

4. The battery life prediction and health assessment system based on BMS expansion parameter analysis according to claim 3 is characterized in that: The sliding window mechanism is implemented by the following steps: Based on the battery operation cycle density estimation strategy, the composite residual code sequence is divided into multiple life cycle stage subintervals. Within each life cycle stage subinterval, a density-aware hierarchical enhancement sampling operation is performed based on the variance of the change rate of the expanded feature dimension in the composite residual code sequence within a local time period to obtain a preliminary enhanced life cycle stage structure degradation sample set; The initially enhanced life cycle stage structural degradation sample set is input into a context consistency filtering mechanism. The mechanism calculates, for each candidate sample window, the consistency of the direction of change of the expansion feature dimension between all time steps within the window and the consistency of the change trend of the electrochemical operating parameters. Based on the weighted results of the consistency of the direction of change of the expansion feature dimension and the consistency of the change trend, candidate windows with inconsistent internal structures or conflicting trends are screened out, and a context consistency sample set is output. The context-consistent sample set is input into the sample retention scoring function, which performs multi-factor weighted fusion based on the deformation residual amplitude of the expanded feature dimension of each candidate sample, the power spectral density energy of the voltage and current signals, and the confidence of the life cycle stage to which the sample belongs. The retained samples are then screened based on the scoring threshold to form the final sample set used to construct the structural 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, which includes the following steps: Based on the health trajectory samples, the distribution mean and distribution variance of the expansion feature dimension on different battery cell samples in each training batch are calculated to obtain a corresponding batch-level distribution deviation index. The batch-level distribution deviation index is used to quantify the statistical distribution difference of the expansion feature dimension between different battery batches; Based on the batch-level distribution deviation indicator, a feature calibration vector is generated that corresponds one-to-one to the inflated feature dimension, and the feature calibration vector and the healthy trajectory sample are subjected to dimension-by-dimension weighted compensation processing to form a calibration feature tensor that has been corrected for batch distribution; Inputting the calibration feature tensor into a multi-head attention channel selection network, the multi-head attention channel selection network assigns corresponding attention weights to each dilated feature dimension in the calibration feature tensor and outputs a channel-weighted embedding representation with channel selection capability; The channel weighted embedding representation is input into a recurrent neural network branch and a Transformer network branch, respectively, wherein the recurrent neural network branch is used to model the historical state dependency of the channel weighted embedding representation, and the Transformer network branch is used to model the non-local interaction relationship between the expanded feature dimension and the charging rate in the channel weighted embedding representation. A structural prior guided term is introduced into the loss function, and the known theoretical functional relationship between the expanded 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 control module is specifically used for: Constructing a residual service life interval mapping table based on the remaining service life, and dynamically adjusting the charging cut-off voltage according to the interval to which the current remaining service life belongs; Inputting the adjusted charge cut-off voltage and the degradation trend score into a lifecycle adaptive current limiting strategy evolution model to generate a corresponding current limiting control curve; Based on the historical maximum expansion amplitude change rate recorded by the processing module, an envelope function is constructed to correct the current limiting control curve, and a power regulation strategy with a limiting constraint is output; When a sudden change occurs in the expansion feature dimension score, the activation order of the protection subunits in the BMS control structure is adjusted according to the power regulation strategy, and the current reduction, charging suppression and heat dissipation control programs are started first.

Images