Battery health state or remaining service life prediction method and system and storage medium

Abstract

The invention discloses a battery health state or remaining service life prediction method and system and a storage medium, and belongs to the field of battery health management prediction. The battery health state or remaining service life prediction method comprises the steps of obtaining a battery life data set and performing preprocessing to obtain a multi-source fusion feature data set; performing correlation evaluation and weighting on the multi-source fused feature data set through a correlation weight generation unit to obtain a weighted feature sequence; an MSP-Mama model is constructed, the weighted feature sequence serves as the input of the MSP-Mama model, and the MSP-Mama model comprises a multi-scale feature extraction module, a selective state space module and a learnable multi-scale fusion module; and predicting the state of health or the remaining service life of the battery based on an MSP-Mamba model. According to the invention, the accuracy, reliability and stability of battery health state or remaining service life prediction can be improved.

Description

Methods, systems, and storage media for predicting battery health or remaining lifespan Technical Field

[0001] This application belongs to the field of battery health management prediction, specifically relating to a method, system, and storage medium for predicting battery health status or remaining service life. Background Technology

[0002] With the rapid development of electric vehicles, energy storage systems, and other fields, battery health management (BHM) and state of health (SOH) prediction have become key technical challenges in battery management systems. During long-term operation and multi-condition cycling, batteries experience capacity decay and performance degradation due to factors such as material aging, accumulation of side reactions, and temperature fluctuations. Accurately predicting the battery's state of health (SOH) is crucial for ensuring system safety, improving energy efficiency, and reducing operation and maintenance costs.

[0003] Among related technologies, battery health state prediction methods are mainly divided into two categories: (1) Pure data-driven methods: such as using deep learning models such as LSTM, GRU, and TCN. These methods learn degradation trends through historical time series data, but ignore the electrochemical degradation mechanism, resulting in insufficient model interpretability and weak generalization ability across batteries. (2) Mechanism and data hybrid methods: Some studies attempt to introduce battery capacity degradation models or aging equations into neural network structures, but usually lack systematic feature fusion and scale adaptive modeling mechanisms.

[0004] Therefore, improving the accuracy, reliability, and stability of battery health status prediction methods has become a pressing technical problem in this field. Summary of the Invention

[0005] In view of the aforementioned problems, the present invention aims to at least partially solve one of the technical problems in the related art. To this end, the present invention provides a method, system, and storage medium for predicting battery health status or remaining lifespan, which can improve the accuracy, reliability, and stability of battery health status or remaining lifespan prediction.

[0006] To address the aforementioned technical problems, this application provides the following: According to a first aspect of this application, an embodiment of this application provides a method for predicting battery health status or remaining service life. The method includes: acquiring a battery life dataset and preprocessing it to obtain a multi-source fused feature dataset; evaluating and weighting the multi-source fused feature dataset using a correlation weight generation unit to obtain a weighted feature sequence; constructing an MSP-Mamba model, wherein the weighted feature sequence serves as the input to the MSP-Mamba model, and the MSP-Mamba model includes a multi-scale feature extraction module, a selective state space module, and a learnable multi-scale fusion module; wherein the multi-scale feature extraction module is used to extract multi-timescale degradation representations during battery degradation through multi-scale convolutional branches; the selective state space module, based on a selective state space reasoning mechanism, is used to perform temporal reasoning and dynamic memory modeling on the multi-timescale degradation representations to form multi-scale state representations; the learnable multi-scale fusion module is used to adaptively fuse the multi-scale state representations and output degradation prediction results; and predicting battery health status or remaining service life based on the MSP-Mamba model.

[0007] In an optional implementation, obtaining the multi-source fused feature dataset specifically includes: acquiring degradation information data covering the entire lifespan of the battery as the original data; constructing a battery physical prior generation module based on the battery physical degradation mechanism, calculating relevant mechanism features through the battery physical prior generation module to form a physical prior that complements the original data; and performing multi-source feature fusion and standardization to obtain the multi-source fused feature dataset.

[0008] In an optional implementation, the raw data includes one or more of the battery's voltage, current, temperature, capacity, and SOH data.

[0009] In an optional implementation, the relevant mechanistic features include one or more of the following: degradation rate, SEI film growth-related characterizations, theoretical capacity, capacity deviation rate, and SOH change.

[0010] In an optional implementation, obtaining the weighted feature sequence specifically includes: the correlation weight generation unit performs correlation evaluation and channel-by-channel weighting on the multi-source fused feature dataset to suppress low-correlation or noisy channels and enhance effective channels consistent with the degradation process; wherein, within a preset sliding window, the correlation between the degradation prior and the multi-source features is calculated based on the Pearson correlation coefficient to generate channel-level weights to enhance highly correlated features and suppress noise.

[0011] In an optional implementation, the construction of the MSP-Mamba model specifically includes: inputting the weighted feature sequence into a linear projection convolution module to obtain a structured hidden feature sequence through channel mapping and feature reshaping; inputting the hidden feature sequence into multi-scale convolution branches constructed with different kernel sizes to extract degradation representations at multiple time scales through a multi-scale feature extraction module; setting an independent selective state space module after each scale convolution branch to form a multi-scale state representation; weighting and aggregating the multi-scale state representations through a learnable multi-scale fusion module to form a unified fused temporal representation; and mapping the window dimension of the fused temporal representation to a single predicted value through an output mapping layer to obtain the predicted result of the energy storage battery health status or remaining service life.

[0012] In an optional implementation, the multiple time scales include short-term, medium-term, and long-term time scales; the hidden feature sequence is... Where b is the batch size; w is the time window length; and d is the time window length. m To hide the channel dimension, each convolutional branch uses a different kernel size to correspond to the short-term, medium-term, and long-term time scales, respectively. The convolutional expression is as follows:

[0013] Then, a nonlinear activation is added to the convolution output of each convolution branch, which is characterized as follows: The state-space modeling process for the above three convolutional branches is represented as follows:

[0014] Where i = 1, 2, 3 for each SSM i Each is an independent MSP-Mamba branch, corresponding to short-term, medium-term, and long-term time scales, respectively; the multi-scale state representation is expressed as: F1, F2, and F3 correspond to short-term, medium-term, and long-term feature representations, respectively; the multi-scale state representation serves as the input to the learnable multi-scale fusion module.

[0015] In an optional implementation, the MSP-Mamba model includes an input selection gate, a state update equation, and an output selection mechanism; wherein, the input selection gate is represented as:

[0016] Among them, F i,t The input features are at time step t; W u These are learnable weights; The Sigmoid activation function controls the degree of selection of input information; the state update equation is expressed as:

[0017] Among them, X t The hidden state is represented by A and B, which are learnable parameters in the state space. The output selection mechanism is expressed as follows:

[0018] Where C is the output projection matrix, used to map the MSP-Mamba hidden states to temporal outputs; for the three time scales, the MSP-Mamba branch at each scale independently executes this inference process, resulting in three sets of sequences with selective state memory capabilities: .

[0019] In an optional implementation, when predicting battery health status or remaining lifespan based on the MSP-Mamba model, three indicators—relative error, mean absolute error, and root mean square error—are used to evaluate the performance of the MSP-Mamba model.

[0020] According to a second aspect of this application, embodiments of this application provide a battery health status or remaining lifespan prediction system, the system having program modules corresponding to the steps of the above-described method, and executing the steps in the above-described battery health status or remaining lifespan prediction method during runtime.

[0021] According to a third aspect of this application, an embodiment of this application provides a computer-readable storage medium storing computer program instructions, which, when read and executed by a processor, perform the aforementioned battery health status or remaining lifespan prediction method.

[0022] Implementing the technical solution of the present invention has at least the following beneficial effects: In the embodiments of this application, the battery health status or remaining service life prediction method provided can improve prediction stability and robustness, enhance the ability to characterize degradation patterns across multiple time scales, enhance interpretability, and improve the generalization ability across batteries / operating conditions, which is conducive to improving the accuracy and stability of prediction and enabling more accurate prediction and estimation of battery health status or remaining service life. Attached Figure Description

[0023] Figure 1 is a flowchart illustrating a method for predicting battery health status or remaining service life according to an embodiment of the present invention.

[0024] Figure 2 is a flowchart illustrating another method for predicting battery health status or remaining service life provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Where specific conditions are not specified in the embodiments, conventional conditions or conditions recommended by the manufacturer shall apply.

[0026] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0027] In related technologies, the battery health status prediction methods still have certain shortcomings, mainly including: (1) lack of integration of physical laws: most traditional deep learning methods are purely data-driven and do not fully consider the physical degradation mechanism of batteries. (2) insufficient modeling of multi-timescale characteristics: the battery degradation process involves short-term fluctuations, medium-term trends and long-term decay, and existing methods often perform poorly in dealing with these multi-scale characteristics. (3) significant interference of feature noise on modeling: there are often sensing errors and operating condition disturbances in the original collected data, resulting in feature dimension redundancy, noise accumulation and prediction instability. Therefore, there is still a need for improvement.

[0028] In view of this, in order to solve the above problems, the technical solution of this application provides a method, system and storage medium for predicting battery health status or remaining service life. The method is an energy storage battery health prediction method based on MSP-Mamba (Multi-Scale Physics-guided Mamba). Based on this method, a battery degradation prediction framework can be constructed, which fundamentally improves the generalization ability and prediction stability of battery degradation prediction in cross-operating conditions, and can effectively overcome the shortcomings of the above-mentioned related technologies. The specific technical solution is described below.

[0029] Referring to Figure 1, in some embodiments, a method for predicting battery health status or remaining service life is provided. The method includes: acquiring a battery life dataset and preprocessing it to obtain a multi-source fused feature dataset; evaluating and weighting the multi-source fused feature dataset through a correlation weight generation unit to obtain a weighted feature sequence; constructing an MSP-Mamba model, with the weighted feature sequence as the input to the MSP-Mamba model. The MSP-Mamba model includes a multi-scale feature extraction module, a selective state space module, and a learnable multi-scale fusion module. The multi-scale feature extraction module is used to extract multi-timescale degradation representations during battery degradation through multi-scale convolutional branches. The selective state space module, based on a selective state space reasoning mechanism, is used to perform temporal reasoning and dynamic memory modeling on the multi-timescale degradation representations to form multi-scale state representations. The learnable multi-scale fusion module is used to adaptively fuse the multi-scale state representations and output degradation prediction results. The battery health status or remaining service life is predicted based on the MSP-Mamba model.

[0030] In this embodiment, the provided prediction method can be used to predict or estimate the state of health (SOH) of the battery, and also to predict or estimate the remaining useful life (RUL) of the battery.

[0031] It should be understood that the State of Health (SOH) and Remaining Life (RUL) of a battery are crucial for evaluating its performance, safety, and reliability. Remaining life primarily refers to the length of time a battery can provide effective energy under its current state of health. Effective SOH or RUL prediction methods can be used to extend battery life, improve battery safety and performance, and support the development of battery applications.

[0032] In this embodiment of the invention, the provided energy storage battery health prediction method based on MSP-Mamba (Multi-Scale Physics-guided Mamba) constructs a battery degradation prediction framework, including: firstly, constructing a prior knowledge module based on the physical degradation mechanism of the battery; designing a correlation weight generation unit to assign weights to different degradation features based on physical priors; then, extracting short-term, medium-term, and long-term degradation characteristics in the battery degradation process through the multi-scale convolutional branches of MSP-Mamba; for the features extracted at each scale, adopting the selective state-space reasoning mechanism of MSP-Mamba, combined with physical prior constraints to achieve accurate modeling of degradation laws at each scale; finally, through the learnable cross-scale fusion module built into the model, adaptively fusing the reasoning results at each scale and outputting the final degradation prediction result.

[0033] This invention, based on the dual dimensions of battery physical mechanisms and multi-scale time-series modeling, fundamentally improves the generalization ability and prediction stability of battery degradation prediction across operating conditions by leveraging the physical guidance and multi-scale state-space reasoning characteristics of the MSP-Mamba model. In other words, it can improve the accuracy and stability of battery state of health (SOH) and battery remaining service life (RUL) prediction, enabling more accurate battery SOH or RUL estimation.

[0034] The MSP-Mamba model combines multi-scale physical knowledge with the next-generation sequence model Mamba for predicting the health of energy storage batteries. This MSP-Mamba model utilizes a multi-scale physical model to reveal the intrinsic mechanisms of battery aging, providing interpretability and physical consistency constraints; and leverages the Mamba model to efficiently process long-sequence data on battery operation (such as voltage, current, and temperature), capturing complex time dependencies and dynamic patterns. By deeply integrating the two, it achieves more accurate, robust, and generalized predictions of State of Health (SOH) and Remaining Life (RUL) than a single data-driven or single-physical model.

[0035] Introducing the MSP-Mamba model into predictions for SOH or RUL offers high accuracy and robustness. Its physical knowledge reduces the model's dependence on data volume, improving prediction stability under unseen conditions or data noise. It also enhances interpretability, as its predictions can be correlated with specific physical processes, rather than being merely "black box" outputs. Furthermore, it possesses early prediction capabilities, potentially capturing subtle degradation signals from early battery cycling through multi-scale physical features, enabling earlier health warnings. Finally, it is computationally efficient: the Mamba module is more efficient than Transformers when processing long-sequence data, making it suitable for application in the embodiments of this application.

[0036] In some implementations, the provided method for predicting the state of health (SOH or relative uptime) of energy storage batteries based on MSP-Mamba mainly includes: calculating mechanistic characteristics such as degradation rate, capacity deviation rate, and SOH change through a Battery Physics-Prior Generator (B-PPG) module to form a physical prior that complements the original observation data, providing mechanistic constraints for MSP-Mamba; and using a correlation weighting unit... The Concurrent Functional Unit (CWU) calculates the correlation between degraded priors and multi-source features based on the Pearson correlation coefficient within a sliding window, generating channel-level weights to enhance highly correlated features, suppress noise, and optimize the input feature quality of MSP-Mamba. It utilizes the multi-scale convolutional branches built into MSP-Mamba to extract short-term, medium-term, and long-term degraded features, and inputs each scale feature into the model's selective state space unit to achieve multi-scale dynamic inference. Finally, through MSP-Mamba's learnable multi-scale fusion mechanism, it achieves adaptive aggregation of representations at different time scales, improving the stability, robustness, and cross-cell generalization ability of predictions from both physical mechanisms and multi-scale temporal modeling dimensions.

[0037] Therefore, the proposed prediction method comprehensively employs physical degradation prior construction, multi-source feature fusion and standardization, correlation weight generation, MSP-Mamba architecture, and a learnable multi-scale fusion mechanism to achieve robust modeling and prediction output of battery degradation processes. This method is applicable to predicting the State of Health (SOH) or Remaining Useful Life (RUL) of energy storage batteries and other rechargeable batteries with cycle degradation characteristics.

[0038] In some specific implementations, step S10, acquiring and preprocessing the battery life dataset to obtain a multi-source fused feature dataset, includes: S110, acquiring degradation information data covering the entire battery life stage as the raw data; S120, constructing a battery physical prior generation module based on the battery physical degradation mechanism, calculating relevant mechanism features through the battery physical prior generation module to form a physical prior that complements the raw data; as an example, the battery physical prior generation module (Battery Physics-Prior Generator, B-PPG) calculates mechanism features such as degradation rate, capacity deviation rate, and SOH change to form a physical prior that complements the raw observation data, providing a mechanism constraint for MSP-Mamba; S130, performing multi-source feature fusion and standardization to obtain a multi-source fused feature dataset. For example, the raw features and physical degradation features are concatenated along the channel dimension.

[0039] Step S10 can be a preparation stage for building the MSP-Mamba model, which can be used to complete the raw data processing, physical degradation prior feature construction, feature fusion, and standardization and correlation weighting.

[0040] To obtain degradation information covering the entire lifespan, the dataset used in this embodiment of the invention is derived from time-series data generated by multiple rounds of cyclic testing of the battery. The testing process includes the collection of charging, discharging, and state monitoring data to construct a multi-dimensional time-series feature covering the entire lifespan of the battery, thereby providing a data foundation for characterizing degradation mechanisms and predicting lifespan.

[0041] Raw battery data typically contains various noise sources, including sensor errors, temperature perturbations, and electrochemical randomness. Directly using this data for neural network training can lead to gradient instability, convergence difficulties, and prediction bias. Therefore, this invention first preprocesses the raw features and the degraded features calculated based on physical modules, and then performs standardization on the fused multi-source features to reduce the impact of numerical scale differences and noise perturbations from different sources on subsequent modeling, thereby improving the consistency and stability of the input representation.

[0042] In terms of constructing physical priors, this embodiment of the invention sets up a battery physical prior generation module, which calculates degradation prior features based on battery capacity decay theory and degradation mechanism. These degradation prior features, i.e., relevant mechanism features, include but are not limited to degradation rate, SEI film growth-related characterization quantities, theoretical capacity, capacity deviation rate, and SOH change. The degradation prior features are then concatenated with the original acquired features along the channel dimension to obtain a fused feature sequence, achieving a joint expression of mechanism features and observational features.

[0043] For example, step S10 includes: experimentally processing two publicly available energy storage battery lifetime datasets to obtain the capacity degradation curves and multidimensional time series features of each battery throughout its entire lifespan.

[0044] The system is divided into a training set and a test set. The degradation data of one battery is used as the test set, and the data of the other batteries is used as the training set.

[0045] The resulting original sequence is composed of information such as voltage, current, temperature, capacity, and state of equilibrium (SOH) collected during battery charge-discharge experiments. Its input tensor dimension is x∈R. (b×w×d) Where: b represents the batch size; w represents the time window length; d is the number of original features. In order to enhance the model's understanding of the internal mechanism of the battery, the battery physical prior generation module in this step calculates the physical degradation feature sequence based on the electrochemical laws and degradation mechanism of the battery, including the following five types of physical quantities: (1) Degradation rate quantity D tAccording to the Arrhenius degradation mechanism, the effect of temperature on the SEI film growth rate is expressed as follows:

[0046] Where: D0 is the initial degradation constant; E a is the activation energy; k is the Boltzmann constant; T t Let t be the temperature at time step t.

[0047] This feature forms a tensor of type D∈R within the time window. (b×w×1) .

[0048] (2) Capacity degradation fitting output P out (t): The linear approximation relationship between degradation rate and capacity is as follows:

[0049] Where C and b are derived from the fitting constants. This yields P∈R. (b×w×1) .

[0050] (3) Theoretical capacity and capacity deviation rate δQ k First, the theoretical capacity C is obtained based on the electrochemical relationship. theo :

[0051] Where: F is the Faraday constant; n is the number of electrons transferred; M is the mass of the active substance.

[0052] Then, the maximum capacity Q is obtained by combining the mass of the cell's active materials. k,max :

[0053] This yields the capacity deviation rate for each cycle:

[0054] Forming a physical characteristic sequence: δQ k ∈R (b×w×1) .

[0055] (4) SOH attenuation ΔSOH k :

[0056] Agent quantity indicating battery health status , where Qk is the battery capacity at the current moment, Q0 is the initial capacity, and ε is a stabilizing term used to prevent the denominator from being zero.

[0057]

[0058] It describes the change in health status between two adjacent time steps.

[0059] Formation of physical characteristic sequence: △SOH k ∈R (b×w×1) .

[0060] (5) Rate of change of SOH R k : Reflects the rate of degradation per unit time:

[0061] Forming a sequence of physical characteristics: R k ∈R (b×w×1) .

[0062] When a physical characteristic includes the above five types of physical quantities, its dimensions are: d phys =5. Concatenate the original features and the physical degradation features along the channel dimension. The dimensions after concatenation are: Where: the original sequence contributes d channels; the physical sequence contributes d channels. phys One channel.

[0063] In some specific implementations, step S20, which involves evaluating and weighting the feature dataset of the multi-source fusion through the correlation weight generation unit to obtain a weighted feature sequence, includes: evaluating the correlation of the feature dataset of the multi-source fusion through the correlation weight generation unit and weighting it channel by channel to suppress low-correlation or noisy channels and enhance effective channels consistent with the degradation process; wherein, within a preset sliding window, the correlation between the degradation prior and the multi-source features is calculated based on the Pearson correlation coefficient to generate channel-level weights to enhance highly correlated features and suppress noise.

[0064] Step S20 can be a preparatory stage for building the MSP-Mamba model and can be used to complete the correlation weighting.

[0065] To reduce the interference of invalid and noisy channels on subsequent deep time series modeling, this embodiment of the invention sets up a correlation weight generation unit. Within a preset sliding window, the Pearson correlation coefficient between each channel of the fused feature and the degenerate prior characterization sequence is calculated, and a correlation weight vector is generated through normalization mapping. The fused features are then weighted channel by channel, so that channels with higher consistency with the degradation process are enhanced, while low-correlation channels are suppressed, thereby obtaining a correlation-enhanced feature sequence.

[0066] For example, the correlation weight generation unit performs correlation evaluation and channel-by-channel weighting on the multi-source fusion features to suppress low-correlation or noisy channels and enhance effective channels consistent with the degradation process. Specifically, this includes: assuming the standardized fusion feature sequence is... The value of the k-th channel after expansion across all samples and time steps is denoted as . The expanded sequence of the target variable (SOH or RUL) is denoted as... N = b × w. Then the Pearson correlation coefficient r between the k-th feature and the target variable is... k Defined as:

[0067] in, and These represent the mean of the k-th feature and the target variable across all samples, respectively. The correlation coefficients of all channels are then combined into a vector. To retain both strong positive and strong negative correlations, the absolute value of the correlation coefficient is taken and normalized to obtain the correlation weight vector W. CWU :

[0068] Subsequently, W CWU Assuming weights along the channel dimension, the standardized feature sequences are weighted according to the channel dimension to obtain the weighted feature sequence X. CWU :

[0069] Therefore, through the above processing, highly correlated features in X CWU The input quality of the subsequent multi-scale feature extraction module is improved, and low-correlation or noisy features are suppressed, thereby enhancing the effective feature expression capability of the prediction model.

[0070] In some specific implementations, step S30, the step of constructing the MSP-Mamba model, includes: inputting the weighted feature sequence into a linear projection convolution module, and obtaining a structured hidden feature sequence through channel mapping and feature reshaping; inputting the hidden feature sequence into multi-scale convolution branches constructed with different convolution kernel sizes, so as to extract degradation representations at multiple time scales through a multi-scale feature extraction module; for example, the multiple time scales include short-term, medium-term and long-term time scales; setting an independent selective state space module after each scale convolution branch to form a multi-scale state representation; weighting and aggregating the multi-scale state representations through a learnable multi-scale fusion module to form a unified fused temporal representation, and mapping the window dimension of the fused temporal representation to a single predicted value through an output mapping layer to obtain the prediction result of the energy storage battery health status or remaining service life.

[0071] Step S30 is the core modeling stage, which can be used to sequentially perform linear projection channel mapping, MSP-Mamba temporal inference, and learnable multi-scale aggregation, and output health status or remaining life prediction values.

[0072] In this embodiment of the invention, during the multi-scale feature extraction and temporal modeling stage, the relevance-enhanced feature sequence is first input into a linear projection convolution module to perform channel dimensionality upscaling and mapping to a unified hidden feature space, thereby obtaining a structured hidden feature sequence. Subsequently, the hidden feature sequence is input into multi-scale convolution branches constructed with different kernel sizes to extract degradation representations at short-term, medium-term, and long-term time scales, respectively. After each scale branch, a selective state-space modeling module of MSP-Mamba is introduced to perform temporal reasoning and dynamic memory modeling on the corresponding scale sequence, thereby representing long-range dependencies and degradation dynamic features at each time scale, forming a multi-scale state representation.

[0073] To achieve adaptive fusion of cross-scale information, this embodiment of the invention sets up a learnable multi-scale fusion mechanism after the MSP-Mamba output. By setting learnable fusion weights α, β, and γ, and satisfying the constraint α+β+γ=1, the state representations of short-term, medium-term, and long-term scales are weighted and aggregated to form a unified fused representation. The fused time-series representation is further mapped to a single predicted value by an output mapping layer to obtain the RUL prediction result or SOH prediction result of the energy storage battery.

[0074] For example, step S30 specifically includes: obtaining a feature sequence X that effectively enhances key information and suppresses noise components through a correlation weight generation unit. CWU This sequence serves as the foundational input for subsequent deep models. To ensure that feature channels have a uniform hidden dimension representation before entering the multi-scale modules of MSP-Mamba, X... CWU The input is a linear projection convolution module, which obtains the hidden feature sequence through channel mapping and feature reshaping. , where d m To hide the channel dimension, and to capture degradation dynamics across different time ranges, three parallel multi-scale convolutional branches are set up. Each branch uses a different kernel size, corresponding to short-term, medium-term, and long-term time scales, respectively. Their convolution expressions are as follows:

[0075] Then, a non-linear activation is added to the convolution output of each convolution branch:

[0076] To further perform in-depth temporal modeling of the aforementioned multi-scale features and retain key degradation information at different time scales, this embodiment of the invention sets up an independent selective state space modeling module after each convolutional branch. This module is a concrete implementation of the MSP-Mamba model, used for selective state updates and dynamic memory modeling of sequence dependencies at the corresponding time scales, thereby extracting degradation state representations at different scales. The state space modeling process of the three branches can be represented as follows:

[0077] Where i = 1, 2, 3 for each SSM i Each is an independent MSP-Mamba branch, corresponding to short-term, medium-term, and long-term time scales, respectively. Ultimately, the processed sequences at the three scales together constitute the model's multi-scale time-series representation, which can be expressed as:

[0078] This multi-scale sequence serves as the input to subsequent learnable multi-scale fusion layers, enabling the model to simultaneously integrate short-term, medium-term, and long-term degradation information within a unified hidden space, and providing a structured and semantically complete feature representation for the final prediction of the energy storage battery's health status or remaining lifespan.

[0079] Furthermore, the obtained three-scale feature sequences are input into the selective state-space reasoning mechanism (selective state-space module) of the MSP-Mamba model to form a state representation with selective memory update capability at each time scale, and to provide a unified temporal representation for subsequent cross-time scale information integration. Specifically, let the outputs of the three branches be F1, F2, and F3, respectively, corresponding to the short-term, medium-term, and long-term feature representations. As a selective state-space model, MSP-Mamba's core consists of input selection gating, state update equations, and output selection mechanisms. For any scale Fi, its internal reasoning can be simplified as follows: (1) MSP-Mamba input selection gate:

[0080] Among them, F i,t The input features are at time step t; W u For learnable weights, This is the Sigmoid activation function, used to control the degree of selection of input information.

[0081] (2) State update equation:

[0082] Among them, X t A represents a hidden state, and A and B are learnable parameters in the state space.

[0083] (3) Output selection mechanism:

[0084] Where C is the output projection matrix, used to map the MSP-Mamba hidden states to temporal outputs. For each of the three time scales, the MSP-Mamba branch executes this inference process independently, resulting in three sets of sequences with selective state memory capabilities:

[0085] To adaptively integrate the features that contribute most to degradation prediction across the three time scales, this method incorporates a learnable multi-scale fusion module after the MSP-Mamba output. Let the learnable weights be:

[0086] The fused multi-scale representation can then be simplified as:

[0087] This fusion operation enables the model to automatically adjust the importance of each time scale based on the data, achieving structurally interpretable and dynamically adjustable multi-scale feature aggregation. The fused sequence needs to have its time dimension reduced to a single predicted value to obtain the final output mapping.

[0088] In some specific implementations, step S40, which involves predicting the battery health status or remaining lifespan based on the MSP-Mamba model, includes using three metrics—relative error, mean absolute error, and root mean square error—to evaluate the performance of the MSP-Mamba model.

[0089] Step S40 is the verification and evaluation stage, which is used to test and comprehensively evaluate the model's predictive performance based on evaluation indicators.

[0090] Specifically, step S40 includes: evaluating the model's performance using the predicted values ​​output by the model, wherein the method uses three metrics—relative error (RE), mean absolute error (MAE), and root mean square error (RMSE)—to evaluate the model's performance. For example:

[0091] Among them, RUL pred and RUL ture These represent the predicted RUL and the actual RUL, respectively.

[0092]

[0093] Where s represents the number of samples in the test set; y ty represents the true label value of the t-th sample; t * This represents the model's prediction result for the t-th sample; MAE represents the average of the absolute values ​​of errors for all samples, and the smaller the value, the more accurate the prediction.

[0094]

[0095] The smaller the values ​​of RE, MAE, and RMSE mentioned above, the better the performance. To evaluate the performance of all models, the method will use leave-one-out evaluation on all data.

[0096] Thus, through the above technical solution, the present invention achieves collaborative characterization of multi-scale degradation dynamics, and improves the effectiveness and robustness of input features by leveraging physical degradation priors and correlation weighting. Combined with selective state-space modeling and learnable cross-scale fusion, it achieves stable time-series prediction output, which is suitable for energy storage battery health management and lifetime prediction application scenarios.

[0097] In some embodiments, a battery health status or remaining lifespan prediction system is also provided. The system has program modules corresponding to the steps of the above-described method, and executes the steps in the battery health status or remaining lifespan prediction method when running.

[0098] In some embodiments, a computer-readable storage medium is also provided, on which computer program instructions are stored, which, when read and executed by a processor, perform the battery health status or remaining lifespan prediction method.

[0099] In summary, this invention realizes a "mechanism-data-temporal" prediction chain guided by physical degradation priors. It constructs a closed-loop process from physical mechanism to deep temporal modeling: the Battery Physical Prior Generation (B-PPG) module generates interpretable degradation prior features based on physical degradation formulas; these features are fused and standardized with the original observed features at the channel dimension; and then fed into a multi-scale selective state-space network for temporal inference and prediction output. In other words, it provides a method for constructing physical degradation priors and their fusion with original features, as well as the input organization of this fused representation in the subsequent multi-scale selective state-space inference chain.

[0100] This invention also employs a channel-wise weighted input mechanism using a correlation weight generation unit (CWU). To address the interference of multi-source feature noise and redundant channels on time-series modeling, this invention sets up a correlation weight generation unit (CWU): within a preset sliding window, it calculates the Pearson correlation between the original features and degenerate prior features, and generates a weight vector through normalization mapping. This vector is then used to weight the fused features channel-wise, achieving the input of "enhancing key channels and suppressing noisy channels".

[0101] This invention employs a multi-scale physically guided Mamba structure (MSP-Mamba), deeply integrating multi-scale representation learning with physical prior guidance. It captures degradation dynamics across short, medium, and long-term time dimensions through multi-scale branches, while leveraging physical constraints to ensure that features at each scale conform to battery degradation patterns. This invention proposes an overall MSP-Mamba architecture that integrates physical degradation mechanisms with multi-scale temporal modeling. By allocating feature weights based on battery physical degradation prior constraints, this invention simultaneously captures short, medium, and long-term degradation dynamics through multi-scale parallel branches, and then allows each scale to independently complete temporal inference. This achieves a synergistic improvement in both "physical interpretability" and "multi-scale representation capability," overcoming the limitations of traditional models that either lack physical basis or struggle to comprehensively characterize complex degradation processes.

[0102] This invention employs learnable cross-scale fusion. To address the issue of varying dependence on short / medium / long-term information at different degradation stages, this invention sets up a learnable fusion module after the MSP-Mamba output. This module adaptively aggregates state representations at different scales using learnable weights, and then maps the window dimension to a single predicted value output SOH or RUL via an output mapping layer.

[0103] Therefore, based on the above settings, the main beneficial effects of the present invention include: (1) Improved prediction stability and robustness: By organizing the input with "physical degradation prior and correlation channel weighting", the interference of redundant and noise channels on time series modeling is reduced, so that the prediction output can maintain a more stable trend consistency when there are operating condition disturbances or data noise.

[0104] (2) Enhanced ability to represent degradation patterns at multiple time scales: Parallel multi-scale convolution extracts short / medium / long-term degradation features, and dynamic memory modeling is performed by MSP-Mamba implementation at each scale, enabling the model to simultaneously characterize short-term fluctuations, medium-term trends and long-term decay, thereby improving the modeling ability for complex non-stationary degradation processes.

[0105] (3) Enhanced interpretability: Physical degradation prior features provide a mechanistic explanation for the prediction; correlation weights and cross-scale fusion weights can reflect the relative contributions of different channels or different time scales to the prediction, thereby improving the interpretability and analyzability of the model output.

[0106] (4) Improved generalization ability across batteries / operating conditions: In the cross-battery partitioning experiment of the public dataset, this technology shows better or more stable results in terms of error index and stability compared with the traditional deep time series model, which reflects the adaptability to different batteries and different operating conditions.

[0107] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for predicting battery health status or remaining service life, characterized in that, The method includes: acquiring a battery life dataset and preprocessing it to obtain a multi-source fused feature dataset; evaluating and weighting the multi-source fused feature dataset using a correlation weight generation unit to obtain a weighted feature sequence; constructing an MSP-Mamba model, with the weighted feature sequence serving as the input to the MSP-Mamba model, wherein the MSP-Mamba model includes a multi-scale feature extraction module, a selective state space module, and a learnable multi-scale fusion module; wherein the multi-scale feature extraction module is used to extract multi-timescale degradation representations during battery degradation through multi-scale convolutional branches; the selective state space module, based on a selective state space reasoning mechanism, is used to perform temporal reasoning and dynamic memory modeling on the multi-timescale degradation representations to form multi-scale state representations; the learnable multi-scale fusion module is used to adaptively fuse the multi-scale state representations and output degradation prediction results; and predicting the battery health status or remaining service life based on the MSP-Mamba model.

2. The method for predicting battery health status or remaining service life according to claim 1, characterized in that, The process of obtaining the multi-source fused feature dataset specifically includes: acquiring degradation information data covering the entire lifespan of the battery as the original data; constructing a battery physical prior generation module based on the battery physical degradation mechanism, calculating relevant mechanism features through the battery physical prior generation module to form a physical prior that complements the original data; and performing multi-source feature fusion and standardization to obtain the multi-source fused feature dataset.

3. The method for predicting battery health status or remaining service life according to claim 2, characterized in that, The raw data includes one or more of the following: battery voltage, current, temperature, capacity, and SOH data; and / or, the relevant mechanistic features include one or more of the following: degradation rate, SEI film growth-related characterizations, theoretical capacity, capacity deviation rate, and SOH change.

4. The method for predicting battery health status or remaining service life according to claim 1, characterized in that, The obtained weighted feature sequence specifically includes: the correlation weight generation unit performs correlation evaluation and channel-by-channel weighting on the multi-source fused feature dataset to suppress low-correlation or noisy channels and enhance effective channels consistent with the degradation process; wherein, within a preset sliding window, the correlation between degradation priors and multi-source features is calculated based on the Pearson correlation coefficient to generate channel-level weights to enhance highly correlated features and suppress noise.

5. The method for predicting battery health status or remaining service life according to any one of claims 1 to 4, characterized in that, The construction of the MSP-Mamba model specifically includes: inputting the weighted feature sequence into a linear projection convolution module to obtain a structured hidden feature sequence through channel mapping and feature reshaping; inputting the hidden feature sequence into multi-scale convolution branches constructed with different convolution kernel sizes to extract degradation representations at multiple time scales through a multi-scale feature extraction module; setting an independent selective state space module after each scale convolution branch to form a multi-scale state representation; weighting and aggregating the multi-scale state representations through a learnable multi-scale fusion module to form a unified fused temporal representation; and mapping the window dimension of the fused temporal representation to a single predicted value through an output mapping layer to obtain the predicted result of the energy storage battery health status or remaining service life.

6. The method for predicting battery health status or remaining service life according to claim 5, characterized in that, The multiple time scales include short-term, medium-term, and long-term time scales; the hidden feature sequence is... Where b is the batch size; w is the time window length; and d is the time window length. m To hide the channel dimension, each convolutional branch uses a different kernel size to correspond to the short-term, medium-term, and long-term time scales, respectively. The convolutional expression is as follows: Then, a nonlinear activation is added to the convolution output of each convolution branch, which is characterized as follows: The state-space modeling process for the above three convolutional branches is represented as follows: Where i = 1, 2, 3 for each SSM i Each is an independent MSP-Mamba branch, corresponding to short-term, medium-term, and long-term time scales, respectively; the multi-scale state representation is expressed as: F1, F2, and F3 correspond to short-term, medium-term, and long-term feature representations, respectively; the multi-scale state representation serves as the input to the learnable multi-scale fusion module.

7. The method for predicting battery health status or remaining service life according to claim 6, characterized in that, The MSP-Mamba model includes an input selection gate, a state update equation, and an output selection mechanism; wherein, the input selection gate is represented as: Among them, F i,t The input features are at time step t; W u These are learnable weights; The Sigmoid activation function controls the degree of selection of input information; the state update equation is expressed as: Among them, X t The hidden state is represented by A and B, which are learnable parameters in the state space. The output selection mechanism is expressed as follows: Where C is the output projection matrix, used to map the MSP-Mamba hidden states to temporal outputs; for the three time scales, the MSP-Mamba branch at each scale independently executes this inference process, resulting in three sets of sequences with selective state memory capabilities: 。 8. The method for predicting battery health status or remaining service life according to claim 1, characterized in that, In predicting battery health status or remaining lifespan based on the MSP-Mamba model, three indicators—relative error, mean absolute error, and root mean square error—are used to evaluate the performance of the MSP-Mamba model.

9. A battery health status or remaining service life prediction system, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 8, and executes the steps in the battery health status or remaining lifespan prediction method described above when running.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when read and executed by a processor, perform the battery health status or remaining service life prediction method according to any one of claims 1 to 8.

Images