A Power Battery Life Prediction Method Based on Mechanism-Aware Sparse Coding and WTA Prediction

Abstract

This invention belongs to the field of power battery life prediction technology, and particularly relates to a power battery life prediction method based on mechanism-aware sparse coding and WTA prediction. First, multi-source operating data is collected and clustered into typical operating conditions. A sparse autoencoder with dynamic sparse masking, two-order sparse regularization, and physical consistency constraints is constructed to extract low-dimensional core aging features. Then, a multi-branch prediction structure adapted to the operating conditions is established, and a WTA competitive training mechanism is used to update only the optimal branch parameters to achieve specialized probabilistic prediction. Finally, the sliding window is dynamically adjusted based on feature sparse entropy, and the multi-branch results are fused through Kalman filtering and Bayesian rules to output the final RUL with confidence intervals, triggering multi-condition maintenance decisions. This invention solves the problem in existing technologies where the lack of deep integration of time-series sparse modeling, probabilistic prediction, and power battery degradation mechanisms leads to poor power battery life prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of power battery life prediction technology, and particularly relates to a power battery life prediction method based on mechanism-aware sparse coding and WTA prediction. Background Technology

[0002] With the large-scale popularization of the new energy vehicle industry, the State of Health (SOH) estimation and remaining life prediction of power batteries, as core energy storage components, are key technical aspects for ensuring vehicle operation safety, optimizing battery lifecycle management, and reducing operation and maintenance costs. Currently, power battery remaining life prediction technologies are mainly divided into three categories: 1. Physicochemical modeling method: Based on the electrochemical reaction mechanism, a complex model containing thermodynamic and kinetic equations is constructed, which can explain the battery aging process in essence. However, it has the drawbacks of high difficulty in parameter identification, high computational complexity, and difficulty in adapting to the variable operating conditions of real vehicles. 2. Equivalent circuit model method: The external characteristics of the battery are simulated by circuit components such as resistors and capacitors. The model structure is simple and the calculation speed is fast, but the generalization ability is weak. It cannot accurately represent the strong nonlinear decay characteristics during the battery aging process, and the prediction accuracy decreases significantly with the deepening of battery aging. 3. Data-driven approach: This approach learns battery degradation patterns from historical data using machine learning and deep learning algorithms. Representative methods include deep autoencoders, Long Short-Term Memory (LSTM) networks, and Transformers. For example, patent publication number CN120180059A, titled "A Power Battery Degradation Prediction Method Based on Deep Autoencoders," describes the construction of a deep autoencoder with fixed sparse constraints and the use of a single BiLSTM+Attention network to predict the remaining lifespan of the power battery. However, it has the following technical drawbacks: Data is disconnected from the attenuation mechanism: the model is often trained using static cyclic data from the laboratory, ignoring the coupling relationship between real vehicle multi-source heterogeneous data (such as altitude, driving mode, temperature fluctuation) and battery aging, resulting in low prediction accuracy under real vehicle conditions; The model lacks targeted optimization: the general sparse coding and time-series prediction model are not designed in conjunction with battery degradation characteristics, and the fixed sparse constraints are prone to loss of core aging features. In addition, the model parameters are redundant (usually reaching millions), which cannot meet the deployment requirements of low computing power and low latency in the vehicle. The prediction results lack uncertainty quantification: only a single-point RUL prediction value is output, which cannot reflect the lifespan fluctuation range under different aging scenarios and is difficult to support vehicle safety decisions and operation and maintenance scheduling. Insufficient dynamic adaptation capability: The use of a fixed sliding window and a single feature set cannot cope with the changes in the degradation mode throughout the battery's entire life cycle (early, middle, and late stages), resulting in uneven prediction accuracy at different stages.

[0003] Therefore, the root cause of the above defects is that the existing technology does not deeply integrate time-series sparse modeling, probability prediction and power battery degradation mechanism, but simply applies a general algorithm framework, lacks targeted innovative design for battery aging characteristics, and cannot simultaneously meet the engineering requirements of high precision, lightweight, scenario adaptation and uncertainty quantification. Summary of the Invention

[0004] The technical problem solved by this invention is to provide a power battery life prediction method based on mechanism-aware sparse coding and WTA prediction, so as to solve the problem that the existing technology does not deeply integrate time-series sparse modeling, probability prediction and power battery degradation mechanism, resulting in poor power battery life prediction accuracy.

[0005] The basic solution provided by this invention is a power battery life prediction method based on mechanism-aware sparse coding and WTA prediction, comprising: S1: Collect multi-source operation data of the entire life cycle of power batteries of new energy vehicles, preprocess the operation data to obtain the characteristic data of power batteries, and use a preset clustering algorithm to cluster the preprocessed operation data into typical working conditions; S2: Construct a sparse autoencoder consisting of an input layer, an encoder with embedded dynamic sparse coding, a decoder, and an output layer. The input layer matches feature data, the encoder side applies channel sparsity loss, the decoder side applies parameter sparsity loss, and physical consistency loss is introduced at the same time. The output layer outputs a low-dimensional sparse feature sequence. S3: Construct multiple differential prediction branches corresponding to typical working conditions. Each differential prediction branch performs RUL prediction based on the input being a low-dimensional sparse feature sequence. Based on the WTA competitive training mechanism, only the branch parameters with the smallest prediction error for the current sample are updated to obtain multiple differential prediction branches that have been trained. S4: Calculate the feature sparse entropy based on the low-dimensional sparse feature sequence output by the sparse autoencoder and dynamically adjust the sliding window; S5: Extract the low-dimensional sparse feature sequence of the adjusted sliding window length, input it into the trained differential prediction branch, and output the prediction result of RUL prediction value and prediction confidence of each differential prediction branch, and call Kalman filter to correct the prediction result. S6: Using a preset Bayesian fusion rule, the matching degree between typical working conditions and differentiated prediction branches, and the confidence level of the output of differentiated prediction branches, the prediction results of each differentiated prediction branch are weighted and fused to obtain the final RUL prediction value and prediction confidence level. S7: Construct a maintenance decision rule base with multiple conditions linked. Based on the final RUL prediction value and prediction confidence, perform rule matching and triggering with the maintenance decision rule base to output the maintenance decision result.

[0006] Furthermore, S1 includes: S1-1: Collect multi-source operational data of the entire life cycle of power batteries for new energy vehicles, including core battery parameters, operating environment parameters, vehicle status parameters, and offline testing data; S1-2: The collected multi-source running data are sequentially processed by missing value filling and feature normalization to obtain preprocessed feature data; S1-3: The K-means algorithm is used to cluster the preprocessed feature data into typical working conditions. The clustering objective function is:

[0007] in, The clustering objective function is... The number of clusters, For the i-th cluster, Let i be the centroid of the i-th cluster. These are the feature values ​​of the preprocessed feature data.

[0008] Furthermore, the typical operating conditions include low temperature high rate, normal temperature cycling, high temperature static placement, intermittent discharge, and long-term storage.

[0009] Furthermore, S2 includes: S2-1: Construct a sparse autoencoder consisting of an input layer, an encoder, a decoder, and an output layer; S2-2: In a sparse autoencoder, the input layer receives and carries the preprocessed feature data; S2-3: In the sparse autoencoder, the encoder embeds a dynamic sparse mask to perform high-dimensional feature compression, sparsification, and core aging feature filtering on the input layer's feature data, outputting a low-dimensional sparse feature vector. Channel sparsity loss is then applied to the output low-dimensional sparse feature vector to suppress redundant channels. The expression is as follows:

[0010] in, This is a penalty for channel sparsity. For the first The layer encoder outputs a feature vector. This represents the total number of layers in the encoder. It is an L1 norm. For the first The number of neurons in the layer encoder; S2-4: In a sparse autoencoder, the decoder reconstructs the low-dimensional sparse features output by the encoder back to the original dimension, and applies a parametric sparse loss to the decoder weights, pruning redundant parameters. The expression is:

[0011] in, For parameter sparsity penalty term, For the first Layer decoder weight matrix, This represents the number of decoder layers. It is an L2 norm matrix. For the first The number of input neurons in the layer decoder. For the first The number of output neurons in the layer decoder For the first The connection weights from the i-th input neuron to the j-th output neuron in the layer decoder; S2-5: Introduce physical consistency constraints of capacity and internal resistance into the loss functions of the encoder and decoder, expressed as follows:

[0012] in, , These are the weighting coefficients. , The input features are estimated values ​​for capacity and internal resistance. , These are the reconstructed capacitance and internal resistance values; S2-6: The reconstruction loss, channel sparsity loss, parameter sparsity loss, and physical consistency loss are fused together, and the expression is:

[0013] in: For the total loss function, The reconstruction loss is expressed as the reconstruction error. , , The regularization coefficient; S2-7: The output layer directly outputs the low-dimensional sparse feature sequence and reconstruction error reconstructed by the decoder; S2-8: Training and optimizing sparse autoencoders using an improved Adam optimizer.

[0014] Furthermore, S3 includes: S3-1: Construct differentiated prediction branches adapted to typical working conditions. The input of each differentiated prediction branch is the low-dimensional sparse feature sequence output by the sparse autoencoder. , Let be the sliding window length, and d be the feature dimension; each differential prediction branch includes a bidirectional LSTM layer, an attention layer, and a fully connected output layer, expressed as: Bidirectional LSTM layer:

[0015] in, Let k be the hidden state sequence of the k-th branch. Here are the LSTM network parameters for the k-th branch; Attention layer:

[0016]

[0017] in, For attention weight vectors, , These are the trainable parameters for the attention layer. For context vectors, Let be the hidden state at the i-th time step in the k-th branch; Fully connected output layer:

[0018]

[0019] in, , These are the trainable parameters for the fully connected layer. The RUL prediction value output by the k-th branch. This corresponds to the prediction confidence level, i.e., the standard deviation estimate. This represents the average predicted value of the k-th branch within the current window; S3-2: Construct the dynamic competitive weights for each differentiated prediction branch based on the basic competitive weights and the operating condition matching weights. The expression is:

[0020]

[0021]

[0022] in, For dynamic competitive weights, Based on the basic competitive weight, Weights are assigned to different operating conditions; This is the adjustment function for the decay phase. This represents the estimated battery health status at time t. This represents the average prediction accuracy of the k-th branch over the most recent N time steps. This is the typical working condition prototype corresponding to the k-th branch. This is a similarity adjustment parameter; S3-3: In each forward propagation, each differentiated prediction branch independently outputs a predicted value. The weighted loss is calculated using the following expression:

[0023] in, The weighted loss for the k-th branch. This is the actual RUL value; S3-4: Select the differential prediction branch with the minimum loss as the winner, expressed as:

[0024] in, The winning branch for the current sample; S3-5: Update only the network parameters of the winner branch, while keeping the parameters of the other differential prediction branches unchanged. The expression is:

[0025] in, For the network parameters of the winning branch at time t, For learning rate, The gradient of the loss function for the winner's branch; S3-6: Introduce a weight adjustment function continuously associated with SOH to adaptively adjust the competitive weights of each differentiated prediction branch at different decay stages. The expression is:

[0026] in, The decay phase adjustment function for branch k is expressed as:

[0027] in, The original dynamic competitive weights; The adjusted competition weight for the k-th branch; S3-7: Train the differentiated prediction branch according to the adjusted competition weights to obtain the trained differentiated prediction branch.

[0028] Furthermore, S4 includes: S4-1: Obtain the low-dimensional sparse feature sequence output by the sparse autoencoder and extract the low-dimensional feature vector z; S4-2: Calculate the energy proportion of each dimension of the eigenvector z, the expression is:

[0029] Where i = 1, 2, ..., d; This represents the energy percentage of the i-th feature. This represents the numerical value of the j-th feature. Represents the numerical value of the i-th feature; This represents the energy value of the i-th feature. S4-3: Calculate the sparse entropy, the expression is:

[0030] in, It is sparse entropy; S4-4: Dynamically adjust the sliding window based on sparse entropy, the expression is:

[0031] in, This represents the length of the sliding window.

[0032] Furthermore, S5 includes: S5-1: Based on the length of the sliding window Obtain the low-dimensional sparse feature sequence of the sliding window length times and input it into the differential prediction branch respectively. Each differential prediction branch outputs the prediction result including the RUL prediction value and the prediction confidence. S5-2: Define the health state of the power battery as a system state variable. ,in, This represents the estimated battery health status at time t. This represents the estimated remaining lifespan of the power battery at time t. Indicates the SOH decay rate; S5-3: Using the prediction results of each differentiated prediction branch as the observation input, construct a discrete linear state-space model, where the state equation expression is:

[0033] The expression for the observation equation is:

[0034] in, Here is the state transition matrix. For the observation matrix, For process noise, Let Q and R be the observation noise, respectively, and let Q and R be the covariance matrices of the process noise and the observation noise, respectively. S5-4: Prediction-update recursive process based on Kalman filtering, utilizing the state estimate from the previous time step. Covariance And the state-space model predicts the prior estimate at the current time step, and the prediction process expression is:

[0035]

[0036] in, For state prior estimation, To estimate the covariance a priori, Here is the state transition matrix. This is the transpose of the state transition matrix. Let be the covariance matrix of the process noise; The update process involves using the prediction results of each differentiated prediction branch as the observed values. Calculate Kalman gain Then, the prior estimate is weighted and corrected using Kalman gain to obtain the posterior estimate. The Kalman gain calculation expression is as follows:

[0037] The posterior estimation expression is:

[0038]

[0039] in, For the transpose of the observation matrix, The covariance matrix of the observed noise, For posterior state estimation, To estimate the covariance in the posterior time, It is an identity matrix.

[0040] Furthermore, S6 includes: S6-1: Calculate the original weights of each differentiated prediction branch, and normalize the original weights to obtain the final fused weights. The expression for calculating the original weights is:

[0041] in, The working condition matching degree of the k-th differentiated prediction branch is the same as the working condition matching weight in S3-2. Let k be the prediction confidence level of the k-th differential prediction branch. The standard deviation of the output corresponding to the differentiated prediction branch; The final fusion weight expression is:

[0042] in, Total number of branches; S6-2: Calculate the weighted average of the predicted values ​​from each differentiated forecast branch to obtain the final RUL predicted value, expressed as:

[0043] in, The RUL prediction value for the k-th differential prediction branch. This is the final RUL prediction value; S6-3: The fusion prediction confidence of each differentiated branch is calculated using the weighted variance formula to obtain the final prediction confidence, expressed as follows:

[0044] in, To determine the final prediction confidence level, The final prediction confidence level for the k-th differentiated prediction branch; S6-4: Based on the final RUL prediction value and final prediction confidence The 95% confidence interval is obtained, expressed as follows: .

[0045] The technical principle of this invention lies in the following: Based on the collaborative working principle of mechanism-aware sparse coding, WTA multi-branch competitive learning, dynamic sparse window, and probabilistic fusion decision-making, this invention achieves quantifiable remaining life prediction of power batteries across the entire life cycle and multiple operating conditions; the overall principle is as follows: First, multi-source information covering the entire lifecycle is constructed by collecting core battery parameters, environmental parameters, vehicle status, and offline testing data. This information is then processed using SOH segmented adaptive interpolation and differential normalization to eliminate data gaps, dimensional differences, and time-series misalignments. Next, K... Means clustering divides real-vehicle operating scenarios into typical working conditions, enabling the model to learn professionally for real-world scenarios and fundamentally solve the problem of data being disconnected from real-vehicle operating conditions. Next, a symmetric sparse autoencoder is constructed. On the encoder side, dynamic sparse masking and channel sparsity constraints are used to activate only the feature channels that are strongly related to the battery aging mechanism and suppress redundant features. On the decoder side, parameter sparsity constraints are used to achieve network lightweighting. At the same time, capacity and internal resistance physical consistency loss are introduced to bind the feature extraction process to the physical law of battery degradation, achieving the four goals of "dimensionality reduction, noise reduction, lightweighting, and mechanism consistency", and solving the problems of general sparse coding features having no physical meaning and excessive model redundancy. Subsequently, a multi-branch prediction structure corresponding to typical working conditions is constructed. Through the winner-take-all (WTA) competitive training mechanism, only the branch with the smallest prediction error for the current sample is updated with parameters, so that each branch can evolve into a specialized predictor for specific working conditions. At the same time, the RUL prediction value and confidence level are output synchronously to achieve probabilistic prediction and solve the industry pain points that a single model cannot adapt to multiple working conditions and cannot quantify the uncertainty of prediction. Then, based on the low-dimensional features output by the sparse autoencoder, the feature sparse entropy is calculated, and the entropy value is used to measure the stability of battery degradation: small entropy indicates stable degradation and a long window is used; large entropy indicates severe degradation and a short window is used. By adaptively adjusting the window length, the model maintains consistent prediction sensitivity throughout the entire battery life cycle, solving the problems of early-stage lag and late-stage distortion caused by a fixed window. Finally, Kalman filtering is used to smooth the prediction results of each branch over time to suppress noise and jumps. Then, a Bayesian fusion rule is applied, weighting both the working condition matching degree and the prediction confidence, to obtain the globally optimal RUL prediction value and its 95% confidence interval. Finally, a safety strategy is output based on a multi-condition maintenance decision library, forming a "prediction..." Correction Fusion A complete closed loop for decision-making.

[0046] The beneficial effects of this invention are as follows: 1. This invention introduces multi-source heterogeneous data acquisition and working condition adaptive clustering, enabling the model to learn in accordance with real vehicle operating scenarios; at the same time, it introduces physical consistency constraints of capacity and internal resistance in the sparse coding stage, so that the feature extraction process is deeply bound to the physical law of battery decay, avoiding the black box fitting defects caused by pure data-driven approach, and greatly improving the prediction accuracy and generalization ability under complex and variable real vehicle operating conditions. 2. This invention uses a two-stage sparsity mechanism that combines encoder channel sparsity with decoder parameter sparsity, combined with a dynamic sparse mask that adapts to the battery health status. This effectively suppresses redundant features and redundant network parameters while ensuring that core aging features are not lost, making the model lighter overall and more efficient inference. It can be stably deployed in low-computing environments on the vehicle side, meeting the real-time and lightweight requirements of the vehicle battery management system. 3. This invention uses a multi-branch probabilistic prediction structure to simultaneously output lifetime prediction results and prediction confidence, and combines Bayesian fusion rules to form a comprehensive prediction result with a reliable interval. This can intuitively reflect the stability of the prediction results and potential fluctuation risks, providing a quantifiable and reliable basis for decisions such as vehicle safety control, graded early warning, and operation and maintenance scheduling, and avoiding the decision risks caused by a single prediction value. 4. This invention dynamically adjusts the sliding window length based on sparse entropy, enabling the model to maintain appropriate temporal awareness at different battery degradation stages. Combined with an adaptive competitive learning mechanism and dynamic weight adjustment, the model can maintain stable and sensitive prediction performance in the early stage of battery health, the middle stage of degradation, and the late stage of degradation, fundamentally improving the shortcomings of traditional methods that show large differences in performance at different degradation stages. 5. This invention adopts the WTA multi-branch competitive learning architecture, which enables each prediction branch to autonomously form a specialized modeling capability for specific working conditions during training. It can automatically adapt to changes in working conditions caused by different operating environments and different usage habits, and has stronger scene adaptability and better prediction robustness compared with a single model structure. Attached Figure Description

[0047] Figure 1 This is an overall flowchart of an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the implementation steps of an embodiment of the present invention; Figure 3 This is a schematic diagram of the algorithm flow according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the data preprocessing process according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the sparse autoencoder architecture according to an embodiment of the present invention. Detailed Implementation

[0048] The following detailed description illustrates the specific implementation method: The implementation examples are basically as follows Figure 1 , Figure 2 and Figure 3 As shown: A power battery life prediction method based on mechanism-aware sparse coding and WTA prediction includes: S1: Collect multi-source operational data of the entire life cycle of new energy vehicle power batteries, preprocess the operational data to obtain the characteristic data of the power batteries, and use a preset clustering algorithm to cluster the preprocessed operational data into typical operating conditions; wherein, S1 includes: S1-1: Collect multi-source operational data throughout the entire lifecycle of new energy vehicle power batteries, including core battery parameters, operating environment parameters, vehicle status parameters, and offline testing data; among which: Battery core parameters: Collect charging and discharging current, total voltage, single cell voltage, DC internal resistance, available capacity, depth of charge, and depth of discharge parameters of the power battery. The collection frequency is 1Hz, and the data comes from the real-time monitoring module of the vehicle battery management system (BMS). Operating environment parameters: Ambient temperature, ambient humidity, altitude, and atmospheric pressure parameters are collected during vehicle operation at a frequency of 30 minutes / time. The data comes from the vehicle's environmental sensors and navigation and positioning system. Vehicle status parameters: Collect the vehicle's cumulative mileage, current driving mode (economy / standard / sport / hill), charging frequency, charging time, electric braking regeneration current, and vehicle load parameters. The data is collected every 1 minute and comes from the vehicle control unit (VCU). Offline testing data: Collects periodic capacity calibration values, AC internal resistance test values, single cell voltage consistency data, and aging fault records of power batteries. The collection cycle is once a month, and the data comes from offline maintenance and testing equipment for new energy vehicles.

[0049] All of the above data are accompanied by a unified timestamp for subsequent data time alignment processing.

[0050] S1-2: The collected multi-source operational data are sequentially processed with missing value imputation and feature normalization to obtain preprocessed feature data; such as Figure 4 As shown: Missing value imputation: A piecewise adaptive interpolation method based on battery state of health (SOH) is used to impute missing data. The appropriate interpolation method is selected according to the current SOH value of the battery. When the battery SOH > 80% (in the early stage of battery health, the degradation characteristics are gradual), the missing values ​​are filled by linear interpolation. When 60% < SOH ≤ 80% (mid-term battery degradation, with a non-linear accelerating trend in degradation characteristics), a hybrid interpolation method combining linear interpolation and cubic spline interpolation is used to fill in the missing values. When SOH≤60% (the end of battery degradation, when the degradation characteristics are drastic and fluctuate greatly), cubic spline interpolation is used to fill in the missing values.

[0051] Feature normalization: Different normalization methods are used for different types of physical quantities to eliminate the impact of dimensional differences on model training. For parameters with clearly defined physical upper and lower limits, such as charging and discharging current, total voltage, individual cell voltage, and capacity, the min-max normalization method is used to map the eigenvalues ​​to... The normalization formula for the interval is:

[0052] In the formula, x is the original eigenvalue. , These are the minimum and maximum values ​​of the feature in the training set, respectively. These are the normalized eigenvalues; For parameters such as ambient temperature, altitude, and atmospheric pressure that do not have clearly defined physical upper and lower limits, the Z-score normalization method is used. The normalization formula is as follows:

[0053] In the formula, This is the mean of the feature in the training set. is the standard deviation of this feature in the training set.

[0054] It also includes time alignment processing: using 10-second intervals as a uniform time interval, all multi-source data are aligned according to timestamps. For parameters with multiple sample values ​​within the same time interval, the average value is taken as the feature value of that time interval. For time intervals without sample values, the missing value filling method described above is used to fill in the missing values, and finally, standardized feature data with continuous time series and uniform dimensions are obtained.

[0055] S1-3: The K-means algorithm is used to cluster the preprocessed feature data into typical operating conditions. In this embodiment, there are 5 types of typical operating conditions: low temperature high rate, normal temperature cycling, high temperature static storage, intermittent discharge, and long-term storage. The clustering process is as follows: Clustering input feature selection: Five features that can significantly distinguish different operating scenarios are selected as clustering input features: average charge / discharge rate, average ambient temperature, charge / discharge time ratio, resting time ratio, and electric braking recovery intensity. Clustering parameter settings: Set the number of clusters K=5, corresponding to the five most common typical operating conditions in actual vehicle operation: low temperature high rate operation, normal temperature cycle operation, high temperature static operation, intermittent discharge operation, and long-term parking operation; set the maximum number of iterations to 100, and the convergence threshold to [missing information]. That is, the clustering process terminates when the change in centroid between two consecutive iterations is less than the convergence threshold; The clustering objective function is:

[0056] in, The clustering objective function is... The number of clusters, For the i-th cluster, Let i be the centroid of the i-th cluster. These are the feature values ​​of the preprocessed feature data; The clustering process is as follows: Step 1: Randomly select 5 samples from the preprocessed feature data as the initial cluster centroids. , , ..., ; Step 2: Calculate the Euclidean distance from each sample x to the centroids of each cluster, and assign the sample to the nearest cluster. middle; Step 3: Recalculate the centroid of each cluster. The centroid is calculated as the mean of all samples within the cluster. Step 4: Repeat steps 2 and 3 until the change in cluster centroids is less than the convergence threshold or the maximum number of iterations is reached; Clustering result processing: After clustering, the feature data of each time step are labeled with the corresponding working condition label to form a standardized dataset with working condition label, which serves as the basis for working condition matching in subsequent WTA multi-branch prediction.

[0057] In step S1 above, the technical approach of this application is as follows: First, obtain the full life cycle operation information of the power battery from four dimensions: battery body, vehicle operation, external environment, and offline testing. Then, through missing value imputation, normalization, time alignment, and typical operating condition clustering, transform discrete, heterogeneous, and data of different frequencies into a continuous, unified, and standardized feature dataset with operating condition labels. This processing method enables subsequent models to not only utilize direct health characterization parameters such as battery current, voltage, capacity, and internal resistance, but also combine external stimuli such as temperature, driving mode, charging behavior, and vehicle load, thereby more completely describing the power battery life degradation process. Compared with existing technologies, existing power battery life prediction methods usually rely only on single or a few battery parameters such as capacity, voltage, current, or internal resistance, and often use fixed interpolation, fixed normalization, and manual division of operating conditions, which are difficult to adapt to the characteristics of complex data sources, different sampling frequencies, and frequent changes in operating conditions in real vehicle operation. The difference in this application lies in the following aspects: Firstly, missing value imputation does not employ a uniform method, but rather uses linear interpolation, mixed interpolation, or cubic spline interpolation depending on the stage of SOH, so that the data repair strategy matches the battery degradation stage. Secondly, K-means is used to automatically divide the operating data into typical operating conditions such as low temperature high rate, normal temperature cycling, high temperature static storage, intermittent discharge, and long-term parking, providing a basis for operating condition matching for subsequent multi-branch prediction, and avoiding the problem of insufficient adaptability of traditional single models to complex operating conditions.

[0058] S2: Construct a sparse autoencoder consisting of an input layer, an encoder embedding dynamic sparse coding, a decoder, and an output layer. Feature data is matched through the input layer. Channel sparsity loss is applied to the encoder, and parametric sparsity loss is applied to the decoder. Physical consistency loss is also introduced. The output layer outputs a low-dimensional sparse feature sequence. The structure of the sparse autoencoder is as follows: Figure 5 As shown, S2 includes: S2-1: Construct a sparse autoencoder consisting of an input layer, an encoder, a decoder, and an output layer. The encoder is used for high-dimensional feature compression and core aging feature selection, while the decoder is used for low-dimensional feature reconstruction and lightweighting of model parameters. The input and output layers have the same dimensions, matching the dimensions of the preprocessed features, such as 32-dimensional or 45-dimensional.

[0059] S2-2: In a sparse autoencoder, the input layer receives and carries preprocessed feature data; the specific processing procedure is as follows: The number of neurons in the input layer is perfectly matched with the dimension of the preprocessed features. For example, when the dimension of the preprocessed features is 32, the input layer is set to have 32 neurons. The input layer does not perform any nonlinear transformations; it only performs dimension alignment and standardization of the feature data and inputs the feature vectors in a uniform format sequentially to the encoder according to time steps. Input features include normalized multi-source parameters such as charging and discharging current, total voltage, individual cell voltage, internal resistance, capacity, differential entropy, ambient temperature, and altitude.

[0060] S2-3: In the sparse autoencoder, the encoder embeds a dynamic sparse mask to perform high-dimensional feature compression, sparsification, and core aging feature filtering on the input layer's feature data, outputting a low-dimensional sparse feature vector. Channel sparsity loss is then applied to the output low-dimensional sparse feature vector to suppress redundant channels. Specifically, the implementation is as follows: 1. Network Structure Design: The encoder uses a 3-4 layer fully connected network with the number of neurons decreasing sequentially. For example, when the input layer is 32-dimensional, the number of encoder neurons is 32→24→16→8; when the input layer is 45-dimensional, the number is 45→36→24→12. Each layer of the encoder uses the ReLU activation function, introducing a nonlinear transformation to capture the nonlinear characteristics of battery aging.

[0061] 2. Dynamic Sparse Mask Embedding: A dynamic sparse mask matrix is ​​embedded at the output of each layer of the encoder, and the sparsity is adaptively adjusted segment by segment according to the current state of battery health (SOH). When SOH > 80% (early stage of battery health, with gradual degradation), the sparsity rate is set to 0.05-0.1, activating only a small number of characteristic channels related to basic aging. When 60% < SOH ≤ 80% (mid-stage of battery degradation, accelerated degradation), the sparsity rate is set to 0.1-0.2, and the number of activated channels is moderately increased to capture enhanced nonlinear characteristics; When SOH≤60% (the end of battery degradation, with severe degradation), the sparsity rate is set to 0.2-0.3 to further activate more feature channels to fully characterize complex aging modes.

[0062] Dynamic sparse masking is achieved by setting the output of neurons below a threshold to zero, retaining only the feature channels that are strongly correlated with core aging processes such as lithium dendrite growth, electrolyte decomposition, and SEI film thickening.

[0063] 3. Channel Sparsity Loss Construction: Apply channel sparsity loss constraints to the output features of each layer of the encoder to suppress redundant feature channels. The expression for channel sparsity loss is as follows:

[0064] in, This is a penalty for channel sparsity. For the first The layer encoder outputs a feature vector. This represents the total number of layers in the encoder. It is an L1 norm. For the first The number of neurons in the layer encoder; The role of channel sparsity loss is to promote the sparse representation of the encoder's output features, so that the output of most neurons approaches zero and only a few core neurons are activated, thereby achieving effective compression of feature dimensions.

[0065] S2-4: In a sparse autoencoder, the decoder reconstructs the low-dimensional sparse features output by the encoder back to the original dimension, and applies a parametric sparse loss to the decoder weights to prune redundant parameters. Specifically: 1. Network Structure Design: The decoder uses the same number of fully connected networks as the encoder, with the number of neurons increasing in an incremental fashion, symmetrical to the encoder. For example, if the encoder's sequence is 32→24→16→8, the decoder's sequence is 8→16→24→32; if the encoder's sequence is 45→36→24→12, the decoder's sequence is 12→24→36→45. Each layer of the decoder uses the ReLU activation function, and the output layer uses a linear activation function to ensure that the numerical range of the reconstructed features is consistent with the input.

[0066] 2. Construction of sparse parameter loss: Apply sparse parameter loss constraints to the weight matrices of each layer of the decoder, prune invalid or redundant connection weights, and achieve lightweight model parameters. The expression for sparse parameter loss is:

[0067] in, For parameter sparsity penalty term, For the first Layer decoder weight matrix, This represents the number of decoder layers. It is an L2 norm matrix. For the first The number of input neurons in the layer decoder. For the first The number of output neurons in the layer decoder For the first The connection weights from the i-th input neuron to the j-th output neuron in the layer decoder; The purpose of parameter sparsity is to penalize parameters with large absolute values ​​in the weight matrix, pushing the weights to converge to zero, thereby pruning invalid connections and reducing the number of model parameters and computational complexity.

[0068] S2-5: Introduce physical consistency constraints of capacity and internal resistance into the loss functions of the encoder and decoder, expressed as follows:

[0069] in, This is a penalty for physical consistency. , For weighting coefficients, the preferred value is... , ; , The input features are estimated values ​​for capacity and internal resistance. , These are the reconstructed capacitance and internal resistance values; Physical consistency loss constrains the consistency of the core health parameters of the battery before and after encoding, enabling the features learned by the sparse autoencoder to be deeply bound to the physical process of battery degradation, thereby improving the interpretability of features and the generalization ability of the model.

[0070] S2-6: The reconstruction loss, channel sparsity loss, parameter sparsity loss, and physical consistency loss are fused together, and the expression is:

[0071] in: For the total loss function, For reconstruction loss, mean squared error (MSE) is used to measure the difference between the input features and the reconstructed features, expressed as:

[0072] in, As input features, To reconstruct features, The number of samples; , , The regularization coefficient is denoted by , and the preferred value is . , , The optimal value can be determined through 5-fold cross-validation; S2-7: The output layer directly outputs the low-dimensional sparse feature sequence and reconstruction error reconstructed by the decoder; the number of neurons in the output layer is the same as that in the input layer, and it outputs the complete feature vector after reconstruction, which is used to calculate the reconstruction loss and physical consistency loss; at the same time, it outputs the low-dimensional sparse feature vector of the last layer of the encoder to form a low-dimensional sparse feature sequence, which is used as the input for the WTA multi-branch prediction in the subsequent S3 step. S2-8: The sparse autoencoder is trained and optimized using an improved Adam optimizer. The specific implementation is as follows: Data partitioning: The dataset with working condition labels after step S1 is divided into training set, validation set and test set in a ratio of 7:2:1; Optimizer settings: An improved Adam optimizer is used, with the learning rate dynamically adjusted based on the battery state of equilibrium (SOH): when SOH ≥ 80%, the learning rate is set to 0.001; when SOH < 60%, the learning rate is reduced to 0.0001; the first-order moment decay coefficient of the optimizer is also specified. =0.9, second-order moment attenuation coefficient =0.999, numerical stability coefficient ; Training process: Set the batch size to 64-128, the training rounds to 100-150, and use an early stopping mechanism to prevent overfitting. Terminate training when the validation set loss no longer decreases for 10 consecutive rounds. Hyperparameter optimization: The optimal regularization coefficient, sparsity range, learning rate decay factor and other hyperparameters are determined by 5-fold cross-validation, and finally the trained mechanism-aware sparse autoencoder is obtained.

[0073] In step S2 above, the technical idea of ​​this application is to compress the high-dimensional multi-source features obtained in S1 using a sparse autoencoder, and to introduce dynamic sparse mask and channel sparse loss on the encoder side, and introduce parameter sparse loss on the decoder side. At the same time, the feature reconstruction process is constrained by the physical consistency loss of capacity and internal resistance, so that the final output of the model is a feature sequence that is both low-dimensional and sparse, and can reflect the real aging mechanism of the battery.

[0074] Compared to existing technologies, traditional autoencoders or sparse coding methods primarily focus on reconstruction errors and mathematical dimensionality reduction, easily extracting abstract features lacking physical interpretation. This results in models that, while able to fit training data, have insufficient generalization ability under different batteries, operating conditions, or degradation stages. This application differs in several ways: First, the sparsity rate is not fixed but dynamically correlated with SOH (State of Health). A lower sparsity rate is used in the early stages of battery health, retaining only basic aging features; the sparsity rate is increased in the later stages of degradation to enhance the ability to express complex aging patterns. Second, feature channel sparsity is applied on the encoder side, and model parameter sparsity is applied on the decoder side, achieving the dual effects of "feature lightweighting" and "model lightweighting." Third, physical consistency constraints on capacity and internal resistance are introduced, establishing a correspondence between low-dimensional features and aging mechanisms such as lithium dendrite growth, SEI film thickening, and electrolyte decomposition, thereby improving the interpretability and engineering reliability of the prediction process.

[0075] S3: Construct multiple differentiated prediction branches corresponding to typical working conditions. Each differentiated prediction branch performs RUL prediction based on a low-dimensional sparse feature sequence as input, and updates only the parameters of the branch with the smallest prediction error for the current sample based on the WTA competitive training mechanism, resulting in multiple trained differentiated prediction branches; wherein, S3 includes: S3-1: Construct differentiated prediction branches adapted to typical operating conditions. Each differentiated prediction branch independently completes time-series feature modeling and lifetime prediction. The input is a low-dimensional sparse feature sequence output by a sparse autoencoder. , Let be the sliding window length, and d be the feature dimension; each differential prediction branch includes a bidirectional LSTM layer, an attention layer, and a fully connected output layer, where: A bidirectional LSTM layer is used to capture the long-term and short-term temporal dependencies in the battery degradation process, while utilizing both forward and reverse temporal information to improve prediction accuracy. The expression is as follows:

[0076] in, Let k be the hidden state sequence of the k-th branch. The parameters of the LSTM network for the k-th branch include the weights and biases of the input gate, forget gate, cell state gate, and output gate. The bidirectional LSTM layer consists of a forward LSTM and a backward LSTM. The forward LSTM processes the feature sequence in forward time order, and the backward LSTM processes the feature sequence in reverse time order. Finally, the hidden states in the two directions are concatenated to obtain the complete hidden state sequence.

[0077] The attention layer is used to adaptively assign weights to the hidden states at different time steps, highlighting temporal information that is more important for lifetime prediction. Its calculation process consists of two steps: The first step is to calculate the attention score, expressed as:

[0078] The second step is to generate the context vector, expressed as follows:

[0079] in, For attention weight vectors, , These are the trainable parameters for the attention layer. For context vectors, Let be the hidden state at the i-th time step in the k-th branch; The fully connected output layer is used to map the context vector to RUL predictions and prediction confidence (standard deviation estimates), as shown in the following expression:

[0080]

[0081] in, , These are the trainable parameters for the fully connected layer. The RUL prediction value output by the k-th branch. This represents the corresponding prediction confidence level, which is the standard deviation of the predicted values ​​within the current window. The smaller the value, the more reliable the prediction result. This represents the average predicted value of the k-th branch within the current window; S3-2: Based on the basic competitive weight and the working condition matching weight, construct the dynamic competitive weight for each differentiated prediction branch. This weight is used to measure the fit of each differentiated prediction branch to the current sample. The expression is:

[0082] in, For dynamic competitive weights, Based on the basic competitive weight, Weights are assigned to different operating conditions; The basic competitive weights are adaptively adjusted based on the historical prediction accuracy of the branch in the current decay phase, and the expression is:

[0083] in, This is a decay phase adjustment function used to weakly guide branch weights at different battery aging stages. This represents the estimated battery health status at time t. Let N be the average prediction accuracy of the k-th branch over the most recent N time steps, where N is preferably between 20 and 50. The working condition matching weight is calculated based on the Euclidean distance between the current input features and the typical working condition prototype corresponding to the branch, and the expression is:

[0084] in, The centroid of the k-th cluster, obtained from the clustering in step S1, is determined to be the typical working condition prototype corresponding to the k-th branch. The similarity adjustment parameter is preferably set to 0.1-0.5, which is used to control the decay rate of the weights. The greater the working condition matching weight, the more similar the current sample is to the typical working condition corresponding to that branch, and the higher the reliability of the prediction result of that branch.

[0085] S3-3: In each forward propagation, each differentiated prediction branch independently outputs a predicted value. The weighted loss is calculated using the following expression:

[0086] in, The weighted loss for the k-th branch. This is the actual RUL value; S3-4: Employing a winner-take-all (WTA) competition mechanism, the differential prediction branch with the minimum weighted loss is selected as the winner, expressed as:

[0087] in, The winning branch for the current sample; The total number of differentiated prediction branches is equal to the number of typical operating conditions, and the preferred value is 5, which corresponds to five typical operating conditions: low temperature high rate, normal temperature cycle, high temperature static, intermittent discharge and long-term parking.

[0088] The winning branch is the branch that predicts the most accurately and best fits the current sample.

[0089] S3-5: Perform backpropagation and gradient update only on the network parameters of the winner branch, while keeping the parameters of the other differential prediction branches unchanged. The expression is:

[0090] in, For the network parameters of the winning branch at time t, The learning rate is preferably set to 0.001. The gradient of the loss function for the winner's branch; By updating only the parameters of the winning branch, each branch will gradually focus on the working conditions it is best at through repeated competition, forming specialized modeling capabilities and avoiding mutual interference between branches.

[0091] S3-6: To further improve the model's predictive performance throughout the entire lifecycle of the power battery, a weight adjustment function continuously correlated with the battery's state of health (SOH) is introduced. This function adaptively adjusts the competing weights of each differentiated prediction branch at different decay stages. The expression is as follows:

[0092] in, The original dynamic competition weights are calculated in step S3-2; The adjusted competition weight for the k-th branch; The decay phase adjustment function for branch k is expressed as:

[0093] Among them, the safety threshold-related branch corresponds to the branches that easily lead to rapid battery aging, such as low temperature and high rate operation, and high temperature and static storage. This adjustment function only serves as a weak guide for the branch competition process and does not change the basic principle of the WTA mechanism of "using the branch with the smallest prediction error as the update object".

[0094] S3-7: Repeat S3-3 to S3-6 to iterate training on all training samples until the training termination condition is met, and obtain the differentiated prediction branch after training is completed. The training termination condition includes that the validation set loss no longer decreases for 10 consecutive rounds, the preset maximum number of training rounds is reached (preferably 100-150 rounds), and the prediction accuracy of each branch under the corresponding typical working condition reaches the preset threshold.

[0095] In step S3 above, the technical idea of ​​this application is as follows: Differentiated prediction branches are constructed for different typical working conditions. Each branch includes a bidirectional LSTM layer, an attention layer and a fully connected output layer, which are used to learn the temporal decay law under the corresponding working condition from the low-dimensional sparse feature sequence. During the training process, a WTA competition mechanism is introduced. Each branch outputs the prediction result at the same time, but only the branch with the smallest weighted loss is selected as the winner and its parameters are updated, so that each branch gradually forms a professional prediction ability in repeated competition.

[0096] Compared to existing technologies, traditional power battery RUL prediction methods typically employ a single LSTM, a single Transformer, or a single regression model to uniformly model all operating conditions, which easily leads to interference between samples from different operating conditions. For example, the capacity drop patterns under low temperature and high rate conditions, the SEI film thickening patterns under high temperature and static conditions, and the self-discharge patterns under long-term parking conditions differ significantly. If a single model is used for uniform learning, the model may perform reasonably well on average, but the prediction error will be large under extreme operating conditions. The difference in this application is that it uses a WTA mechanism to allow the low temperature and high rate branch, the high temperature and static branch, and the normal temperature cycling branch to learn their own most effective degradation patterns, avoiding parameter interference caused by all branches being updated by the same sample at the same time. At the same time, the competitive weights also combine the operating condition matching degree and historical prediction accuracy, so that the branch selection not only depends on the instantaneous error but also considers the similarity between the current sample and typical operating conditions, thereby improving the prediction stability and adaptability under complex real-world vehicle operating conditions.

[0097] S4: Calculate the feature sparse entropy based on the low-dimensional sparse feature sequence output by the sparse autoencoder and dynamically adjust the sliding window; where S4 includes: S4-1: Obtain the low-dimensional sparse feature sequence output by the sparse autoencoder and extract the low-dimensional feature vector z; the specific implementation is as follows: Input source: The low-dimensional sparse feature sequence output from the last layer of the sparse encoder in step S2. Redundant information has been removed from this sequence, and only features strongly correlated with the battery core aging process are retained. Feature vector definition: Extracting the feature vector at the current time t , where d is the dimension of the low-dimensional sparse feature, which is consistent with the number of neurons in the last layer of the encoder of the sparse autoencoder; This represents the value of the i-th dimension feature at the current moment, corresponding to the characterization intensity of a specific aging mode of the power battery, such as the degree of lithium dendrite growth, electrolyte decomposition rate, and SEI film thickness.

[0098] S4-2: Calculate the energy percentage of each dimension of the feature vector z to quantify the contribution of each aging mode to the current battery state. The expression is:

[0099] Where i = 1, 2, ..., d; This represents the energy percentage of the i-th feature. This represents the numerical value of the j-th feature. Represents the numerical value of the i-th feature; The energy value represents the i-th feature. The larger the energy value, the more dominant the aging mode corresponding to that dimension is at the current moment. S4-3: Calculating Feature Sparse Entropy Based on Feature Energy Distribution The entropy value is used to quantify the stability and complexity of the battery degradation mode, and the expression is:

[0100] in, It is the sparse entropy; the physical meaning of feature sparsity is: when When the value approaches 0, the characteristic energy is highly concentrated in a few dimensions, indicating that the battery degradation mode is simple, the change is stable, and it is in the early stage of health or the stable degradation stage. when Approaching When the characteristic energy is uniformly distributed across all dimensions, it indicates that the battery degradation mode is complex and multiple aging mechanisms are coupled, indicating that it is in the late stage of degradation or a stage of drastic changes in operating conditions.

[0101] S4-4: Based on sparse entropy, dynamically adjust the sliding window according to preset segmentation rules. The expression is:

[0102] in, The length of the sliding window; the design basis and technical effects of each window length are as follows: when During the steady decay phase, a long sliding window of length 10 is used to capture the long-term decay trend of the battery, suppress instantaneous noise interference, and improve the stability of the prediction results. when During the transitional decay phase, a medium sliding window of length 5 is used to balance trend capture capability and response speed, adapting to the gradual changes in the decay pattern. when During the severe degradation phase, a short sliding window of length 3 is used to quickly respond to sudden changes in battery state, avoid prediction lag and distortion caused by long windows, and improve the prediction accuracy of critical degradation phases.

[0103] In step S4 above, the technical idea of ​​this application is as follows: using the low-dimensional sparse features output by the sparse autoencoder, the energy ratio of each dimension of the features is calculated, and the feature sparse entropy is further calculated to characterize the concentration and complexity of the current battery degradation mode; when the feature energy is concentrated and the degradation mode is stable, a longer sliding window is used to enhance the trend capture capability; when the feature energy is dispersed and the aging mechanism is coupled more strongly, the sliding window is shortened to improve the response speed to abrupt changes.

[0104] Compared to existing technologies, current lifespan prediction methods mostly employ fixed-length sliding windows, such as taking a fixed number of recent cycles or time steps as input. This approach may result in windows that are too short during the stable decay phase, leading to insufficient trend information; conversely, windows may be too long during the late decay phase or abrupt changes in operating conditions, causing prediction lag. The difference in this application is that the sliding window length is not manually fixed, but dynamically determined by the energy distribution and sparse entropy of low-dimensional sparse features, allowing the window length to automatically adjust as the battery aging state changes. Therefore, the model can use a long window to suppress noise in the early healthy phase, balance trend and response speed during the transition phase, and quickly capture abnormal changes through a short window during the severe decay phase, thereby improving the continuity of the entire lifespan prediction and the sensitivity to critical decay stages.

[0105] S5: Extract the low-dimensional sparse feature sequence of the adjusted sliding window length, input it into the trained differential prediction branch, and output the RUL prediction value and prediction confidence of each differential prediction branch. Then, call Kalman filtering to correct the prediction results. S5 includes: S5-1: Based on the length of the sliding window The process involves obtaining low-dimensional sparse feature sequences at time steps of the sliding window length and inputting them into differential prediction branches. Each differential prediction branch outputs a prediction result including the RUL prediction value and prediction confidence. The specific implementation is as follows: Input source: Low-dimensional sparse feature sequence output by the S2 step mechanism-aware sparse autoencoder, which has completed feature dimensionality reduction and core aging feature screening; Window truncation rule: Take the current time t as the endpoint and truncate the continuous window backwards. The low-dimensional sparse features at each time step form the input feature sequence. , where d is the dimension of the low-dimensional sparse feature; Branch prediction output: The input feature sequence is fed into each of the differential prediction branches. Each branch independently completes time series modeling and prediction, and the output includes the RUL prediction value. With prediction confidence The original prediction results are given, where k = 1, 2, ..., K, and K is the total number of differential prediction branches.

[0106] S5-2: To construct a state-space model that conforms to the physical laws of battery degradation, the health state of the power battery is defined as a system state variable. ,in, This represents the estimated battery health status at time t. This represents the estimated remaining lifespan of the power battery at time t. Indicates the SOH decay rate; The three parameters mentioned above are selected as system state variables based on the following: SOH represents the current health status of the battery, RUL represents the remaining battery life, and the decay rate represents the aging trend of the battery. The combination of the three can comprehensively and accurately describe the dynamic health status of the battery, which meets the requirements of Kalman filtering for the completeness of state variables.

[0107] S5-3: Using the prediction results of each differentiated prediction branch as the observation input, construct a discrete linear state-space model, where the state equation expression is:

[0108] The expression for the observation equation is:

[0109] in, Here is the state transition matrix. For the observation matrix, The process noise is used to characterize the random fluctuations during battery aging, and it follows a Gaussian distribution with a mean of 0 and a covariance matrix of Q. The observation noise is used to characterize the prediction error of each prediction branch, and follows a Gaussian distribution with a mean of 0 and a covariance matrix of R; Q and R are the covariance matrices of the process noise and the observation noise, respectively. S5-4: Based on the constructed state-space model, perform the Kalman filter prediction-update recursive process, utilizing the state estimate from the previous time step. Covariance The state-space model predicts the prior state estimate and prior covariance matrix at the current time step. The prediction process is expressed as follows:

[0110]

[0111] in, For state prior estimation, To estimate the covariance a priori, Here is the state transition matrix. This is the transpose of the state transition matrix. Let be the covariance matrix of the process noise; The update process involves using the prediction results of each differentiated prediction branch as the observed values. Calculate Kalman gain Then, the prior estimate is weighted and corrected using Kalman gain to obtain the posterior estimate. The Kalman gain calculation expression is as follows:

[0112] The physical meaning of Kalman gain is: when the observation noise is low, the gain is large, and the correction effect of the observation on the state estimate is stronger; when the process noise is low, the gain is small, and the reliability of the prior state estimate is higher.

[0113] The posterior state estimate is calculated using the following expression:

[0114] The updated posterior estimate of the covariance matrix is ​​expressed as follows:

[0115] in, For the transpose of the observation matrix, The covariance matrix of the observed noise, For posterior state estimation, To estimate the covariance in the posterior time, It is the identity matrix; Through the above recursive process, Kalman filtering can automatically fuse the physical trend of battery degradation with the observation results of multi-branch prediction, effectively suppressing instantaneous noise and abnormal fluctuations, making the RUL prediction curve smoother and more continuous, avoiding false warnings, and at the same time outputting the uncertainty of state estimation, providing a reliable basis for subsequent Bayesian probabilistic fusion.

[0116] In step S5 above, the technical idea of ​​this application is as follows: the sliding window features determined in S4 are input into the trained multiple differentiated prediction branches to obtain the RUL prediction value and prediction confidence of each branch; then, the state space model is constructed using SOH, RUL and SOH decay rate as system state variables, and the multi-branch prediction results are recursively corrected in the time dimension by Kalman filtering so that the prediction results not only conform to the model observation output, but also conform to the physical law of battery health state evolution over time.

[0117] Compared to existing technologies, current methods often directly use the neural network output as the final RUL prediction value, which is easily affected by instantaneous sampling noise, abnormal operating condition disturbances, or single prediction errors of the model, leading to unreasonable jumps in the prediction curve. This application differs in that it does not simply use the branch output as the final result, but instead uses it as the observation input to a Kalman filter framework. It integrates the previous health status, current observation results, and noise covariance information through a prediction-update recursive process. This preserves the neural network's ability to learn complex nonlinear decay laws while utilizing the state-space model to constrain the continuity of RUL changes, reducing spurious fluctuations and false alarms, and improving the smoothness and engineering usability of the prediction curve.

[0118] S6: Using a pre-defined Bayesian fusion rule, the matching degree between typical working conditions and differentiated prediction branches, and the confidence level of the output of differentiated prediction branches, the prediction results of each differentiated prediction branch are weighted and fused to obtain the final RUL prediction value and prediction confidence level; wherein, S6 includes: S6-1: Calculate the fusion weight of each differentiated prediction branch based on both the working condition matching degree and the prediction confidence, so that the branch that is more suitable for the current working condition and whose prediction results are more reliable receives a higher weight contribution; the original fusion weight of each differentiated prediction branch is determined by the product of the working condition matching degree and the prediction confidence, and the expression is:

[0119] in, The original fusion weights for the k-th differential prediction branch are... The working condition matching degree of the k-th differentiated prediction branch is the same as the working condition matching weight in S3-2. Let k be the prediction confidence level of the k-th differential prediction branch. The standard deviation of the output corresponding to the differentiated prediction branch; To ensure that the sum of the fusion weights of all branches is 1, the original fusion weights are normalized, and the final fusion weight expression is:

[0120] in, Total number of branches; S6-2: Based on the normalized fusion weights, a weighted average is calculated for the predicted values ​​of each differentiated prediction branch to obtain the final RUL predicted value. The expression is as follows:

[0121] in, The RUL prediction value for the k-th differential prediction branch. This is the final RUL prediction value; This weighted average method can fully utilize the specialized prediction advantages of each branch: when the vehicle is in normal temperature cycle conditions, the normal temperature cycle branch will receive the highest weight and dominate the final prediction result; when the vehicle is in low temperature high magnification conditions, the weight of the low temperature high magnification branch will be automatically increased to ensure prediction accuracy under complex conditions.

[0122] S6-3: The fusion prediction confidence of each differentiated branch is calculated using the weighted variance formula to obtain the final prediction confidence. This takes into account both the prediction uncertainty of each branch itself and the dispersion of prediction results between branches. The expression is:

[0123] in, To determine the final prediction confidence level, The final prediction confidence level for the k-th differentiated prediction branch; The squared deviation term between the predicted value of the k-th branch and the final fused value is used to quantify the degree of dispersion between the prediction results of different branches; The technical significance of introducing the squared deviation term is that when the prediction results of each branch are highly consistent, the deviation term is small, and the uncertainty after fusion is mainly determined by the prediction variance of each branch itself; when the prediction results of each branch are significantly different, the deviation term will increase significantly, reflecting the overall uncertainty of the model prediction under the current working conditions, and avoiding decision-making risks caused by branch divergence.

[0124] S6-4: Based on the final RUL prediction value and final prediction confidence The 95% confidence interval is obtained, expressed as follows: 1.96 is the 95th percentile of the standard normal distribution.

[0125] In step S6 above, the technical approach of this application is as follows: based on the Kalman filter correction, a Bayesian fusion rule is further used to perform spatially weighted fusion of the results of each differentiated prediction branch. The fusion weights consider two factors simultaneously: first, the degree of matching between the current sample and the typical working conditions corresponding to each branch; and second, the confidence level of the prediction results output by each branch. Therefore, the branch that is more suitable for the current working conditions and has lower prediction uncertainty occupies a higher weight in the final RUL prediction.

[0126] Compared to existing technologies, traditional multi-model fusion methods typically employ simple averaging, fixed weights, or selection of only a single optimal model output, making it difficult to simultaneously consider both operating condition adaptability and prediction reliability. When vehicle operating conditions change, fixed-weight fusion cannot adjust the contributions of each model in a timely manner; while selecting only a single branch may ignore the effective information provided by other branches. This application differs in that it achieves dynamic allocation of fusion weights through a two-dimensional weight calculation method of "operating condition matching degree + prediction confidence degree." Simultaneously, it introduces the inherent uncertainty of each branch and the prediction bias between branches into the final confidence degree calculation, ensuring that the final output includes not only the RUL point estimate but also the confidence interval, thereby reflecting the reliability of the prediction results and providing a probabilistic basis for subsequent maintenance decisions.

[0127] S7: Construct a maintenance decision rule base with multiple conditions linked. Based on the final RUL prediction value and prediction confidence, the rule matching is triggered with the maintenance decision rule base to output the maintenance decision result. Specifically, based on the fused RUL prediction value and its 95% confidence interval, combined with the current operating condition type, battery health stage (SOH interval) and prediction confidence of each branch, a maintenance decision rule base with multiple conditions linked is established.

[0128] This rule base divides the remaining battery life into multiple risk level ranges, with each range associated with a set of triggering conditions (such as RUL threshold, confidence lower limit, specific operating condition identifier, SOH range, etc.) and corresponding decision outputs. Decision outputs cover various types, including usage strategy adjustments, testing and calibration reminders, system diagnostic warnings, and operation and maintenance scheduling suggestions, and can be flexibly configured according to battery model, vehicle type, or user needs.

[0129] For example: When the predicted RUL is greater than 1000 cycles and the system is operating under ambient temperature cycling conditions, it is recommended to perform capacity calibration every 100 cycles. When the RUL is between 500-1000 cycles, the confidence level is ≥90%, and a low temperature environment is detected, the system pushes a usage suggestion to "reduce the charging rate to 0.8C and increase the battery warm-up time". When the RUL drops to 200-500 cycles, the confidence level is ≥95%, and high-temperature fast charging occurs frequently, the system triggers a reminder that "the fast charging temperature is too high, it is recommended to adjust the fast charging power to 0.6C" and starts the thermal management system self-check. When the RUL is below 100 cycles and the confidence level is ≥98%, the system will forcibly trigger a maintenance alarm regardless of the operating conditions, limit the vehicle's output power, and guide the user to return to the site for maintenance immediately.

[0130] The above thresholds and recommendations can be dynamically adjusted according to the specific battery type (such as ternary lithium, lithium iron phosphate), vehicle application (passenger car, commercial vehicle) and safety level requirements, thereby realizing an integrated closed loop of prediction and decision-making, and improving engineering practicality and safety.

[0131] In step S7 above, the technical idea of ​​this application is to input the final RUL prediction value, prediction confidence, current operating condition type, SOH range and branch prediction reliability obtained in the above steps into the maintenance decision rule base, and output decision results such as usage strategy adjustment, detection and calibration reminder, system diagnosis warning or forced maintenance alarm through multi-condition linkage matching, thereby realizing closed-loop control from life prediction to operation and maintenance decision.

[0132] Compared to existing technologies, current power battery life prediction schemes typically only output Remaining Life (RUL) values ​​or State of Health (SOH) estimates, lacking a decision-making mechanism that integrates with actual vehicle control, maintenance reminders, and safety management. This makes it difficult to directly translate prediction results into engineering applications. The difference in this application is that maintenance decisions are not triggered solely by a single RUL threshold, but rather comprehensively consider the RUL range, confidence level, operating condition type, and SOH stage. For example, in low-temperature environments, the system not only determines the remaining lifespan but also provides usage suggestions such as reducing the charging rate and increasing preheating time; when high-temperature fast charging is frequent and RUL decreases, a thermal management self-check can be triggered; when RUL is below the safety threshold and the confidence level is high, a maintenance alarm and power limitation are forcibly triggered. Therefore, this application achieves a complete closed loop of "data acquisition—mechanism feature extraction—lifespan prediction—uncertainty assessment—maintenance decision-making," enhancing the practical application value and safety assurance capabilities of power battery life prediction methods.

[0133] In summary, this application does not simply combine sparse coding, LSTM prediction, WTA mechanism, Kalman filtering, and Bayesian fusion in parallel. Instead, it establishes a progressively layered technical chain around the degradation process of the entire life cycle of a power battery: First, it addresses the problem of complex and heterogeneous input data through multi-source data fusion and clustering of typical operating conditions; second, it solves the problems of high-dimensional feature redundancy and the lack of physical meaning of black-box features through mechanism-aware sparse autoencoders; third, it addresses the problems of large differences in degradation patterns under different operating conditions and insufficient generalization ability of single models through WTA multi-branch prediction; subsequently, it addresses the problems of prediction lag, instantaneous fluctuations, and insufficient quantification of uncertainty through sparse entropy dynamic windows, Kalman filtering, and Bayesian fusion; finally, it realizes the transformation of prediction results into engineering operation and maintenance measures through a multi-condition maintenance decision rule base. Therefore, compared with existing technologies, this application has substantial improvements in adaptability to complex operating conditions, full life cycle prediction accuracy, lightweight model deployment, prediction interpretability, and closed-loop maintenance decision-making capabilities. To better illustrate the principles of the technical solution of this application, the specific process is described in detail through the following two examples: Example 1: Passenger Vehicle Power Battery Degradation and RUL Prediction 1. Actual usage environment Application scenario: A certain brand of pure electric passenger vehicle (battery type: ternary lithium battery, rated capacity 50Ah, rated voltage 350V), daily driving scenarios include urban commuting (normal temperature cycle), short-distance winter trips (low temperature high rate), and long-distance highway trips (intermittent discharge), which needs to be adapted to the deployment of vehicle-mounted BMS (computing power ≤1TOPS).

[0134] 2. Specific Implementation (1) Data acquisition and preprocessing A. Data collection frequency: Battery core parameters (current / voltage / internal resistance) and vehicle status parameters (driving mode / mileage) are collected every 10 seconds; environmental parameters (temperature / humidity) are collected every hour; and offline test data (capacity calibration) is collected once a month. B. Preprocessing details: Missing value filling adopts "SOH>80% linear interpolation, SOH≤80% cubic spline interpolation"; after feature normalization, 32 core features are retained (including differential entropy of charge and discharge curve, internal resistance change rate, temperature fluctuation coefficient, etc.); five typical scenarios are obtained by working condition clustering, and they are passed into the model through working condition identification vector (5-dimensional one-hot code).

[0135] (2) Training of decay-aware sparse autoencoders A. Model parameters: Input layer 32-dimensional, encoder neurons 32→24→16→8, decoder 8→16→24→32; dynamic sparse mask sparsity range 0.05-0.3, adjusted in real time with SOH (triggered by SOH estimation value collected in real time by BMS). B. Training settings: batch size=64, training epoch=100, initial learning rate 0.001, automatically reduced to 0.0001 when SOH≤60%; cross-validation determined. =0.01、 =0.005, the parameter size after model training is ≤6000.

[0136] (3) Deployment of WTA multi-branch probability prediction module A. Branch settings: 5 scene branches, each branch has 64 LSTM hidden layer nodes and 32 Attention dimensions; B. Competitive Training: The WTA loss function only updates the optimal branch parameters. In winter (ambient temperature ≤ -10℃), the weight of the "low temperature high magnification branch" is increased to 1.2, and in summer (ambient temperature ≥ 40℃), the weight of the "high temperature static branch" is increased to 1.3. C. Dynamic window: The sparse entropy threshold is set to 0.3 and 0.7. The window size is 10 cycles for urban commuting scenarios (sparse entropy ≈ 0.2) and 3 cycles for winter short-distance scenarios (sparse entropy ≈ 0.8).

[0137] (4) Integration of decision output A. Output results: Real-time output of RUL value (e.g., "Remaining life 235 cycles") and 95% confidence interval (e.g., "220-250 cycles"). When RUL < 50 cycles, push "Recommended charging rate reduced to 0.8C to avoid low-temperature fast charging" to the vehicle terminal via BMS. B. Online update: Model parameter fine-tuning is triggered every 10 cycles, optimizing sparse mask and WTA branch weights based on the latest collected data.

[0138] Example 2: Commercial Vehicle Power Battery Degradation and RUL Prediction 1. Actual usage environment Application scenario: A certain brand of electric bus (battery type: lithium iron phosphate battery, rated capacity 252Ah, rated voltage 532.5V), mainly used for urban passenger transport, with operating conditions covering "stop-at-station" (normal temperature cycle) for urban loop passenger transport, midday fast charging (high temperature static placement), nighttime after returning to the depot (long-term parking), and intermittent electric braking (intermittent discharge) when entering and leaving stations. It needs to withstand ambient temperature of 0℃-50℃ and support data interaction with a remote battery monitoring platform.

[0139] 2. Specific Implementation: (1) Data acquisition and preprocessing A. Data collection frequency: Battery core parameters (charging / discharging current / voltage / internal resistance, electric braking regeneration current) 1Hz (to meet the high-frequency start-stop data requirements of "stop-at-station"); Environmental parameters (temperature / humidity) 30min / time (to accurately capture high temperature fluctuations during midday fast charging); Vehicle status parameters (passenger capacity, station entry / exit signs, electric braking trigger frequency) 1min / time; Offline testing data (capacity calibration, electric braking system operating condition testing) once a month (in line with the centralized maintenance cycle of public transportation). B. Preprocessing Details: Missing value imputation; linear interpolation is used for "stop-at-station" conditions (high-frequency start-stop); zero-order hold interpolation is used for long-term parking at night; and cubic spline interpolation is used for midday fast charging conditions. Feature engineering: three bus-specific features are added: "electrical braking recovery efficiency," "fast charging temperature rise rate," and "inter-station discharge depth," expanding the feature dimensions to 45 (including original core features + bus-specific features). Condition clustering: four core conditions are clustered (stop-at-station ambient temperature cycle, midday fast charging high temperature standby, long-term parking at night, and intermittent discharge upon entering and exiting stations), and the condition identifier vector is passed into the model through a 4-dimensional unique hot code. Data synchronization: the system is linked with the bus operation and dispatch system to align dispatch data such as "departure time, arrival time, and fast charging reservation time" with battery data time, supporting accurate scenario identification.

[0140] (2) Training of decay-aware sparse autoencoders A. Model Parameters: Input layer 45-dimensional (adapted to 45-dimensional bus-specific features), encoder neurons decrease according to the trend of "45→36→24→12", decoder neurons increase according to the trend of "12→24→36→45"; the sparsity range of the dynamic sparse mask is adjusted to 0.1-0.35 (lithium iron phosphate batteries have a long cycle life and more gradual decay, and the overall sparsity is higher than that of passenger cars, ensuring that core features are not lost), the sparsity triggering condition is associated with "number of fast charging times and frequency of electric braking" (when fast charging ≥100 times or electric braking triggers ≥300 times per day, the sparsity is increased to 0.3-0.35); B. Training settings: batch size=128 (adapting to massive bus operation data), training epoch=120, initial learning rate 0.0015 (for lithium iron phosphate batteries with larger data volume, appropriately increasing the initial learning rate accelerates convergence), automatically decreasing to 0.0002 when SOH≤70% (the retirement threshold for bus batteries is higher than that for passenger vehicles); the two-order sparse regularization coefficients are determined through cross-validation. =0.012、 =0.006, the parameter size after model training is ≤7500 (meets the real-time computing requirements of the remote monitoring platform).

[0141] (3) Deployment of WTA multi-branch probability prediction module A. Branch Setup: Adapting to 4 core operating conditions of public transportation, 4 dedicated scenario branches are constructed. Each branch has 80 LSTM hidden layer nodes (higher than passenger cars, improving the ability to capture features in complex operating conditions) and 40 attention dimensions. The functional positioning of each branch is as follows: Branch 1 (Station-to-Station Stop-and-Rest Temperature Cyclic): Optimizes capacity decay prediction under frequent start-stop conditions; Branch 2 (Midday Fast Charging High Temperature Rest): Focuses on adapting to accelerated aging caused by temperature rise during fast charging; Branch 3 (Nighttime Long-Term Parking): Targets self-discharge and capacity loss prediction under resting conditions; Branch 4 (Intermittent Discharge at Station Entry and Exit): Adapts to the decay mode coupled with electric braking recovery and intermittent discharge. B. Competitive Training: The WTA loss function focuses on the dual indicators of "operating condition adaptability + prediction accuracy". During the midday period (11:00-14:00, peak fast charging), the weight of branch 2 is increased to 1.4; during the morning and evening peak periods (7:00-9:00, 17:00-19:00, dense parking at every station), the weight of branch 1 is increased to 1.3; and during the night period (22:00-5:00 the next day, long-term parking), the weight of branch 3 is increased to 1.2. C. Dynamic window: The sparse entropy threshold is adjusted to 0.35 and 0.75. The window size for the station-stop operation (sparse entropy ≈ 0.3) is set to 12 cycles (to adapt to the daily average operation of 12 loop cycles of buses). The window size for the midday fast charging operation (sparse entropy ≈ 0.8) is set to 4 cycles (the fast charging time is about 1 hour, corresponding to 4 data collection cycles).

[0142] (4) Integrating decision output with remote interaction A. Real-time output: RUL value (e.g., "Remaining life 420 cycles"), 95% confidence interval (e.g., "400-440 cycles"), battery health level (A / B / C / D four levels, associated with bus operation and maintenance standards, C level and below triggers operation and maintenance work order dispatch). B. Targeted warnings: When RUL < 500 cycles, push "Recommend return to the field for capacity calibration" to the remote monitoring platform; when the prediction error of branch 2 is ≥ 5% during midday fast charging, push "Fast charging temperature is too high, it is recommended to adjust the fast charging power to 0.6C"; when the electric braking recovery efficiency decreases by ≥ 3% for 3 consecutive days, push "Electric braking system operating condition test". C. Remote Interaction: Supports remote monitoring platform to call model real-time prediction results, query historical attenuation curves, and analyze operating condition adaptability. Data transmission adopts an encrypted protocol (compliant with public transportation operation data security requirements). A battery health status report is synchronized to the platform every 30 minutes, supporting fleet-level battery management and maintenance scheduling.

[0143] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting the lifespan of a power battery based on mechanism-aware sparse coding and WTA prediction, characterized in that: include: S1: Collect multi-source operational data of the entire life cycle of power batteries for new energy vehicles, and obtain the characteristic data of power batteries after preprocessing the operational data; The preprocessed running data is clustered into typical working conditions using a preset clustering algorithm. S2: Construct a sparse autoencoder consisting of an input layer, an encoder with embedded dynamic sparse coding, a decoder, and an output layer. The input layer matches feature data, the encoder side applies channel sparsity loss, the decoder side applies parameter sparsity loss, and physical consistency loss is introduced at the same time. The output layer outputs a low-dimensional sparse feature sequence. S3: Construct multiple differential prediction branches corresponding to typical working conditions. Each differential prediction branch performs RUL prediction based on the input being a low-dimensional sparse feature sequence. Based on the WTA competitive training mechanism, only the branch parameters with the smallest prediction error for the current sample are updated to obtain multiple differential prediction branches that have been trained. S4: Calculate the feature sparse entropy based on the low-dimensional sparse feature sequence output by the sparse autoencoder and dynamically adjust the sliding window; S5: Extract the low-dimensional sparse feature sequence of the adjusted sliding window length, input it into the trained differential prediction branch, and output the prediction result of RUL prediction value and prediction confidence of each differential prediction branch, and call Kalman filter to correct the prediction result. S6: Using a preset Bayesian fusion rule, the matching degree between typical working conditions and differentiated prediction branches, and the confidence level of the output of differentiated prediction branches, the prediction results of each differentiated prediction branch are weighted and fused to obtain the final RUL prediction value and prediction confidence level. S7: Construct a maintenance decision rule base with multiple conditions linked, and trigger rule matching based on the final RUL prediction value and prediction confidence with the maintenance decision rule base to output the maintenance decision result; S3 includes: S3-1: Construct differentiated prediction branches adapted to typical working conditions. The input of each differentiated prediction branch is the low-dimensional sparse feature sequence output by the sparse autoencoder. , Let be the sliding window length, and d be the feature dimension; each differential prediction branch includes a bidirectional LSTM layer, an attention layer, and a fully connected output layer, expressed as: Bidirectional LSTM layer: in, Let k be the hidden state sequence of the k-th branch. Here are the LSTM network parameters for the k-th branch; Attention layer: in, For attention weight vectors, , These are the trainable parameters for the attention layer. For context vectors, Let be the hidden state at the i-th time step in the k-th branch; Fully connected output layer: in, , These are the trainable parameters for the fully connected layer. The RUL prediction value output by the k-th branch. This corresponds to the prediction confidence level, i.e., the standard deviation estimate. This represents the average predicted value of the k-th branch within the current window; S3-2: Construct the dynamic competitive weights for each differentiated prediction branch based on the basic competitive weights and the operating condition matching weights. The expression is: in, For dynamic competitive weights, Based on the basic competitive weight, Weights are assigned to different operating conditions; This is the adjustment function for the decay phase. This represents the estimated battery health status at time t. This represents the average prediction accuracy of the k-th branch over the most recent N time steps. This is the typical working condition prototype corresponding to the k-th branch. This is a similarity adjustment parameter; S3-3: In each forward propagation, each differentiated prediction branch independently outputs a predicted value. The weighted loss is calculated using the following expression: in, The weighted loss for the k-th branch. This is the actual RUL value; S3-4: Select the differential prediction branch with the minimum loss as the winner, expressed as: in, The winning branch for the current sample; S3-5: Update only the network parameters of the winner branch, while keeping the parameters of the other differential prediction branches unchanged. The expression is: in, For the network parameters of the winning branch at time t, For learning rate, The gradient of the loss function for the winner's branch; S3-6: Introduce a weight adjustment function continuously associated with SOH to adaptively adjust the competitive weights of each differentiated prediction branch at different decay stages. The expression is: in, The decay phase adjustment function for branch k is expressed as: in, The original dynamic competitive weights; The adjusted competition weight for the k-th branch; S3-7: Train the differentiated prediction branch according to the adjusted competition weights to obtain the trained differentiated prediction branch.

2. The method for predicting the lifespan of a power battery based on mechanism-aware sparse coding and WTA prediction according to claim 1, characterized in that: S1 includes: S1-1: Collect multi-source operational data of the entire life cycle of power batteries for new energy vehicles, including core battery parameters, operating environment parameters, vehicle status parameters, and offline testing data; S1-2: The collected multi-source running data are sequentially processed by missing value filling and feature normalization to obtain preprocessed feature data; S1-3: The K-means algorithm is used to cluster the preprocessed feature data into typical working conditions. The clustering objective function is: in, The clustering objective function is... The number of clusters, For the i-th cluster, Let i be the centroid of the i-th cluster. These are the feature values ​​of the preprocessed feature data.

3. The method for predicting the lifespan of a power battery based on mechanism-aware sparse coding and WTA prediction according to claim 2, characterized in that: The typical operating conditions include low temperature high rate, normal temperature cycling, high temperature static placement, intermittent discharge, and long-term storage.

4. The method for predicting the lifespan of a power battery based on mechanism-aware sparse coding and WTA prediction according to claim 3, characterized in that: S2 includes: S2-1: Construct a sparse autoencoder consisting of an input layer, an encoder, a decoder, and an output layer; S2-2: In a sparse autoencoder, the input layer receives and carries the preprocessed feature data; S2-3: In the sparse autoencoder, the encoder embeds a dynamic sparse mask to perform high-dimensional feature compression, sparsification, and core aging feature filtering on the input layer's feature data, outputting a low-dimensional sparse feature vector. Channel sparsity loss is then applied to the output low-dimensional sparse feature vector to suppress redundant channels. The expression is as follows: in, This is a penalty for channel sparsity. For the first The layer encoder outputs a feature vector. This represents the total number of layers in the encoder. It is an L1 norm. For the first The number of neurons in the layer encoder; S2-4: In a sparse autoencoder, the decoder reconstructs the low-dimensional sparse features output by the encoder back to the original dimension, and applies a parametric sparse loss to the decoder weights, pruning redundant parameters. The expression is: in, For parameter sparsity penalty term, For the first Layer decoder weight matrix, This represents the number of decoder layers. It is an L2 norm matrix. For the first The number of input neurons in the layer decoder. For the first The number of output neurons in the layer decoder For the first The connection weights from the i-th input neuron to the j-th output neuron in the layer decoder; S2-5: Introduce physical consistency constraints of capacity and internal resistance into the loss functions of the encoder and decoder, expressed as follows: in, , These are the weighting coefficients. , The input features are estimated values ​​for capacity and internal resistance. , These are the reconstructed capacitance and internal resistance values; S2-6: The reconstruction loss, channel sparsity loss, parameter sparsity loss, and physical consistency loss are fused together, and the expression is: in: For the total loss function, The reconstruction loss is expressed as the reconstruction error. , , The regularization coefficient; S2-7: The output layer directly outputs the low-dimensional sparse feature sequence and reconstruction error reconstructed by the decoder; S2-8: Training and optimizing sparse autoencoders using an improved Adam optimizer.

5. The method for predicting the lifespan of a power battery based on mechanism-aware sparse coding and WTA prediction according to claim 4, characterized in that: S4 includes: S4-1: Obtain the low-dimensional sparse feature sequence output by the sparse autoencoder and extract the low-dimensional feature vector z; S4-2: Calculate the energy proportion of each dimension of the eigenvector z, the expression is: Where i = 1, 2, ..., d; This represents the energy percentage of the i-th feature. This represents the numerical value of the j-th feature. Represents the numerical value of the i-th feature; This represents the energy value of the i-th feature. S4-3: Calculate the sparse entropy, the expression is: in, It is sparse entropy; S4-4: Dynamically adjust the sliding window based on sparse entropy, the expression is: in, This represents the length of the sliding window.

6. The method for predicting the lifespan of a power battery based on mechanism-aware sparse coding and WTA prediction according to claim 5, characterized in that: S5 includes: S5-1: Based on the length of the sliding window Obtain the low-dimensional sparse feature sequence of the sliding window length times and input it into the differential prediction branch respectively. Each differential prediction branch outputs the prediction result including the RUL prediction value and the prediction confidence. S5-2: Define the health state of the power battery as a system state variable. ,in, This represents the estimated battery health status at time t. This represents the estimated remaining lifespan of the power battery at time t. Indicates the SOH decay rate; S5-3: Using the prediction results of each differentiated prediction branch as the observation input, construct a discrete linear state-space model, where the state equation expression is: The expression for the observation equation is: in, Here is the state transition matrix. For the observation matrix, For process noise, Let Q and R be the observation noise, respectively, and let Q and R be the covariance matrices of the process noise and the observation noise, respectively. S5-4: Prediction-update recursive process based on Kalman filtering, utilizing the state estimate from the previous time step. Covariance And the state-space model predicts the prior estimate at the current time step, and the prediction process expression is: in, For state prior estimation, To estimate the covariance a priori, Here is the state transition matrix. This is the transpose of the state transition matrix. Let be the covariance matrix of the process noise; The update process involves using the prediction results of each differentiated prediction branch as the observed values. Calculate Kalman gain Then, the prior estimate is weighted and corrected using Kalman gain to obtain the posterior estimate. The Kalman gain calculation expression is as follows: The posterior estimation expression is: in, For the transpose of the observation matrix, The covariance matrix of the observed noise, For posterior state estimation, To estimate the covariance in the posterior time, It is an identity matrix.

7. The method for predicting the lifespan of a power battery based on mechanism-aware sparse coding and WTA prediction according to claim 6, characterized in that: S6 includes: S6-1: Calculate the original weights of each differentiated prediction branch, and normalize the original weights to obtain the final fused weights. The expression for calculating the original weights is: in, The working condition matching degree of the k-th differentiated prediction branch is the same as the working condition matching weight in S3-2. Let k be the prediction confidence level of the k-th differential prediction branch. The standard deviation of the output corresponding to the differentiated prediction branch; The final fusion weight expression is: in, Total number of branches; S6-2: Calculate the weighted average of the predicted values ​​from each differentiated forecast branch to obtain the final RUL predicted value, expressed as: in, The RUL prediction value for the k-th differential prediction branch. This is the final RUL prediction value; S6-3: The fusion prediction confidence of each differentiated branch is calculated using the weighted variance formula to obtain the final prediction confidence, expressed as follows: in, To determine the final prediction confidence level, The final prediction confidence level for the k-th differentiated prediction branch; S6-4: Based on the final RUL prediction value and final prediction confidence The 95% confidence interval is obtained, expressed as follows: .

Images