Power battery attenuation influence factor quantitative identification method based on ensemble learning

Abstract

The invention relates to the technical field of power battery performance analysis, discloses a power battery attenuation influence factor quantitative identification method based on ensemble learning, and aims to solve the problems of weak data relevance, fuzzy feature importance and poor model adaptability in a traditional method. Comprising the steps of collecting and constructing a multi-source data set; capturing degradation characteristics of different time scales, and screening key influence factors of the whole life cycle in combination with a Top-K space attention mechanism; performing weighted fusion on static attributes and dynamic time sequence characteristics through a gating fusion module to enhance data relevance; a Stacking ensemble learning model is constructed, a decision tree is taken as a base model, linear regression is taken as a meta-model, sliding window preheating training and Bayesian optimization are combined to realize hyper-parameter tuning and weight updating, and model adaptability is improved; and respectively quantifying global and local feature importance through a feature splitting gain sum and an SHAP value, generating an optimization maintenance report containing an interactive thermodynamic diagram and a three-dimensional network diagram, and realizing accurate identification and control of attenuation influence factors.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicles, and specifically to a method for quantitatively identifying factors affecting the degradation of power batteries based on ensemble learning. Background Technology

[0002] With the global energy structure transformation and the rapid development of the new energy vehicle industry, the performance degradation law of power batteries, as the core power source of new energy vehicles, directly determines the vehicle's driving range, safety performance, and service life, and is a key factor affecting the overall vehicle economy and user experience. Currently, the number of new energy vehicles continues to rise, and the scale of retired power batteries is expanding year by year. Accurately identifying the core influencing factors of battery degradation and building reliable degradation prediction models have become core requirements to support battery design optimization, operation and maintenance strategy formulation, and tiered utilization, and are also key technological bottlenecks that the industry urgently needs to overcome.

[0003] However, existing methods for analyzing factors affecting battery degradation still have several technical limitations, making it difficult to meet the high requirements for accuracy, dynamic adaptability, and engineering guidance in practical applications. Specific problems are as follows: First, feature engineering is crude, resulting in insufficient capture and weak correlation of key degradation information. During use, power batteries generate massive amounts of multi-source, heterogeneous spatiotemporal sequence data, which contains crucial information about the degradation of the battery's internal structure. However, existing methods have significant shortcomings in feature construction: on the one hand, they lack multi-scale spatiotemporal feature extraction capabilities, making it difficult to automatically capture key degradation patterns at sub-cycle, single-cycle, and multi-cycle scales from raw voltage and current time-series data; on the other hand, they lack effective feature screening and fusion mechanisms, failing to dynamically identify the core factors most relevant to the current degradation stage from the high-dimensional feature space, and also struggling to adaptively fuse static attributes (such as battery model) with dynamic temporal behavior, leading to weak correlation of model input features and the coexistence of information redundancy and noise interference.

[0004] Secondly, the models lack dynamic adaptability, making it difficult to balance long-term sequence modeling capabilities with computational efficiency. The entire life cycle of a power battery is a typical long-term series problem, with its degradation pattern dynamically evolving with usage time, environmental conditions, and user habits. Existing deep learning-based time series models (such as LSTM and GRU) face a contradiction between computational resource requirements and the linear growth of sequence length when dealing with battery data spanning several years. This forces existing research to reduce temporal resolution or shorten prediction duration, thereby losing long-term dependencies. Furthermore, most existing ensemble learning models are trained statically offline, lacking a training mechanism that can efficiently handle long sequences and perform incremental-full co-optimization throughout the entire life cycle. This results in the model's inability to adapt to the dynamic evolution of degradation patterns in real time, leading to a high risk of long-term prediction failure.

[0005] Third, the accuracy of characterizing and attributing nonlinear coupling relationships is insufficient. Battery degradation is a complex nonlinear physicochemical process influenced by multiple factors, including temperature, charge / discharge rate, cycle count, and SOC range, with significant dynamic interactions between these factors. Traditional time-series models (such as RNNs and LSTMs) are completely incapable of characterizing such complex relationships. Existing improved models based on attention mechanisms, when capturing spatiotemporal dependencies, fail to effectively filter out irrelevant spatiotemporal neighbor information, incorporating irrelevant or weakly correlated features and time-series nodes into the calculation. This introduces noise and dilutes the influence weights of key factors, limiting the model's prediction accuracy. Furthermore, the subsequent quantification of feature importance results are ambiguous, failing to provide clear and accurate guidance for design optimization.

[0006] In summary, current methods for analyzing factors affecting battery degradation have systematic shortcomings in feature engineering, long-sequence modeling capabilities, and the characterization and accurate attribution of complex nonlinear relationships. These shortcomings fail to meet the needs of the new energy vehicle industry for accurate battery performance evaluation, reliable life prediction, and engineering optimization guidance. There is an urgent need to build a new analytical method that can integrate advanced spatiotemporal feature learning, efficient long-sequence modeling, and accurate attribution. Summary of the Invention

[0007] The present invention aims to provide a quantitative identification method for the factors affecting the degradation of power batteries based on ensemble learning, in order to solve the problems of insufficient data correlation mining, fuzzy feature importance, and poor model adaptability in the traditional identification of factors affecting the degradation of power batteries, and to realize the accurate quantitative identification of factors affecting degradation and their engineering application.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for quantitatively identifying factors affecting battery degradation based on ensemble learning, comprising: The data acquisition and processing steps involve accessing basic vehicle information, BMS data, and charging / discharging equipment data to form a multi-source raw dataset. The multi-scale temporal feature construction step involves extracting temporal features at different time scales from the multi-source original dataset to form an initial feature set. The feature dynamic selection and fusion step adopts a feature selection mechanism based on Top-K spatial attention to dynamically select key features from the initial feature set, and adaptively fuses static attribute features and dynamic temporal features through a gated fusion module to generate an optimized high-dimensional feature set. The steps for building and optimizing the ensemble learning model include constructing a Stacking ensemble learning framework with multiple decision trees as the base model and linear regression as the meta-model; using a sliding window preheating training paradigm based on battery lifecycle data, and combining Bayesian optimization algorithm for hyperparameter optimization and dynamic updating of model weights. The feature importance quantification output step quantifies the global feature importance by summing the feature splitting gains, and uses SHAP to calculate the local feature importance, outputting the feature importance ranking; The engineering application output steps generate a design optimization and maintenance strategy report based on the feature importance ranking results, and present key influencing factors through interactive heat maps and three-dimensional feature influence network diagrams to guide the improvement of power battery design and adjustment of usage strategies.

[0009] The principles and advantages of this scheme are as follows: In practical applications, it addresses the deficiencies at the data level by breaking through the limitations of discrete sampling, deeply mining the correlation of time-series data, and enriching the information dimensions through multi-source data fusion, laying a high-quality data foundation for subsequent analysis. Secondly, it achieves accurate mining and deep fusion in the feature engineering stage, abandoning the static feature extraction mode. Combining the degradation characteristics of different stages of the battery's entire life cycle, it automatically extracts features containing key degradation information from the original data through "multi-scale spatiotemporal feature construction," overcoming the limitations of manual feature engineering. Furthermore, it uses the "Top-K attention mechanism" to accurately screen high-dimensional features, effectively filtering out noise and redundancy, and highlighting core influencing factors. Finally, the "gated fusion module" adaptively integrates static attributes and dynamic behaviors, fundamentally enhancing the correlation and representation capabilities of the data, laying a solid foundation for solving the problem of ambiguous feature importance. In terms of modeling, it adopts a Stacking ensemble learning strategy, achieving efficient long-sequence modeling and dynamic adaptation at the model level. It is the first to introduce the "sliding window preheating training paradigm" into the field of battery degradation prediction. This paradigm significantly reduces computational memory overhead by segmenting long sequences into short windows for training, while utilizing a "warm-up" technique to inherit historical states, ensuring the capture of long-term dependencies. This mechanism, combined with Bayesian hyperparameter optimization, enables the model to perform efficient online learning and self-correction throughout the entire battery lifecycle, completely solving the problem of poor dynamic adaptability in traditional static models. This solution forms an end-to-end closed-loop technical system. Starting from multi-source heterogeneous data, through advanced feature engineering, powerful ensemble learning models, and efficient training strategies, it ultimately outputs accurate quantitative attribution results through interpretable algorithms (split gain and SHAP), directly transforming them into executable engineering optimization reports. The entire process achieves a complete closed loop from data to decision-making, promoting the practical application of power battery degradation analysis from theoretical research to engineering practice.

[0010] Preferably, as an improvement, the multi-scale spatiotemporal feature construction step adopts an encoder-decoder network structure to extract degenerate features at sub-cycle, single-cycle, and multi-cycle scales from multi-source original datasets; the network structure includes an encoder composed of convolutional layers and pooling layers, and a decoder composed of deconvolutional layers, which fuse features at different scales through skip connections.

[0011] Technical effects: The encoder-decoder structure enables the systematic extraction of multi-scale degradation features. Convolutional and pooling layers effectively capture key local information, while deconvolutional layers accurately restore feature dimensions. Skip connections break down the transmission barriers between features of different scales, ensuring efficient fusion of shallow detailed features and deep abstract features, solving the limitations of single-scale feature extraction, and providing more comprehensive and accurate feature support for subsequent analysis.

[0012] Preferably, as an improvement, the feature dynamic filtering and fusion step includes: The initial feature set is used to construct a feature network from the various sensors and operating parameters of the battery. A Top-K spatial attention mechanism is employed for the target feature nodes. Calculate its relationship with other nodes in the feature network. Attention score:

[0013] in, Let i be the embedding representation of feature node i. For scoring functions, The total number of features; Sort the results in descending order based on the attention scores, and retain the top-ranked results. The feature nodes are used to form a subset of key features; Selected key dynamic features Static attribute features Input is sent to the gating fusion module and processed by gating weights. Perform adaptive fusion and output the fused feature vector. :

[0014]

[0015] in, It is the Sigmoid activation function. and For learnable parameters, This indicates a splicing operation. This indicates element-wise multiplication.

[0016] Technical effects: The Top-K spatial attention mechanism achieves dynamic screening of key factors by quantifying feature correlation, eliminating redundant features and focusing on core influencing factors, thereby improving feature quality and computational efficiency; the gated fusion module uses adaptive weights to balance the contribution of dynamic temporal features and static attributes, solving the compatibility problem of the fusion of the two types of features, strengthening data correlation, inputting more representative fused features into the model, and improving subsequent prediction accuracy.

[0017] Preferably, as an improvement, the training of the base model includes: initializing with a GBDT tree as the base model, using Bootstrap resampling to allocate a differentiated training sample set for each base model, and simultaneously differentiating input features through feature weighted vectors; training the GBDT base model, where the first tree directly fits the target variable, and each subsequent tree fits the prediction residual of the previous tree, with the fitting model being:

[0018] in, This represents the predicted value of the ensemble model after the m-th iteration. These are the model predictions from the previous iteration. For learning rate, For the currently constructed decision tree model; the optimal split point is determined using the split gain criterion:

[0019] in, denoted as , where is the variance of the node samples, N is the number of samples in the parent node, and left / right are the left and right child nodes after the split.

[0020] Technical effects: The combination of Bootstrap resampling and feature weighted vectors assigns differentiated samples and features to each base model, enhancing the diversity of base models and reducing the homogenization error of ensemble learning; the GBDT tree fitting residual method gradually reduces the model prediction error; the split gain criterion accurately selects the best split point, improving the fitting accuracy of individual decision trees and laying the foundation for the performance of the ensemble model.

[0021] Preferably, as an improvement, the fusion of the meta-model includes: taking the prediction results of the base model as input, learning the optimal integration strategy through the meta-model, and outputting the final predicted value.

[0022] in The final predicted values ​​are integrated for Stacking, where g is a linear regression meta-model. This is the predicted value of the k-th base model.

[0023] Technical effects: The linear regression meta-model integrates the output of multiple base models, avoiding the limitations of a single base model, giving full play to the advantages of each base model, further reducing prediction errors, and improving the stability and accuracy of the final prediction results; it forms an architecture in which the base model fits local patterns and the meta-model integrates global optimization, enhancing the model's ability to capture complex decay relationships.

[0024] Preferably, as an improvement, the hyperparameter optimization using the Bayesian optimization algorithm includes: defining an objective function with the validation set Huber loss as the optimization objective, characterizing the nonlinear mapping relationship between hyperparameters and the objective function through Gaussian process modeling, and guiding the next round of hyperparameter selection through the Expected Improvement acquisition function.

[0025] Technical effects: Using Huber loss as the objective function reduces the interference of outliers on the optimization results and ensures that the hyperparameter optimization direction aligns with actual prediction needs; the combination of Gaussian process modeling and EI acquisition function avoids the blindness of grid / random search, efficiently explores the hyperparameter space, quickly locates the optimal hyperparameter combination, and improves the model optimization efficiency and performance stability.

[0026] Preferably, as an improvement, the dynamic update of the model weights includes adjusting the learning rate based on the battery life cycle stage and performing incremental full-weight collaborative updates. The model for adjusting the learning rate is as follows:

[0027]

[0028]

[0029] in, , This is the upper limit of the learning rate. To standardize the degree of drift, The sensitivity coefficient, This represents the maximum allowable drift threshold during the stable period. This is the sigmoid function.

[0030] Technical effects: Differentiated learning rates are designed for different lifecycle stages (cosine annealing during the stable period, and high sensitivity during the break-in / aging period) to enable the model to quickly respond to changes in decay patterns at each stage and avoid weight oscillations or update lags; the learning rate is dynamically adjusted by standardizing the degree of drift and using the sigmoid function to ensure the rationality and smoothness of the learning rate adjustment and improve the model's adaptability to the dynamic decay process.

[0031] Preferably, as an improvement, the global feature importance is quantified by summing the feature splitting gains, and the model is as follows:

[0032] in, Score the importance of feature j. For the total number of decision trees, Let be the set of nodes that are split using feature j in the T-th tree. The amount by which the loss function is reduced after node s splits.

[0033] Technical benefits: Based on the calculation of the sum of split gains, it directly reflects the contribution of features in all decision tree splitting processes. The quantitative results are intuitive and interpretable, and the key factors driving battery degradation at the overall level are clearly identified. It covers all decision trees and splitting nodes, avoiding the randomness of a single node or tree, and ensuring the comprehensiveness and reliability of global feature importance assessment.

[0034] Preferably, as an improvement, the model for calculating the importance of local features using SHAP is as follows:

[0035] in, Representative feature j pairs of samples The predicted contribution value, where F represents the set of all features, and S is a subset of F that does not contain feature j. To predict the input x by adding feature j to the feature subset S, These are the model predictions using only a subset of features S.

[0036] Technical effects: By comparing feature subsets, the positive and negative contributions of a single feature to the prediction results of a specific sample can be accurately quantified, the dominant cause of a single decay event can be explained, and feature importance analysis at the micro-case level can be achieved; the results can be traced back to specific samples, providing direct evidence for tracing the source of abnormal decay events (such as a sudden drop in capacity), and supplementing the details of the global analysis. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the analysis process for factors affecting the degradation of power batteries according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the model application path in an embodiment of the present invention; Figure 3 This is a schematic diagram showing the ranking of feature importance in an embodiment of the present invention; Figure 4 This is a schematic diagram of the quantitative curve of the charging times characteristic in an embodiment of the present invention. Detailed Implementation

[0038] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 , 2 As shown, a method for quantitatively identifying factors affecting the degradation of power batteries based on ensemble learning includes: data acquisition and processing steps, multi-scale spatiotemporal feature construction steps, feature dynamic screening and fusion steps, ensemble learning model construction and hyperparameter optimization steps, feature importance quantitative output steps, and engineering application output steps.

[0039] The data acquisition and processing steps involve accessing vehicle basic information, BMS data, and charging / discharging equipment data to form a multi-source raw dataset. Specifically, vehicle basic information includes VIN (Vehicle Identification Number, used to uniquely identify a vehicle), vehicle age, rated capacity, and vehicle type; BMS data includes sampling time t and total voltage. Current Individual voltage probe temperature SOC (State of Charge) and total mileage Charging and discharging equipment includes the accumulated energy of the device during charging. Charging rate Charging cutoff voltage BMS required voltage Current demand The first pair of missing values ​​(voltage and current data) in the original data were filled using linear interpolation. For outliers, a method based on... Principle, eliminate those exceeding Data points within a range.

[0040] The multi-scale spatiotemporal feature construction steps, based on an encoder-decoder network structure, extract degradation features at multiple time scales (sub-cycle, single-cycle, multi-cycle) from the original time-series data to construct an initial feature set. The encoder part of this network consists of alternating layers of one-dimensional convolutions and pooling operations, capturing long-period patterns by gradually expanding the receptive field. Its calculation formula is as follows:

[0041] in, For the first l Layered sensory field, The kernel size is [size]. For the first i Layer stride. The decoder consists of deconvolutional layers, responsible for recovering the temporal dimension and reconstructing detailed information.

[0042] In addition, by using skip connections to fuse the low-level local features of the encoder with the high-level abstract features, the multi-scale expressive ability of the model is enhanced. Under this network framework, different degradation stages correspond to different scale feature representations: (1) In the initial break-in stage, the battery performance is not yet stable. At this time, the short-term fluctuations of voltage are captured by the mean and variance of the sliding window with a smaller time window, and the rate of improvement is quantified by calculating the trend slope of the charging and discharging efficiency, so as to reflect the initial adaptability and performance optimization trend of the battery; (2) After entering the stable use stage, the battery performance tends to be stable. The key features turn to the long-term trend slope of voltage and current and the sliding window statistics with a longer time window. These features can effectively reflect the slow change of overall performance and early signs of decay; (3) When the battery enters the aging stage, the performance begins to decay rapidly. At this time, it is necessary to extract the rate of voltage drop and the trend and fluctuation features of internal resistance increase within a shorter time window in order to accurately capture its rapid degradation process. The encoder captures temporal patterns from subcyclic fluctuations to multicyclic trends through layers of convolution and pooling, while the decoder outputs a comprehensive feature representation that integrates multi-scale information through the combined effects of upsampling and skip connections.

[0043] The feature dynamic selection and fusion step treats the initial feature set, composed of the battery sensor data, operating parameters, and multi-scale features constructed above, as a feature network, and applies a Top-K spatial attention mechanism to dynamically select the initial feature set. For each feature node... i Calculate its relationship with other nodes k Attention score:

[0044]

[0045] Before retaining attention score K The features of % constitute a subset of key dynamic features. By fusing static attribute features and dynamic temporal features through a gating fusion module, an optimized high-dimensional feature set is generated. :

[0046]

[0047] in These are static attribute characteristics (such as vehicle age and rated capacity). It is the Sigmoid activation function. This represents element-wise multiplication. This mechanism adaptively assesses the importance of each feature channel based on the battery's real-time state and focuses on the top-K most discriminative features at the current moment. This achieves selective information enhancement and redundancy suppression, resulting in high-quality feature fusion for degradation assessment or lifespan prediction. Specifically, the battery's current lifecycle stage is determined by historical data (such as cycle count, capacity decay rate, and internal resistance change trend), based on the following criteria: Initial break-in stage: cycle count ≤ 100, capacity decay rate < 5%, charge / discharge efficiency fluctuation range > 3%; Stable use stage: cycle count 100-800, capacity decay rate 5%-20%, charge / discharge efficiency stable at 95%-98% (fluctuation range < 1%), internal resistance change rate < 0.5mΩ / month; Aging stage: cycle count > 800, capacity decay rate > 20%, internal resistance change rate > 1mΩ / month, voltage plateau drop > 0.1V (compared to the initial state). Based on the above stage determination, the extraction of specific time-series features is customized for activation or deactivation. For example, during the aging phase, the system automatically enables voltage drop rate calculation based on a 1-minute window (typically 0.02V per hour) to quickly capture signals of accelerated performance degradation; this feature is suppressed during the stable phase due to the gradual voltage changes. This dynamic selection mechanism effectively reduces noise interference and feature redundancy, improving model training efficiency and long-term adaptability.

[0048] This embodiment also includes steps for generating interactive features and selecting features using a Bayesian network. To capture multi-parameter coupling effects, the system generates a series of interactive features (such as the temperature-current joint effect factor, internal resistance-temperature correlation term, and SOC-charging rate product term). To select a key and non-redundant subset from a large number of candidate interactive features, a Bayesian network is used for structure learning and feature optimization, as follows: (1) Network structure learning: A constraint-based PC algorithm is used to initialize an undirected complete graph containing all interaction features and the target variable (capacity decay rate). By performing a conditional independence test (judgment criterion: P(X,Y|Z) = P(X|Z)P(Y|Z)), irrelevant edges are gradually deleted, and the remaining undirected edges are transformed into directed edges according to the causal Markov and Faithfulness conditions, finally forming a Bayesian network structure graph that reflects the causal dependencies between features.

[0049] (2) Feature selection and quantification: Based on the learned network structure, calculate the comprehensive feature importance score of each interaction feature X and target Y:

[0050] in, For conditional mutual information, quantization is performed given other features. Under the conditions and The degree of dependence, Let X be the sum of the strengths of all causal paths from X to Y. These are the weighting coefficients for the causal path strength. Redundancy is also evaluated: if two features... and satisfy ( If a redundancy threshold (ranging from 0.2 to 0.5) is set, then those with lower importance scores are removed. This process ultimately outputs a set of key interaction features with strong explanatory power and low redundancy.

[0051] The steps for building and optimizing the ensemble learning model involve constructing an ensemble (Stacking) learning framework with gradient boosting decision trees as the base model and linear regression as the meta-model.

[0052] Base model training: In this embodiment, Gradient Boosting Decision Tree (GBDT) is the primary model, and residual fitting is used to enhance the model's ability to learn decay patterns. Multiple GBDT trees are initialized as base models, the number of base models is set, and Bootstrap resampling is used to allocate differentiated training sample sets to each base model. Simultaneously, through feature weighted vectors (j is the feature index) Differentiate input features to enhance model diversity and maximize model output diversity:

[0053] Where Diversity represents the model diversity measure (the larger the value, the better the diversity), k is the number of base models, and Dis is the difference measure function (i.e., the Pearson correlation coefficient). Let be the prediction vector of the i-th base model.

[0054] During training, the first tree directly fits the target variable (capacity decay rate), and subsequent trees (m-th trees) fit the prediction residuals of the previous tree.

[0055] in, This represents the predicted value of the ensemble model after the m-th iteration. These are the model predictions from the previous iteration. To control the contribution of each tree as the learning rate, in this embodiment... Set to 0.1, For the currently constructed decision tree model, the optimal split point is determined by minimizing the weighted variance to maximize the split gain value, generated using the split gain criterion.

[0056] in, Let N be the variance of the node samples, N be the number of parent node samples, and left / right be the left and right child nodes after the split. The tree's fitting ability is improved by maximizing the split gain.

[0057] Meta-model fusion: combining the predictions of all base models As input features, the optimal integration weights are learned through a meta-model (linear regression), and the final predicted value is output:

[0058] in The final predicted values ​​are integrated for Stacking, where g is a linear regression meta-model. This is the predicted value of the k-th base model.

[0059] Hyperparameter Bayesian Optimization: This intelligent search method, employing Bayesian optimization and Gaussian process modeling, replaces traditional grid / random search to efficiently locate the optimal hyperparameter combination. Specifically, it defines the Bayesian optimization objective function as follows: using the validation set Huber loss as the optimization objective (balancing outlier sensitivity):

[0060] in, For the Huber loss on the validation set, the loss function is... for:

[0061] in, This is a hyperparameter, initially set to 1.0.

[0062] Gaussian process modeling:

[0063] Where f is the objective function (i.e., the loss on the validation set). It is a mean function. This is the kernel function (Matern 5 / 2 kernel). This involves combining hyperparameters (tree depth, learning rate). A Gaussian process is used to characterize the nonlinear mapping between hyperparameters and the objective function, reducing the number of evaluations of the real model. The Expected Improvement (EI) collection function is used to guide the next round of hyperparameter selection, balancing the exploration of new regions with the utilization of known optimal regions. The expected improvement values ​​for candidate hyperparameters are calculated using the EI function.

[0064] in This is the current optimal verification loss. This represents the expected loss for the new hyperparameter combination.

[0065] Searching in the hyperparameter space makes The largest The loss function is calculated by substituting the data into the real model, and the dataset D is added to retrain the GP model. Simultaneously, overfitting detection and adjustment steps are implemented: overfitting is considered to occur when the validation set loss increases for three consecutive times or when the validation set loss exceeds the training set loss plus a tolerance threshold of 0.02. If overfitting is detected, the learning rate is adjusted by multiplying the current learning rate by a decay factor of 0.8 or reducing the current tree depth by 1 (the total depth should not be less than the set minimum depth of 3).

[0066] The dynamic update of model weights includes adjusting the learning rate based on the battery life cycle stage and performing incremental full-scale collaborative updates. The model for adjusting the learning rate is as follows:

[0067]

[0068]

[0069] in, , This is the upper limit of the learning rate. To standardize the degree of drift, The sensitivity coefficient, This represents the maximum allowable drift threshold during the stable period. This is the sigmoid function.

[0070] The adjustment logic is: set a medium sensitivity level during the break-in period. To avoid overfitting to new noise; smooth adjustment during the stable period using a cosine annealing strategy to prevent weight oscillations; and a high drift sensitivity setting during the aging period. (Triggered by slight drift) .

[0071] Establish an incremental-to-full collaborative update mechanism: Based on an online learning mechanism, implement periodic incremental training to enable the model to dynamically adapt to changes in battery degradation patterns. Set the incremental update period (days) to [value missing]. Full update cycle (month) cumulative drift threshold .

[0072] The incremental update mechanism is as follows: Input the latest k data points (k=100) of the same vehicle model, and use sliding window gradient descent:

[0073] The full update mechanism is as follows: [This refers to the complete historical collection.] With new data Heavy training eliminates incremental error accumulation:

[0074] As battery usage conditions evolve and degradation characteristics gradually become apparent, the model weights are dynamically updated to ensure that the model always possesses good prediction accuracy and generalization ability, and to promptly reflect the latest status of power battery degradation and the changing trends of influencing factors.

[0075] The feature importance quantification step quantifies the contribution of each factor to battery degradation by summing the feature split gains. During GBDT model training, the sum of the gains brought by each feature during decision tree splits is calculated; the larger the gain value, the more significant the impact of that feature on battery degradation. Figure 3 , Figure 4 As shown, based on the feature importance ranking results, interactive heatmaps and three-dimensional feature influence network diagrams can intuitively present key influencing factors, providing clear directions for battery maintenance strategy formulation and battery performance optimization. For example, focusing on high-contribution factors such as temperature and current, and taking corresponding control measures to slow down battery degradation. Specifically: The importance of global features is quantified by summing the feature splitting gains, as shown in the following model:

[0076] in, Score the importance of feature j. For the total number of decision trees, Let be the set of nodes that are split using feature j in the T-th tree. This represents the reduction (gain) in the loss function after node s splits. The output steps are: during GBDT training, calculate the split gain for each feature in each tree. Sum the splitting gains for the same feature across all trees and divide by the total number of trees. Obtain the standardized importance, sort them in descending order of importance, and output the feature ranking (such as temperature, current ratio, pressure difference, etc.).

[0077] We use SHAP (SHapley Additive exPlanations) to calculate the importance of local features and interpret the predictions for individual samples:

[0078] in, Representative feature j pairs of samples The predicted contribution value, where F represents the set of all features, and S is a subset of F that does not contain feature j. To predict the input x by adding feature j to the feature subset S, This is the model prediction using only a subset of features S. The output steps are: for the feature decay event sample x, calculate the predictions for each feature. ;according to Sort the data to identify dominant factors (e.g., the SHAP value of extreme low temperature is significantly higher than the mean in a sudden capacity drop); compare it with global importance to mark local anomalies.

[0079] Engineering application output steps: Generate a design optimization and maintenance strategy report based on the feature importance results, present key influencing factors through interactive heatmaps and 3D feature influence network diagrams, and guide the improvement of power battery design and adjustment of usage strategies.

[0080] The solution has been implemented as follows: (I) Data and Feature Engineering Innovation: Constructing a Feature Processing Mechanism that Combines Dynamics and Causality At the temporal feature processing level, this scheme overcomes the limitations of traditional static feature extraction modes. Through multi-scale spatiotemporal feature construction steps, combined with an encoder-decoder network structure, a dynamic feature selection mechanism based on the entire battery life cycle is established. According to the different degradation characteristics of the three stages of initial break-in, stable use, and aging, the feature extraction strategy is adaptively adjusted: during the break-in period, the focus is on voltage fluctuations and changes in charge and discharge efficiency; during the stable period, the focus is on the long-term trend slope; and during the aging period, the focus is on capturing the voltage decay rate and the increase in internal resistance, so as to achieve accurate matching between features and battery state.

[0081] At the feature selection level, an innovative causal feature selection method based on Bayesian networks is introduced. A causal relationship graph between features is constructed using the PC algorithm, and the importance of features is quantified by combining conditional mutual information and causal path strength. This effectively identifies and eliminates redundant features, uncovering the key interaction factors that truly drive battery degradation, and providing high-quality feature input for the model.

[0082] (II) Innovation in Modeling Methods: Achieving Deep Integration of Ensemble Learning and Bayesian Optimization Employing the Stacking ensemble learning framework, this approach achieves accurate nonlinear modeling through a "multi-base model collaboration + meta-model fusion" architecture. Multiple GBDT trees are used as base models for residual fitting, while linear regression serves as the meta-model for output integration. This approach effectively captures the complex nonlinear relationship between battery degradation and multiple factors, while also mitigating the overfitting risk associated with a single model.

[0083] In terms of hyperparameter optimization, Bayesian optimization is used instead of traditional grid search. A Gaussian process is used to model the complex mapping relationship between hyperparameters and the loss function, and the EI acquisition function is used to intelligently guide hyperparameter selection, accurately identifying the optimal combination of key parameters such as tree depth and learning rate, significantly improving model optimization efficiency and performance stability.

[0084] (III) Model Adaptive Innovation: Establishing a Dynamic Update System Based on Lifecycle Stages A dynamic learning rate adjustment strategy based on battery lifecycle stages was designed. According to the adjustment model: a medium sensitivity setting is used during the break-in period, cosine annealing is used to maintain stability during the stabilization period, and a high sensitivity mechanism is activated during the aging period. Combined with real-time monitoring of standardized drift, the model can quickly respond to changes in degradation patterns at different stages.

[0085] A sliding window preheating training and incremental-full collaborative update mechanism was constructed. Periodic incremental training adapts to short-term data drift, while regular full retraining eliminates error accumulation. This dual-guarantee mechanism effectively solves the failure problem of traditional static models in long-term prediction, ensuring that the model maintains high prediction accuracy throughout the dynamic evolution of battery degradation patterns.

[0086] (iv) Engineering value innovation: forming a multi-dimensional feature importance explanation and application system A multi-dimensional feature importance quantification analysis framework was established, based on SHAP local feature importance calculation and global feature importance quantification methods. It also provides global importance analysis based on split gain and local interpretation capabilities based on SHAP values. This framework can identify key factors driving decay at the macroscopic level and explain the dominant causes of individual decay events at the microscopic level, achieving a comprehensive analysis from macroscopic patterns to microscopic cases.

[0087] Unlike traditional models that only output predictions, this solution provides direct decision support for engineering practice through feature importance ranking and visualization charts. Engineers can optimize battery thermal management design and adjust charging strategies based on high-contribution factors, transforming model analysis results into implementable performance improvement solutions, effectively bridging the gap between modeling analysis and engineering applications.

[0088] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements 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 quantitatively identifying factors affecting the degradation of power batteries based on ensemble learning, characterized in that, include: The data acquisition and processing steps involve accessing basic vehicle information, BMS data, and charging / discharging equipment data to form a multi-source raw dataset. The multi-scale temporal feature construction step involves extracting temporal features at different time scales from the multi-source original dataset to form an initial feature set. The feature dynamic selection and fusion step adopts a feature selection mechanism based on Top-K spatial attention to dynamically select key features from the initial feature set, and adaptively fuses static attribute features and dynamic temporal features through a gated fusion module to generate an optimized high-dimensional feature set. The steps for building and optimizing the ensemble learning model include constructing a Stacking ensemble learning framework with multiple decision trees as the base model and linear regression as the meta-model; using a sliding window preheating training paradigm based on battery lifecycle data, and combining Bayesian optimization algorithm for hyperparameter optimization and dynamic updating of model weights. The feature importance quantification output step quantifies the global feature importance by summing the feature splitting gains, and uses SHAP to calculate the local feature importance, outputting the feature importance ranking; The engineering application output steps generate a design optimization and maintenance strategy report based on the feature importance ranking results, and present key influencing factors through interactive heat maps and three-dimensional feature influence network diagrams to guide the improvement of power battery design and adjustment of usage strategies.

2. The method according to claim 1, characterized in that, The multi-scale spatiotemporal feature construction step adopts an encoder-decoder network structure to extract degradation features at sub-cycle, single-cycle, and multi-cycle scales from multi-source original datasets. The network structure includes an encoder consisting of convolutional and pooling layers, and a decoder consisting of deconvolutional layers, which fuses features at different scales through skip connections.

3. The method according to claim 1, wherein, The feature dynamic filtering and fusion steps include: The various sensors of the battery in the initial feature set are constructed into a feature network with the working condition parameters, a Top-K spatial attention mechanism is adopted, and the attention score of the target feature node with other nodes in the feature network is calculated: wherein, is the embedding representation of feature node i, is the scoring function, is the total number of features; According to the attention scores, arranging in descending order, retaining the top feature nodes, forming a key feature subset; The screened key dynamic features and static attribute features are input into a gated fusion module, and adaptively fused through gated weights to output a fused feature vector : wherein, is a Sigmoid activation function, and are learnable parameters, denotes a concatenation operation, denotes an element-wise multiplication.

4. The method according to claim 1, wherein, The training of the base model includes: initialization using GBDT trees as base models; assigning differentiated training sample sets to each base model using Bootstrap resampling; and differentiating input features through feature weighted vectors. GBDT base model training is then performed, with the first tree directly fitting the target variable, and each subsequent tree fitting the prediction residual of the previous tree. The fitting model is as follows: in, This represents the predicted value of the ensemble model after the m-th iteration. These are the model predictions from the previous iteration. For learning rate, For the currently constructed decision tree model; the optimal split point is determined using the split gain criterion: in, denoted as , where is the variance of the node samples, N is the number of samples in the parent node, and left / right are the left and right child nodes after the split.

5. The method according to claim 1, wherein, The fusion of the meta-model includes: taking the prediction results of the base model as input, learning the optimal integration strategy through the meta-model, and outputting the final prediction value. wherein is the Stacking ensemble final prediction, g is the linear regression meta-model, is the prediction of the k-th base model.

6. The method according to claim 1, wherein, The hyperparameter optimization using the Bayesian optimization algorithm includes: defining an objective function with the validation set Huber loss as the optimization objective; characterizing the nonlinear mapping relationship between hyperparameters and the objective function through Gaussian process modeling; and guiding the next round of hyperparameter selection through the Expected Improvement acquisition function.

7. The method according to claim 1, wherein, The dynamic update of model weights includes adjusting the learning rate based on the battery life cycle stage and performing incremental full-scale collaborative updates. The model for adjusting the learning rate is as follows: wherein, , is an upper limit for the learning rate, is a degree of standardization drift, is a sensitivity coefficient, is a maximum allowed drift threshold for the stabilization period, is a sigmoid function.

8. The method according to claim 1, wherein, The model for quantifying global feature importance by summing the feature splitting gains is as follows: wherein, is the importance score of feature j, is the total number of decision trees, is the set of nodes in the Tth tree that use feature j to split, is the reduction in the loss function after splitting node s.

9. The method according to claim 1, wherein, The model for calculating the importance of local features using SHAP is as follows: in, Representative feature j pairs of samples The predicted contribution value, where F represents the set of all features, and S is a subset of F that does not contain feature j. To predict the input x by adding feature j to the feature subset S, These are the model predictions using only a subset of features S.

Images