Battery consistency-oriented multi-algorithm collaborative abnormal battery cell positioning method

Abstract

The invention relates to the technical field of battery health management, in particular to a multi-algorithm collaborative abnormal battery cell positioning method for battery consistency. The method comprises the following steps: synchronously acquiring multi-modal data of a battery cell through multiple channels, introducing a short-time perturbation enhancement and sliding window processing technology, carrying out time alignment, filtering, denoising and standardized integration on the data, and constructing a high-sensitivity characteristic matrix; the matrix is analyzed in parallel through multiple algorithms, a multi-dimensional evidence signature of each battery cell is generated, and a holographic evidence matrix is constructed in combination with algorithm uncertainty and cross-algorithm consistency; the weights of the algorithms are dynamically distributed through game type weight scheduling, a consensus graph between the battery cells is constructed, and abnormal clusters are recognized in combination with node centrality analysis and edge weight clustering; and performing hierarchical grading and dynamic risk scoring on the abnormal battery cells based on collaborative reasoning and global influence degree analysis, and generating a priority processing queue. According to the invention, closed-loop optimization and adaptive scheduling are realized, and the consistency and reliability of the battery system are improved.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology, specifically to a multi-algorithm collaborative abnormal cell localization method for battery consistency. Background Technology

[0002] With the rapid development of new energy technologies and the widespread application of battery systems in electric vehicles, energy storage power stations and other fields, the consistency management of battery cells in battery packs has become a key factor affecting system safety and performance. In actual operation, due to differences in manufacturing processes, usage environment and aging degree, battery cells may have inconsistent parameters such as voltage, temperature and internal resistance, which may lead to local abnormalities or even serious failures such as thermal runaway.

[0003] Chinese invention patent application CN116990709A discloses a method for judging the consistency of energy storage batteries. This method uses battery charging and discharging characteristic data, performs dimensionality reduction on the characteristic data, and then uses statistical methods such as unsupervised clustering, anomaly detection, and distance + box plot to detect the data from multiple perspectives. By detecting the data from different angles, the calculation results are more comprehensive and accurate, and the urgency of the problems is ranked by an optimizable coefficient.

[0004] Existing detection methods mostly focus on single parameter or static threshold judgment, making it difficult to fully capture multimodal anomaly features and their correlations. Especially under complex operating conditions, early and weak anomaly signals are easily masked or ignored by noise. Therefore, it is necessary to develop a high-precision positioning method that can integrate multi-source information and realize dynamic collaborative analysis to improve the sensitivity, reliability and support capability of anomaly detection for battery consistency management. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a multi-algorithm collaborative abnormal cell localization method for battery consistency.

[0006] The technical solution of this invention: a multi-algorithm collaborative abnormal cell localization method for battery consistency, comprising the following specific implementation steps: S1. By synchronously acquiring the voltage, temperature, terminal current and internal resistance data of the battery cell through multiple channels, and combining short-time perturbation enhancement and sliding window processing technology, the multi-modal data is time-aligned, filtered and denoised and standardized to construct the input feature matrix. S2. Construct an algorithm pool, analyze the input feature matrix in parallel using multiple algorithms, generate multi-dimensional evidence signatures for each battery cell, and construct a holographic evidence matrix by combining the uncertainty of the algorithm output and cross-algorithm consistency information. S3. Based on the holographic evidence matrix, the contribution weights of each algorithm are dynamically allocated through game-like weight scheduling to construct a weighted consensus graph among battery cells. An abnormal prompt list is generated by combining node centrality analysis and edge weight clustering. S4. Based on the weighted consensus graph and the list of abnormal prompts, the abnormal relationship between cells is quantitatively analyzed through local and global collaborative reasoning. Abnormal clusters are identified and cells within the cluster are classified and graded to generate dynamic risk scores. S5. Based on the hierarchical classification and dynamic risk scoring of the abnormal cluster, calculate the comprehensive priority of each abnormal node, form a priority processing queue, and combine collaborative optimization processing strategies and closed-loop effect feedback to complete the priority processing and optimized scheduling of multi-node abnormalities.

[0007] Preferably, constructing the input feature matrix specifically includes: The terminal voltage, individual cell temperature, and terminal current of all cells in the battery pack are collected synchronously. A perturbation current is introduced to measure the internal resistance, and the instantaneous internal resistance estimate of each cell is calculated based on the internal resistance estimation formula; The acquired multi-channel signals are time-aligned and denoised using sliding window filtering to obtain the filtered voltage signal. By applying perturbation current or voltage pulses and extending the sampling window, the short-time energy and enhanced internal resistance response of each cell are calculated to amplify minute anomaly characteristics. The original and enhanced signals are standardized and integrated with voltage, temperature, internal resistance, short-time energy and perturbation response information to form a unified multimodal feature matrix.

[0008] Preferably, constructing a holographic evidence matrix specifically includes: Construct an algorithm pool encompassing time-domain residual detectors, frequency-domain harmonic structure detectors, local coupling synchronization detectors, and lightweight embedded nonlinear feature models; Each algorithm independently processes the signal of each battery cell, calculates anomaly indicators, and outputs the corresponding uncertainty estimate. Vectorize and integrate all algorithm outputs for each battery cell to generate a multidimensional evidence signature containing anomaly index vectors and uncertainty vectors. The similarity of the multi-algorithm outputs for each battery cell is calculated to quantify the degree of consensus among different algorithms on the same anomaly, and to form a consistency matrix among algorithms. All cell evidence signatures and consistency information are integrated into a final holographic evidence matrix.

[0009] Preferably, the dynamic allocation of contribution weights for each algorithm through game-theoretic weight scheduling specifically includes: Initialize the weights of each algorithm for the multi-algorithm evidence vector of each battery cell; A local game model is constructed, in which each algorithm is regarded as a player in the game, and its payoff function integrates the algorithm's anomaly index, uncertainty, and the mean of consistency with other algorithms. The weight of each algorithm for each battery cell is dynamically adjusted through an iterative optimization method, so that the weight is tilted towards the algorithm that is more reliable and consistent in judging anomalies, until the weight converges or the maximum number of iterations is reached.

[0010] Preferably, constructing a weighted consensus graph among battery cells specifically includes: The weighted similarity of abnormal features among battery cells is calculated using the optimized algorithm weights, and used as the edge weights of the consensus graph. Calculate the overall anomaly intensity of each battery cell as the node self-loop weight of the consensus graph; Based on the node self-loop weights and edge weights of all battery cells, a complete weighted consensus graph is constructed.

[0011] Preferably, the quantitative analysis of abnormal relationships between battery cells through local and global collaborative reasoning specifically includes: For each cell node, the local anomaly impact degree is calculated. This impact degree integrates the node's own weighted anomaly index and the collaborative anomaly impact of its neighboring nodes. Based on the local anomaly impact, the global anomaly impact is obtained by simulating the propagation path of the anomaly between cells through global cumulative calculation.

[0012] Preferably, identifying abnormal clusters and classifying the cells within the cluster into different levels specifically includes: Node pairs with edge weights exceeding a preset threshold are selected from the consensus graph to form a set of highly correlated battery cell pairs; Weighted clustering of highly correlated cell pairs identifies anomalous clusters with significant cooperative relationships; Based on the global anomaly impact of each cell node, the cells within the cluster are classified into Level 1, Level 2, and Level 3 anomalies using a preset global impact threshold.

[0013] Preferably, generating a dynamic risk score specifically includes: Within each anomalous cluster, core nodes and boundary nodes are identified based on the global anomalous impact and node degree of each node. Calculate the internal density metric for each anomalous cluster, which is based on the sum of the weights of all edges within the cluster; By combining the average global impact of core nodes and the cluster density index, a comprehensive risk score is calculated for each abnormal cluster.

[0014] Preferably, a priority processing queue is formed, and a collaborative optimization processing strategy and closed-loop effect feedback are combined, specifically including: The overall priority of each anomalous node is calculated by considering its global anomaly impact, the risk score of its respective anomalous cluster, and its hierarchical level. The abnormal nodes are sorted according to their overall priority, and a dynamic maintenance queue is generated. For nodes in the queue, a collaborative optimization processing strategy is generated by combining the collaborative influence of their neighboring nodes and maintenance resource constraints. After each maintenance operation, the global impact of the node and the cluster risk are updated, and the maintenance queue is iteratively adjusted. The effectiveness of each collaborative process is quantitatively evaluated, and the evaluation results are fed back into risk scoring and priority calculation to achieve closed-loop optimization.

[0015] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a multi-algorithm collaborative anomaly cell localization method for battery consistency. This method significantly improves the detection sensitivity of early, weak anomaly features by fusing multimodal data and parallel analysis of multiple algorithms, and introducing short-term perturbation enhancement technology. This achieves a shift from "passive fault alarm" to "active anomaly early warning," laying a solid foundation for accurate fault prediction and health management of battery systems. Furthermore, it employs a game-theoretic weighted scheduling mechanism to dynamically integrate evidence from multiple algorithms and construct a consensus graph among cells. This not only effectively overcomes the limitations of single detection algorithms, such as susceptibility to operating condition interference and biased judgment, but also achieves a leap from isolated point anomaly detection to collaborative cluster anomaly identification. It can accurately locate the core nodes of anomalies and their impact range; and perform hierarchical classification and dynamic risk quantification of anomaly clusters, thereby providing unprecedented structured and visualized basis for maintenance decisions, greatly enhancing the analytical depth and decision reliability of the fault prediction and health management system; and by establishing a priority processing queue that integrates global impact, cluster risk and maintenance resources, and introducing a closed-loop effect feedback mechanism, it realizes adaptive optimization of maintenance strategies and dynamic configuration of scheduling resources; this forms a continuously improving intelligent operation and maintenance closed loop, ensuring that the battery system consistency management strategy is always in the optimal state, significantly improving the overall reliability, safety and service life of the system. Attached Figure Description

[0016] Figure 1 This is a flowchart of a multi-algorithm collaborative abnormal cell localization method for battery consistency proposed in this invention. Detailed Implementation

[0017] Example 1, as Figure 1 As shown, the present invention proposes a multi-algorithm collaborative abnormal cell localization method for battery consistency, which includes the following specific implementation steps: S1. By synchronously acquiring cell voltage, temperature, terminal current, internal resistance, and transient waveforms through multiple channels, and combining short-time perturbation enhancement and sliding window processing, the multi-modal data is time-aligned, filtered for noise reduction, and standardized and integrated to form a highly comparable and highly sensitive input feature matrix. The specific implementation process is as follows: S11. Simultaneously acquire multi-channel data such as battery pack cell terminal voltage, individual cell temperature, terminal current, and internal resistance, and introduce micro-perturbation current to measure internal resistance to ensure consistent observation time and high data comparability for different cells, providing a reliable source of original signals for anomaly-sensitive characteristics. Specifically: Define the voltage matrix acquisition formula: ; Define the temperature matrix acquisition formula: ; Define the internal resistance estimation formula (based on the perturbation injection method): ; Where V(t) represents the column vector of cell voltages at time t, with dimension N×1, representing the set of all cell voltages in the battery pack; V i (t) represents the terminal voltage of the i-th cell at time t, derived from cell terminal measurements and obtained through high-speed ADC sampling; N represents the total number of cells in the battery pack; T(t) represents the column vector of cell temperatures at time t, with dimension N×1; T i (t) represents the surface or internal temperature of the i-th cell, derived from a temperature sensor; R i (t) represents the estimated instantaneous internal resistance of the i-th cell; This indicates the magnitude of voltage change after applying a short-time perturbation current / voltage. This indicates the magnitude of the corresponding current disturbance; This represents the matrix transpose operation; S12. Time alignment, interpolation, and sliding window filtering are performed on the acquired multi-channel signals for noise reduction. Short-time window sliding technology is introduced to enhance transient anomaly characteristics while reducing high-frequency noise interference. Specifically: The filtering and denoising process is as follows: ; A linear interpolation method is used for time alignment to obtain the aligned voltage signal. ; in, This represents the filtered voltage signal, which retains transient anomalies while suppressing noise. The weight coefficients of the sliding window satisfy the following conditions: M represents the half-width of the sliding window, which determines the smoothness of the filtering. S13. Apply a perturbation current or voltage pulse under safety constraints and extend the sampling window. Combine short-time energy and instantaneous slope to extract enhanced signals, amplify minute abnormal features, and make signals of potential manufacturing or contact defects easier to detect. Specifically: Based on the aligned data, a perturbation current or voltage pulse is applied, while the temperature and voltage sampling windows are extended, amplifying the response of minute defects. The short-time energy can be expressed as: ; Meanwhile, the enhanced internal resistance response is: ; Among them, E i (t) represents the short-time energy of the i-th cell, used to enhance transient anomaly signals; This represents the filtered voltage sample; L represents the short-time window length. This indicates the change in internal resistance after the perturbation is enhanced; This represents the baseline internal resistance without any perturbation. S14. Standardize the original and enhanced signals and construct a multimodal feature matrix, integrating multidimensional information such as voltage, temperature, internal resistance, short-time energy, and perturbation response to form a unified holographic feature matrix that can be directly input into multiple algorithm pools, facilitating collaborative analysis and anomaly localization. Specifically: The multi-channel, multi-feature signals obtained in steps S11–S13 are standardized and integrated into a unified feature matrix: ; ; in, Represents the standardized eigenvalues; Characteristic values, such as voltage, temperature, internal resistance, and short-time energy; and Let X(t) represent the mean and standard deviation of the f-th feature of the i-th cell, respectively, which are estimated from historical data; X(t) represents the integrated multimodal feature matrix; and F represents the feature dimension.

[0018] S2. By analyzing the enhanced multimodal feature matrix output from step S1 using multiple algorithms in parallel, a multidimensional evidence signature is generated for each battery cell. Combining the uncertainty and cross-algorithm consistency information output by the algorithms, a holographic evidence matrix is ​​constructed to achieve sensitive capture and interpretable analysis of minute anomalous signals. The specific implementation process is as follows: S21. Based on the feature matrix output in step S1, construct a pool of complementary algorithms covering time-domain residuals, frequency-domain harmonics, local coupling, and nonlinear embedding. Process each cell signal in parallel to generate multi-angle anomaly index vectors, providing a preliminary, multi-dimensional foundation of anomaly features for subsequent evidence unification and consistency analysis. Specifically: Using the output X(t) of step S1 as input, a multi-algorithm pool is constructed to cover detection algorithms for different types of abnormal signals. The algorithm pool includes, but is not limited to: Short-time residual / statistical detector: Sensitive to transient anomalies or sudden drifts, it calculates voltage, internal resistance, or temperature deviations based on sliding window residual analysis. ; ; in, This represents the uncertainty of the abnormal index of the i-th cell under short-time residual detection, indicating that a large residual fluctuation indicates low reliability. represents the voltage signal of the i-th cell after short-time filtering or enhancement; L represents the sliding window length, which defines the time range for residual calculation and is used for local statistical analysis. It is set according to the signal sampling rate and battery response characteristics. This represents the average voltage within the window, and the short-time local average value, used for comparing transient deviations. This represents variance calculation, measuring the magnitude of residual fluctuation within a window; This represents the transient anomaly index of the i-th cell calculated by the short-time residual / statistical detector; Spectrum / Harmonic Structure Detector: Detects potential oscillations or periodic anomalies by analyzing the spectral changes of voltage and current using local Fast Fourier Transform (FFT). ; ; in, This represents the uncertainty of the i-th cell under spectrum detection, reflecting the prominence of abnormal spectrum characteristics; The power spectral density of the i-th cell at frequency f is derived from the local Fourier transform or wavelet transform, reflecting the energy distribution of the signal in the frequency domain; f represents the frequency index, covering the frequency range of the target voltage or current signal. It represents the maximum peak value in the spectrum, characterizing the most significant frequency domain feature; This represents the total energy of the spectrum and is used for normalization to ensure that the uncertainty value is in the range of 0 to 1. Represents the Fourier transform operator; Local Coupling / Synchronization Detector: Analyzes short-term cosynchronicity between battery cells to detect local inconsistencies caused by abnormal cells. ; ; in, This represents the uncertainty of the i-th cell under synchronization detection, and characterizes the magnitude of the fluctuation in synchronization consistency with other cells. , These represent the voltage signals of the i-th and j-th cells after short-time enhancement processing, respectively. The correlation coefficient between the short-time signals of the i-th and j-th cells is represented, reflecting the strength of synchronization. The set represents the calculation of the correlation coefficient with all cells except itself; The standard deviation is used to measure the fluctuation in synchronization with other battery cells. Large fluctuations indicate low reliability in judging the abnormality of the battery cell. This represents the local synchronization index of the i-th cell; Lightweight embedding / nonlinear feature models: These models extract long-term nonlinear drift patterns based on low-dimensional embeddings or nonlinear autoencoders, making them suitable for capturing cumulative anomalies. ; ; in, This represents the uncertainty of the i-th cell under the nonlinear characteristic model, characterizing the proportion of the model prediction error relative to the fluctuation of the original signal; This represents the normal signal predicted by the nonlinear model. An anomaly index representing a nonlinear embedded or autoencoder model, characterizing prediction error; Therefore: Each algorithm independently calculates cell anomaly indicators. This forms preliminary multi-angle anomaly characteristics; S22. Vectorize and integrate the outputs of each algorithm, while introducing uncertainty estimates for each algorithm to form a unified multi-algorithm evidence signature. This takes into account both the source information and reliability of the algorithms, providing highly comparable input for subsequent cross-algorithm consistency analysis and weighted processing, and enabling interpretable quantification of anomalous signals. To unify the outputs of various algorithms, a multi-dimensional evidence signature (Sign) is generated for each battery cell. i The structure includes: ; ; Among them, Sign i This represents the multi-algorithm evidence signature for the i-th battery cell, containing all algorithm outputs and uncertainty information; U represents the abnormal indicator vector output by algorithm a for the i-th cell; i This represents the uncertainty vector output by each algorithm; A represents the total number of algorithms in the algorithm pool. S23. Calculate the similarity or consistency of the multi-algorithm outputs for each battery cell, quantify the consensus level of different algorithms on the same anomaly, and mark high-confidence or uncertain signals through a consistency matrix to provide a reliable basis for subsequent weight scheduling, thereby enhancing the robustness and interpretability of multi-algorithm collaborative judgment. Based on evidence signatures, calculate the consistency matrix of each battery cell across different algorithms: ; Based on this, the consistency matrix C is obtained. i The consensus of different algorithms on the same cell anomaly is quantified. Low consistency indicates uncertainty or possible interference from local anomaly signals. in, This represents the consistency index between the outputs of algorithm a and algorithm b for the i-th battery cell; C represents a function for calculating vector similarity. i Let A represent the algorithm consistency matrix for the i-th cell, with dimension A×A; S24. Integrate the evidence signatures and consistency information of all battery cells into a holographic evidence matrix, uniformly representing multi-algorithm features, uncertainties, and consistency, providing high-information input for the collaborative reasoning and anomaly localization module, namely: Integrate the evidence signatures and consistency information of all battery cells into a final matrix SE: ; ; Here, SE represents the multi-algorithm evidence signature matrix of the entire battery pack, which includes the algorithm output, confidence level, and inconsistency information of each cell.

[0019] S3. Based on the multi-algorithm evidence matrix in step S2, the contributions of each algorithm are dynamically allocated through game-theoretic weight scheduling to construct a consensus graph among battery cells. Anomaly alerts are generated by combining node centrality and edge weight clustering to complete multi-algorithm collaboration and quantification of cross-cell anomaly relationships. The specific implementation process is as follows: S31. Initialize the weights of the multi-algorithm evidence vector for each cell and construct a local game model. Each algorithm aims to optimize the payoff function. The contribution of anomaly indicators, uncertainty, and cross-algorithm consistency is comprehensively quantified to provide basic conditions and evaluation criteria for iterative optimization. Specifically: Evidence vector for each cell i Initialize the weights of each algorithm: The weight represents the initial contribution value of the algorithm in anomaly detection, and uniform initialization ensures fair participation of the algorithm; Constructing a local game model: Each algorithm is considered a player in the game, and its strategy is to choose weights. To maximize its benefits The payoff function is: ; ; in, This represents the game payoff function of algorithm a under cell i, which comprehensively reflects the degree of contribution, confidence level, and consistency of the algorithm. The dynamic weight of algorithm a under cell i represents the relative importance of the algorithm in the anomaly detection of cell i. This represents the abnormal index output by algorithm a for cell i; This represents the average consistency of algorithm a relative to other algorithms on cell i; S32. The algorithm weights are dynamically adjusted through iterative optimization, with weights tilted towards algorithms that are more reliable in anomaly detection, while also taking into account cross-algorithm consistency, to achieve adaptive weight allocation, i.e.: Weights are optimized using a discrete game iterative optimization algorithm. ; in, This represents the current weight of algorithm a for cell i (the value taken in the t-th iteration); This represents the learning rate, which controls the magnitude of each weight update based on the derivative of the return. This represents the derivative of the payoff function with respect to the weights; This represents the weight of cell i after algorithm a iterates over it. Iteration termination condition: Weight convergence: (Set convergence threshold); reaching the maximum number of iterations T max ; S33. A weighted consensus graph is constructed using optimized algorithm weights and the similarity of anomaly features among battery cells. Nodes represent battery cells, edge weights represent the degree of anomaly consensus, and a self-loop weighting algorithm is used to weight the anomaly strength. This achieves unified quantification of the relationship between multiple algorithm outputs and collaborative anomalies among battery cells, providing a structured foundation for anomaly cluster detection. Specifically: The optimized weights are applied to the similarity of abnormal features among battery cells to construct a weighted graph G: ; Node self-loop weights: ; in, The consensus edge weight between cell i and j represents the weighted similarity of different cells in each algorithm's abnormal mode, and is used to identify collaborative abnormal relationships. G represents the self-loop weight of a node, which characterizes the overall anomaly intensity of the cell itself; G represents the consensus graph matrix, a weighted graph structure composed of all nodes and edge weights, which describes the abnormal collaborative relationship structure between cells in the entire battery system; S34. Enhance the consensus graph with features, including node centrality analysis and edge weight clustering, quantify the importance of cell anomalies and local anomaly clusters, and generate an anomaly alert list that can be directly used for collaborative localization, thus achieving a closed loop of data-weights-graph-alerts. Specifically: Feature extraction and enhancement of the map G: Node centrality analysis: Indicators such as Degree, Eigenvector, and PageRank are used to measure the importance of cell anomalies; Edge-weighted clustering analysis: Identifying local anomalous clusters or potential chain anomalies based on weights and similarities; Based on this, a list of error messages is generated: ; in, This represents the abnormality alert value for cell i. The final abnormality alert value is generated by comprehensively considering the abnormality intensity of cell i itself and its correlation with the abnormalities of other cells. This function represents an anomaly alert function that maps outliers and their correlation to risk levels in the range of 0-1. It should be noted that the exception message function It is a comprehensive function used to quantify and output the degree of abnormal risk of battery cells. Its core function is to normalize and fuse the node self-loop weights obtained by multi-algorithm collaborative identification with the abnormal correlation of other battery cells.

[0020] S4. Based on the game-theoretic optimization consensus graph and anomaly alert list generated in step S3, the abnormal relationships between battery cells are quantitatively analyzed through local and global collaborative reasoning. Anomaly clusters and chain anomalies are identified, hierarchical anomaly levels are formed, and dynamic risk scores and anomaly alerts are generated to achieve closed-loop decision-making for multi-dimensional collaborative anomaly localization. The specific implementation process is as follows: S41. Calculate the local anomaly impact degree for each cell node, combining the node's own weighted anomaly index with neighborhood collaborative information. The local anomaly risk is reflected through neighbor-weighted accumulation, providing a quantitative basis for global collaborative reasoning. Simultaneously, the neighborhood size and weight coefficients can be adjusted to adapt to different module structures. Specifically: For each cell node i, calculate the local anomaly impact degree. The combined effects of node anomalies and neighborhood collaboration: ; in, N(i) represents the local anomaly impact degree of node i; N(i) represents the set of neighbors of node i, which is used to accumulate the cooperative anomaly impact of neighboring cells on node i. This represents the local neighbor contribution weight coefficient, which controls the proportion of contribution of neighbor anomalies to the local impact of a node; This represents the local influence function of neighbor node j in the previous iteration. It smooths out the contribution of neighbor anomalies and avoids excessive interference from single anomaly fluctuations on the local influence. The previous local influence can be directly taken. S42. Based on the local impact, the propagation path of anomalies among battery cells is simulated through global cumulative calculation. Considering the neighbor normalized weighting and the global impact coefficient, chain-like or cross-regional collaborative anomaly identification is achieved, providing a global quantitative indicator for anomaly cluster analysis. Specifically: Based on the local impact level, global collaborative reasoning is performed to simulate the possible diffusion path of anomalies between battery cells: ; in, This represents the global anomaly impact of node i; This represents the global impact coefficient, which adjusts the contribution of cumulative anomalies from neighbors to the global impact. S43. Based on global influence and graph topology, weighted clustering and community detection are performed to form anomaly clusters. The cells within the clusters are then classified into Level 1, Level 2, and Level 3 anomalies according to the local and global influence of each node, achieving a structured hierarchical structure for both single-point and collaborative anomalies. Specifically: Select node pairs (i,j) with significant cooperative relationships from the consensus graph G: ; in, This represents the edge weight threshold, used to filter edges with strong collaborative relationships between nodes in the consensus graph; This represents the set of valid edges after initial screening, i.e., highly correlated cell pairs; Weighted clustering is performed on each subgraph to identify anomalous clusters C. m : ; in, This means finding a subset C from the set of candidate nodes. m This maximizes the weighted coordination anomaly intensity among nodes within the cluster. For each anomalous cluster C m Internal cell nodes according to global impact Layering: ; Among them, Th1 and Th2 represent the global impact thresholds, which are used to classify the anomaly level and are determined by combining historical anomaly data statistics, battery safety strategies or experimental experience. This indicates the anomaly level (Level 1 / Level 2 / Level 3) of node i, used to distinguish between core nodes and auxiliary nodes within the cluster, providing a basis for subsequent collaborative processing and risk prioritization; S44. Calculate a comprehensive risk score for the abnormal cluster, and generate a dynamic abnormality alert list by combining the global impact and the cluster's internal density. Mark the cluster's central cell, potential chain-like abnormal nodes, and their priority levels to achieve a structured and operable anomaly location and risk management closed loop. Specifically: For each anomalous cluster C m Internal nodes, based on global impact Node degree (number of connections) is used to identify core nodes and boundary nodes: ; ; in, Indicates cluster C m The core node set within the cell, which is usually the cell with the most severe anomaly and the greatest collaborative impact; Indicates cluster C m The set of non-core nodes, i.e., boundary nodes; deg(i) represents the connectivity of node i in the consensus graph G, i.e., the number of its neighbors; T c D represents the global influence threshold of the core node; c The threshold represents the node degree, used to filter highly collaborative nodes within the cluster. It is set based on the average number of connections in the cluster or experience. The \ symbol represents the set difference operation, which removes all elements from the right set from the left set. To quantify the degree of abnormal coordination among nodes within the cluster, a cluster density index is defined: ; in, Indicates cluster C m Density index of inter-node cooperation strength; |C m | represents the m-th anomalous cluster C m The number of nodes in the cluster, i.e., the total number of battery cells contained in the cluster; A cluster risk score R is generated by combining the global impact of core nodes and cluster density. m : ; Among them, R m Indicates abnormal cluster C m Comprehensive risk score; This represents the weighting coefficient between node influence and cluster density in the risk score, with a value range of [0,1]. This indicates the number of core nodes in the cluster, i.e., the size of the core node set; Based on risk score R m Sort all abnormal clusters and generate an error message list: Includes cluster number, core node list, boundary node list, risk score, and priority level; By combining the dynamic update list of maintenance resources, real-time monitoring and scheduling can be achieved; Anomaly alerts can be directly entered into the maintenance system or upper-level scheduling policies for alarms, priority troubleshooting, load balancing, or preventative maintenance.

[0021] S5. Based on the abnormal cluster and node hierarchical information generated in step S4, the priority handling and optimized scheduling of multi-node anomalies are completed through comprehensive priority sorting, collaborative optimization strategy generation, dynamic queue maintenance, and closed-loop effect feedback. The specific implementation process is as follows: S51. Based on the overall global impact of the node, the risk score of its associated abnormal cluster, and its hierarchical level, calculate the overall priority of each abnormal node to form a multi-dimensional ranking result, namely: Based on risk score R m And node hierarchical levels, calculate the comprehensive priority P of each abnormal node i. i : ; in, The risk score of the abnormal cluster C(i) to which node i belongs is calculated by step S44 using the influence of the core node and the strength of cluster collaboration. The node hierarchical level indicators (numerical), such as Level 1 anomaly = 3, Level 2 = 2, Level 3 = 1; , and Represents the weighting coefficients, corresponding to the contributions of the node's global influence, cluster risk score, and hierarchical level to the overall priority; P i This represents the overall priority value of the node, used for sorting and generating dynamically maintained queues; S52. Based on the priority ranking results, combined with the collaborative influence relationship between nodes and maintenance resource constraints, a collaborative optimization processing strategy is generated. By minimizing the combination function of anomaly impact and resource cost, multi-node collaborative processing is achieved, prioritizing the reduction of the impact of core nodes while also considering the processing effect of neighboring anomaly nodes, thus realizing the overall optimal anomaly management scheme, namely: For the sorted abnormal nodes, generate a collaborative optimization processing strategy AS. i Simultaneously, abnormal correlations between nodes and maintenance resource constraints are considered: ; Among them, A i This represents the set of neighboring nodes that have a cooperative anomaly with node i; This represents the expected decrease in the global influence of neighbor node j after processing; This represents the amount of resources required to maintain or adjust node j. and The weighting coefficients represent the optimization objective function, adjusting the priority between reducing the impact of anomalies and resource costs; AS i This represents the collaborative optimization strategy for node i, including the processing order, method, and neighboring nodes involved. S53. Generate a dynamically maintained queue with the sorted nodes and collaborative processing strategies. The queue is processed according to priority. After each processing, the node impact and cluster risk are updated, and the queue is iteratively adjusted based on the latest state to achieve real-time dynamic scheduling of anomaly handling. Specifically: Generate a dynamic maintenance queue Q from the results of steps S51 and S52: Queues by priority P i Sorting, while embedding collaborative processing strategy AS i After each maintenance operation, update the node impact. Cluster risk ; The maintenance queue is iteratively updated based on the latest status to achieve closed-loop dynamic scheduling. S54. Quantitatively evaluate the effectiveness of each collaborative processing step, calculate the reduction in the impact of node anomalies, and feed the results back to risk scoring and priority ranking. Dynamically adjust the collaborative processing strategy and weighting coefficients based on the feedback to achieve closed-loop optimization, ensuring the anomaly management method has adaptive and continuous optimization capabilities, and improving the effectiveness of multi-node collaborative anomaly handling and overall battery consistency. The effectiveness of each collaborative processing step is quantitatively evaluated: ; Among them, QE i This indicates the amount by which the impact of the anomaly has decreased; This represents the global impact before node processing, serving as a reference benchmark to measure the processing effect; It represents the global impact after node processing, reflects the anomaly mitigation effect, and is used to provide feedback on adjustment strategies; The evaluation results are fed back into step S44 (risk scoring) and step S51 (priority ranking) to achieve closed-loop optimization: if the processing effect of some nodes is insufficient, the weights can be adjusted. , , Or collaborative processing strategy AS i .

[0022] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for locating abnormal battery cell in a battery consistency-oriented multi-algorithm cooperative manner, characterized in that, The specific implementation steps include the following: S1, by multi-channel synchronous acquisition of the voltage, temperature, terminal current and internal resistance data of the battery cell, and combining short-time perturbation enhancement and sliding window processing technology, time alignment, filtering and denoising and standardization integration of multi-modal data are performed to construct an input feature matrix; S2, an algorithm pool is constructed, the input feature matrix is analyzed by multiple algorithms in parallel to generate a multi-dimensional evidence signature for each battery cell, and the holographic evidence matrix is constructed by combining the uncertainty and cross-algorithm consistency information output by the algorithm; S3, based on the holographic evidence matrix, the contribution weight of each algorithm is dynamically allocated by game type weight scheduling, a weighted consensus graph between battery cells is constructed, and an abnormal prompt list is generated by combining node centrality analysis and edge weight clustering; S4, based on the weighted consensus graph and the abnormal prompt list, the abnormal relationship between battery cells is quantitatively analyzed by local and global collaborative reasoning, the abnormal cluster is identified, and the battery cells in the cluster are hierarchically classified to generate a dynamic risk score; S5, based on the hierarchical classification and dynamic risk score of the abnormal cluster, the comprehensive priority of each abnormal node is calculated to form a priority processing queue, and the priority processing and optimization scheduling of multi-node abnormalities are completed by combining the collaborative optimization processing strategy and the closed-loop effect feedback.

2. The method of claim 1, wherein, The construction of the input feature matrix specifically includes: Synchronously collecting the terminal voltage, monomer temperature and terminal current of all battery cells in the battery pack; Introducing the perturbation current to measure the internal resistance, and calculating the instantaneous internal resistance estimate value of each battery cell based on the internal resistance estimation formula; Time alignment and sliding window filtering and denoising are performed on the collected multi-channel signals to obtain filtered voltage signals; A perturbation current or voltage pulse is applied and the sampling window is extended to calculate the short-time energy and enhanced internal resistance response of each battery cell to amplify the small abnormal features; The original and enhanced signals are standardized, and the voltage, temperature, internal resistance, short-time energy and perturbation response information are integrated to form a unified multi-modal feature matrix.

3. The method of claim 2, wherein, The construction of the holographic evidence matrix specifically includes: An algorithm pool is constructed, which includes a time domain residual detector, a frequency domain harmonic structure detector, a local coupling synchronization detector and a lightweight embedded nonlinear feature model; Each algorithm independently processes the signals of each battery cell, calculates the abnormal index and outputs the corresponding uncertainty estimate; The output vectors of all algorithms for each battery cell are integrated to generate a multi-dimensional evidence signature containing an abnormal index vector and an uncertainty vector; The similarity of the multi-algorithm output of each battery cell is calculated to quantify the consensus degree of different algorithms for the same abnormality to form an inter-algorithm consistency matrix; The evidence signature and consistency information of all battery cells are integrated into the final holographic evidence matrix.

4. The method of claim 3, wherein, The dynamic allocation of the contribution weight of each algorithm by game type weight scheduling specifically includes: The weight of each algorithm is initialized for the multi-algorithm evidence vector of each battery cell; A local game model is constructed, and each algorithm is regarded as a game participant, and its payoff function integrates the abnormal index, uncertainty and consistency average of the algorithm with other algorithms; The weight of each algorithm for each battery cell is dynamically adjusted by an iterative optimization method, so that the weight is inclined to the algorithm with higher credibility and consistency for abnormal judgment, and the weight converges or reaches the maximum iteration times.

5. The method of claim 4, wherein, The constructing of the weighted consensus graph between the battery cells specifically includes: calculating the weighted similarity of the abnormal features between the battery cells by using the optimized algorithm weight, as the edge weight value of the consensus graph; calculating the comprehensive abnormal intensity of each battery cell itself as the node self-loop weight value of the consensus graph; based on the node self-loop weight value and the edge weight value of all battery cells, a complete weighted consensus graph is constructed.

6. The multi-algorithm coordinated abnormal cell positioning method for battery uniformity according to claim 5, characterized in that, The quantitative analysis of the abnormal relationship between the battery cells through local and global collaborative reasoning specifically includes: calculating the local abnormal influence degree of each battery cell node, which integrates the weighted abnormal index of the node itself and the collaborative abnormal influence of its neighbor nodes; based on the local abnormal influence degree, the global cumulative calculation is used to simulate the propagation path of the abnormality between the battery cells, and the global abnormal influence degree of each node is obtained.

7. The method of claim 6, wherein, Identifying abnormal clusters and classifying battery cells in the cluster specifically includes: selecting node pairs with edge weight values exceeding a preset threshold from the consensus graph to form a high correlation battery cell pair set; performing weighted clustering on the high correlation battery cell pairs to identify abnormal clusters with significant collaborative relationship; according to the global abnormal influence degree of each battery cell node, using a preset global influence degree threshold, the battery cells in the cluster are divided into primary abnormality, secondary abnormality and tertiary abnormality.

8. The method of claim 7, wherein, Generating dynamic risk score specifically includes: within each abnormal cluster, according to the global abnormal influence degree of the node and the node degree, the core nodes and boundary nodes are identified; calculating the internal tightness index of each abnormal cluster, which is based on the sum of the weight values of all edges in the cluster; comprehensive average global influence degree of core nodes and cluster tightness index, calculate the comprehensive risk score of each abnormal cluster.

9. The method of claim 8, wherein, Forming a priority processing queue, and combining the collaborative optimization processing strategy and the closed loop effect feedback, specifically includes: comprehensive global abnormal influence degree of node, risk score of abnormal cluster to which it belongs and hierarchical level, calculate the comprehensive priority of each abnormal node; sorting the abnormal nodes according to the comprehensive priority and generating a dynamically maintained queue; for the nodes in the queue, combined with the collaborative influence of its neighbor nodes and the maintenance resource constraints, the collaborative optimization processing strategy is generated; after each maintenance execution, update the global influence degree of the node and the cluster risk, and iteratively adjust the maintenance queue; quantitative evaluation is carried out on the effect of each collaborative processing, and the evaluation results are fed back to the risk score and priority calculation, realizing closed loop optimization.

Images