An energy storage battery multi-level voltage consistency detection method and system for adaptive anomaly detection

Abstract

The application discloses a kind of self-adapting abnormality detection energy storage battery multilevel voltage consistency detection method and system, the method includes: collecting the voltage time series data of each battery in each module in multiple battery clusters under the same operating condition;According to the order of battery level, module level and cluster level, the features of each level are extracted from bottom to top, after completing feature extraction, an adaptive multi-granularity isolation forest model is constructed, and abnormality detection is performed layer by layer from top to bottom according to the order of cluster level, module level and battery level;The construction process of the adaptive multi-granularity isolation forest model includes: in the training process of each level node, introduce feature-aware weighted splitting mechanism;On the basis of weighted splitting, a three-layer detection architecture directly corresponding to the physical level of energy storage battery is formed;Introduce a completely data-driven operating condition adaptive contamination rate mechanism.The application realizes high-precision, low-false alarm real-time monitoring of cluster-module-battery multilevel voltage consistency in large-scale energy storage power station.

Description

Technical Field

[0001] This invention belongs to the field of battery management system state monitoring and anomaly diagnosis technology, specifically a method and system for adaptive anomaly detection of multi-level voltage consistency in energy storage batteries. Background Technology

[0002] Battery systems typically consist of hundreds or thousands of cells connected in series and parallel. However, due to factors such as manufacturing process deviations, differences in assembly, temperature gradients during operation, uneven charge and discharge rates, and inconsistent aging and degradation rates, parameters such as voltage, internal resistance, state of charge (SOC), and state of health (SOH) gradually become inconsistent among the cells. This inconsistency not only reduces the system's usable capacity and cycle efficiency, but in severe cases, it may also induce safety accidents such as local overcharging, internal short circuits, thermal runaway, and even fires and explosions, threatening the safety of people and property and the stable operation of the power grid.

[0003] Existing battery conformity testing technologies mainly include the following: (1) Simple statistical method based on threshold: Judgment is made by setting empirical thresholds such as the mean voltage and standard deviation. This method is simple to implement, but the threshold is highly dependent on human experience. It has poor generalization ability under different battery types and different operating conditions, and is prone to missed judgments or misjudgments, especially with low sensitivity to slight inconsistencies.

[0004] (2) Clustering-based methods: such as K-means, DBSCAN and other unsupervised clustering algorithms, which cluster the parameters of individual battery cells into several classes, and those that deviate from the main class are considered inconsistent. This type of method is sensitive to the number of clusters, requires preset parameters, and the clustering effect is unstable when the data distribution is complex or the inconsistency is weak.

[0005] (3) Dimensionality reduction methods based on principal component analysis (PCA), correlation coefficient, or incremental capacity (IC) curves: anomalies are identified by reconstructing the error or correlation matrix after dimensionality reduction. This method is sensitive to linear relationships, while battery inconsistencies often exhibit strong nonlinear characteristics, leading to insufficient detection accuracy.

[0006] More importantly, all the above methods have the following inherent limitations: First, they treat each cell as an independent entity for analysis, ignoring the objectively existing mutual influence relationships between cells. In actual battery systems, cells form a complex network of mutual influence through electrical connections, thermal conduction, and shared loads. Voltage deviations in one cell can propagate to adjacent cells through internal resistance coupling and thermal effects. Ignoring this mutual influence not only makes it difficult to detect early, weak anomalies, but also makes it impossible to distinguish between "independent single-cell anomalies" and "systemic inconsistencies," two fundamentally different types of anomalies, leading to a fundamental lack of information for operation and maintenance decisions. Second, they do not fully utilize the inherent "cluster-module-cell" hierarchical topology of energy storage battery systems, failing to achieve layer-by-layer localization from the system level. Existing methods struggle to effectively distinguish between the four different types of anomalies: "independent single-cell anomalies," "module-level inconsistencies," "cluster-level inconsistencies," and "systemic inconsistencies." Third, the detection thresholds of existing methods largely rely on manual experience, making it difficult to adapt to changes in operating conditions and data distribution.

[0007] Therefore, there is an urgent need for a voltage consistency detection method that can fully utilize the hierarchical topology, adaptively determine the detection threshold, and balance detection accuracy and computational efficiency. Summary of the Invention

[0008] The purpose of this invention is to provide an adaptive anomaly detection method and system for multi-level voltage consistency detection of energy storage batteries. Through multi-level feature extraction and adaptive multi-granularity isolation forest, it achieves high-precision, low-false-alarm real-time monitoring of multi-level voltage consistency of large-scale energy storage power station clusters-modules-cells, thereby solving fundamental problems in the prior art such as treating cells as independent individuals, not fully utilizing the hierarchical topology, difficulty in distinguishing anomaly types, and relying on manual setting of detection thresholds.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution.

[0010] In a first aspect, the present invention provides an adaptive anomaly detection method for multi-level voltage consistency detection of energy storage batteries, comprising: The voltage time series data of each cell in each module within multiple battery clusters of the energy storage battery under test are collected under the same operating conditions, and the hierarchical topology information of battery cluster-module-cell is obtained at the same time. Features are extracted from the bottom up in the order of cell level, module level and cluster level. After feature extraction is completed, an adaptive multi-granularity isolation forest model is constructed, and anomaly detection is performed layer by layer from top to bottom in the order of cluster level, module level and cell level. The construction process of the adaptive multi-granularity isolation forest model includes: introducing a feature-aware weighted splitting mechanism during the training of nodes at each level, determining the feature selection weight based on the variance contribution rate of each feature in the training dataset, and increasing the probability of features with higher variance contribution rates being randomly selected during node splitting, thus giving features sensitive to cell consistency differences a higher chance of participating in splitting; based on the weighted splitting, a coarse-fine granularity hierarchical strategy is further adopted to construct an adaptive multi-granularity isolation forest, forming a three-layer detection architecture directly corresponding to the physical level of the energy storage battery; and introducing a fully data-driven operating condition adaptive contamination rate mechanism, whereby the contamination rate parameter of the isolation forest is adaptively determined based on the abnormal score distribution of historical normal data under each operating condition, without the need for manual setting of operating condition correction coefficients.

[0011] Furthermore, the voltage time series data is divided into sliding windows and subjected to feature standardization to obtain voltage time series segments.

[0012] Furthermore, at the first cell level, statistical features are extracted from the voltage time series segments of each cell within each window to characterize the basic operating state of a single cell. At the second module level, based on the cell-level statistical features and voltage time series segments, coupling relationship features between cells within a module are extracted to characterize the degree of consistency deviation between cells within a module. At the third cluster level, based on the battery cluster, the mutual information coupling strength in the coupling relationship features is aggregated by module to construct a coupling strength matrix between modules within a cluster. This matrix is ​​normalized and used as edge weights, and modules are used as nodes to construct a weighted undirected graph. The graph topology features of each cluster are extracted and combined with cluster-level aggregated statistics to characterize the overall consistency state and anomaly propagation topology at the cluster level.

[0013] Furthermore, the statistical features include the voltage mean, voltage standard deviation, voltage range, voltage coefficient of variation, and voltage rate of change mean; the coupling relationship features include relative deviation, coupling follow-up deviation, and mutual information coupling strength; the graph topology features include weighted degree centrality, betweenness centrality, and local clustering coefficient; and the cluster-level aggregation statistics include intra-cluster module voltage dispersion, intra-cluster voltage range extreme values, and intra-group correlation coefficient.

[0014] Furthermore, the three-layer detection architecture is as follows: The first layer is a coarse-grained isolation forest at the battery cluster level, which uses the median vector of all module features within each cluster as the aggregation representation of that cluster to quickly screen potential candidate anomalous clusters; the second layer is a medium-grained isolation forest at the module level, which uses the median vector of all cell features within each module as the aggregation representation of that module to screen candidate anomalous modules for candidate anomalous clusters; the third layer is a fine-grained isolation forest at the cell level, which directly uses the feature vector of a single cell for candidate anomalous modules to achieve accurate localization of anomalous cells. By aggregating layer by layer, computational resources are concentrated on candidate anomalous regions, reducing the overall detection computational overhead in large-scale energy storage system scenarios.

[0015] Furthermore, for the aforementioned abnormal cells, a three-level progressive fine judgment is performed sequentially at the cell level, module level, and cluster level: at the cell level, the deviation contribution is calculated to identify independent abnormal cells; at the module level, the proportion of abnormal cells within the module is calculated to determine module-level degradation; and at the cluster level, the distribution ratio of abnormal modules within the cluster is calculated to determine cluster-level degradation. The coupling propagation index is combined to distinguish between isolated deviations and precursors of systemic degradation, and finally, the final anomaly determination result is obtained through time-series persistence constraints.

[0016] Furthermore, the adaptive pollution rate mechanism for each operating condition is as follows: For three operating conditions—charging, discharging, and resting—the abnormal score distribution of historical normal operating data is statistically analyzed for each condition. The median and absolute deviation of the abnormal score distribution for each condition are calculated. The median plus the median absolute deviation by a preset fixed multiple is used as the pollution rate threshold for that condition. The preset fixed multiple is a built-in general statistical constant, calculated independently for different operating conditions. The specific value of the pollution rate threshold is entirely adaptively determined by the data distribution under each operating condition, without the need for manually setting empirical threshold parameters strongly correlated with a specific battery system.

[0017] Furthermore, the method for detecting the multi-level voltage consistency of energy storage batteries also includes: Based on the anomaly scores and anomaly detection results, the status of clusters, modules, and cells is classified into different levels, forming a top-down hierarchical diagnostic conclusion and generating a test report containing visual charts.

[0018] Furthermore, the specific steps for classifying the status of clusters, modules, and cells are as follows: At the cell level, based on the statistical characteristics of the distribution of abnormal scores in historical normal operation data, the cell status is divided into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent, according to a preset four-level classification framework. The classification thresholds are adaptively determined by the distribution characteristics of historical data at the corresponding level. At the module level, the status is divided into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent, based on the proportion of abnormal cells within the module and the overall abnormal score of the module. At the cluster level, the status is divided into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent, based on the spatial distribution of abnormal modules within the cluster and the overall abnormal index of the cluster. The classification thresholds for each level are adaptively determined based on the data distribution, without the need for manually setting fixed values.

[0019] Secondly, the present invention provides an adaptive anomaly detection system for multi-level voltage consistency detection of energy storage batteries, used to implement the above-mentioned multi-level voltage consistency detection method for energy storage batteries, comprising: The data acquisition module is used to collect voltage time series data of each cell in each module within multiple battery clusters of the energy storage battery under test under the same operating conditions, and at the same time obtain the hierarchical topology information of battery cluster-module-cell. The data preprocessing module performs sliding window partitioning and feature standardization on the voltage time series data to obtain voltage time series segments; The multi-level feature extraction module includes: a cell-level statistical feature extraction unit, used to extract the mean voltage, standard deviation, range, coefficient of variation, and mean voltage change rate of each cell; a module-level coupling feature extraction unit, used to extract the relative deviation, coupling following deviation, and mutual information coupling strength between cells within the module; and a cluster-level graph feature extraction unit, used to construct a coupled weighted undirected graph and extract weighted degree centrality, betweenness centrality, local clustering coefficient, and cluster-level aggregation statistics. The adaptive multi-granularity isolation forest detection module integrates a feature-aware weighted splitting unit, a cluster-level coarse-grained isolation unit, a module-level medium-grained isolation unit, a cell-level fine-grained isolation unit, and an operating condition adaptive pollution rate adjustment unit. The pollution rate parameter is adaptively determined based on the abnormal score distribution of historical normal data under each operating condition. The hierarchical progressive judgment module is used to sequentially perform cell-level deviation contribution judgment, module-level abnormal cell proportion judgment, and cluster-level abnormal module distribution judgment. It also combines the coupling propagation index to distinguish between isolated deviations and systematic inconsistencies, and outputs the final anomaly judgment result through time-series persistence constraints.

[0020] The beneficial effects of this invention are as follows: This invention utilizes a "cluster-module-cell" hierarchical topology to construct a multi-level feature system based on cell coupling relationships from the bottom up, solving the problems of ignoring the mutual influence between cells and failing to detect weak anomalies; Based on weighted splitting, this invention employs a coarse-fine granular layering strategy to construct an adaptive multi-granularity isolation forest, forming a three-layer detection architecture directly corresponding to the physical level of the energy storage battery, and introduces a fully data-driven adaptive contamination rate mechanism to solve the problems of detection thresholds relying on human experience and being unable to generalize across operating conditions; Through multi-level feature extraction and adaptive multi-granularity isolation forest, this invention achieves high-precision, low-false-alarm real-time monitoring of multi-level voltage consistency across clusters, modules, and cells in large-scale energy storage power stations. Attached Figure Description

[0021] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a flowchart of an adaptive anomaly detection method for multi-level voltage consistency detection of energy storage batteries according to the present invention. Figure 2 This is a schematic diagram of the three-layer detection architecture of the adaptive multi-granularity isolation forest of the present invention; Figure 3 This is a cluster-module layer hierarchical coupling anomaly screening diagram for the present invention; Figure 4 This is a graph showing the difference between the voltage trajectories of a typical abnormal cell and a group of cells. Figure 5 This is a schematic diagram of the multi-level voltage consistency detection system for energy storage batteries according to the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] Example 1 This embodiment describes an adaptive anomaly detection method for multi-level voltage consistency detection in energy storage batteries. Figure 1 As shown, the steps are as follows: Step 1) Collect voltage time series data of each cell in each module within multiple battery clusters under the same operating conditions from the Battery Management System (BMS), and simultaneously obtain the hierarchical topology information of battery cluster-module-cell. Step 2) Divide the voltage time series data into sliding windows and standardize the features extracted in each window to obtain voltage time series segments; Step 3) Extract statistical features from the voltage time series segments of each cell within each window to characterize the basic operating state of a single cell; Step 4) Calculate the coupling relationship characteristics between cells within a module, using the module as the unit, to characterize the degree of consistency deviation between cells within the module. Step 5) Using battery clusters as units, aggregate the mutual information coupling strength in the coupling relationship features by module to construct a coupling strength matrix between modules within a cluster. Based on the coupling strength matrix, construct a weighted undirected graph, extract the graph topology features of each cluster, and combine the voltage dispersion of modules within the cluster, the extreme value of voltage range within the cluster, and the correlation coefficient within the group to characterize the overall consistency state and anomaly propagation topology at the cluster level. Step 6) After completing the feature extraction at the cell level, module level and cluster level, an adaptive multi-granularity isolation forest model is used for anomaly detection. This includes running a coarse-grained isolation forest at the battery cluster level to screen candidate abnormal clusters, running a medium-grained isolation forest at the module level within the candidate clusters to screen candidate abnormal modules, and running a fine-grained isolation forest at the cell level within the candidate modules to screen candidate abnormal cells. Step 7): For the candidate abnormal cells output in Step 6), perform a three-level progressive fine judgment at the cell level, module level, and cluster level in sequence: at the cell level, calculate the deviation contribution to identify independent abnormal cells; at the module level, calculate the proportion of abnormal cells within the module to determine module-level degradation; at the cluster level, calculate the distribution ratio of abnormal modules within the cluster to determine cluster-level degradation; and combine the coupling propagation index to distinguish between isolated deviations and systemic degradation precursors. Finally, output the final anomaly judgment result through time-series persistence constraints. Step 8) Based on the anomaly score and anomaly detection results, classify the status of clusters, modules and cells respectively, form a top-down hierarchical diagnostic conclusion, and generate a test report containing visual charts.

[0025] In step 1), the data is indexed by timestamps, with each column corresponding to the voltage sampling value of a single battery cell, forming... N × T The voltage matrix, where N Indicates the number of battery cells. T This indicates the number of sampling points. The collected data serves as the basic input for subsequent consistency analysis and anomaly detection.

[0026] In step 2), the Z-score standardization method is used to perform dimensionless processing on each feature dimension. The standardized parameters are saved in the training phase and reused in the online inference phase.

[0027] Based on the above multi-level feature system, features are extracted progressively at the three levels of cell level, module level and cluster level in steps 3) to 5) to provide multi-granularity feature input for subsequent top-down hierarchical detection.

[0028] At the first cell level, statistical features are extracted for the voltage time series segments of each cell within each window, including basic statistics such as voltage mean, voltage standard deviation, voltage range, voltage coefficient of variation, and average voltage change within the window. These are the basic statistical distribution features used to characterize the operating status of a single cell.

[0029] At the second-level module, based on cell-level statistical characteristics and voltage time series segments, coupling relationship characteristics between cells within a module are extracted, taking the module as the unit. As a key level connecting clusters and cells, the coupling relationship characteristics of the module can reflect the propagation state of anomalies within the cell group.

[0030] First, to measure the degree of deviation of an individual cell within the module group, a relative deviation is constructed. This indicator can reflect the deviation of the target cell from the median level of its group.

[0031]

[0032] in, Let be the average voltage of cell i, median(V) be the median of the average voltage of all cells, and MAD(V) be the absolute deviation of the median.

[0033] Furthermore, to characterize the consistency between the dynamic change trend of the battery cells and the population, the coupling following deviation is calculated. The larger the value, the more severe the deviation from the group trend.

[0034] in, Indicates battery cell i With battery cells j The Pearson correlation coefficient between voltage time series. When the target cell follows the same trend as the overall trend, the coupling following deviation is close to zero; as it gradually deviates from the overall trend, this indicator will increase significantly.

[0035] Furthermore, to further capture the nonlinear correlation between cells, a k-nearest neighbor mutual information estimator is used to calculate the mutual information coupling strength. :

[0036] in, Indicates battery cell i With battery cells jMutual information between cells can effectively characterize the nonlinear information transmission capability and the strength of coupling relationship between cells.

[0037] At the third cluster level, taking battery clusters as units, the mutual information coupling strength in the coupling relationship features is aggregated by module to construct a coupling strength matrix between modules within a cluster. This matrix is ​​normalized and used as edge weights, and modules are used as nodes to construct a weighted undirected graph. The graph topology features of each cluster are extracted and combined with cluster-level aggregated statistics to characterize the overall consistency state and anomaly propagation topology at the cluster level. The cluster-level aggregated statistics include the voltage dispersion of modules within the cluster, the extreme values ​​of voltage ranges within the cluster, and the correlation coefficient within the cluster.

[0038] A weighted undirected graph G=(V,E,W) is constructed based on the coupling strength matrix of all cells in the cluster, where the node set V represents all cells in the cluster, and the edge weight W is the normalized coupling strength matrix between cells. Graph features for cluster-level state representation are extracted based on this graph.

[0039] First, extract the structural centrality index of the graph – weighted degree centrality. :

[0040] in, Indicates battery cell i With battery cells j The normalized coupling weights between them. This metric is used to measure the propagation capability of the impact of individual cell deviations on the overall network.

[0041] Based on this, betweenness centrality based on the shortest path is further extracted. :

[0042] in, s and t Let be any two distinct nodes in the graph. For nodes s arrive t The total number of shortest paths, The number of paths passing through node i is used to characterize the global bridging role of the battery cell in the deviation propagation network.

[0043] Recalculate the local clustering coefficients:

[0044] in, ka For nodes a The number of neighbors of a node, E(N(a)) is the number of neighbors of a node. a The actual number of edges between the neighbors of a node, where Ca is the number of nodes. a The local clustering coefficient is used to reflect abnormal local clustering tendencies.

[0045] After feature extraction, all features are standardized. The Z-score standardization method is used.

[0046] in, μ and σ These represent the mean and standard deviation of the corresponding features in the training set, respectively. By standardizing the dimensions and numerical ranges of each feature, the stability and convergence efficiency of the subsequent model training process are improved. The standardized parameters are saved during the training phase and reused during the online inference phase.

[0047] It should be noted that feature extraction and anomaly detection in this invention adopt opposite hierarchical directions. Feature extraction proceeds from bottom to top, from cell level to module level to cluster level. This is because the calculation of module-level coupling features requires voltage statistics of individual cells as input, while the construction of cluster-level graph features requires the module-level coupling strength matrix as the source of edge weights. The feature system naturally relies on the aggregation of low-level features to high-level features layer by layer, so a bottom-up order must be followed. Anomaly detection is performed layer by layer from top to bottom, from cluster to module to cell. If all cells are detected one by one, the computational complexity increases quadratically with the number of cells, which is difficult to meet the real-time requirements of large-scale energy storage systems. By adopting a top-down strategy, most normal areas are first screened out at the cluster level with coarse granularity, then modules within candidate anomaly clusters are screened with medium granularity, and finally cells within a small number of candidate anomaly modules are precisely located with fine granularity. This allows computational resources to be concentrated on suspected anomaly areas, significantly reducing the overall detection overhead.

[0048] In step 6), after completing the bottom-up feature extraction at the cell level, module level, and cluster level, an adaptive multi-granularity isolation forest model is constructed, and anomaly detection is performed according to the top-down hierarchical progression strategy.

[0049] The construction process of the adaptive multi-granularity isolation forest model includes: introducing a feature-aware weighted splitting mechanism during the training of nodes at each level, instead of using the completely random equal-probability splitting method of traditional isolation forests, determining the feature selection weights based on the variance contribution rate of each feature in the training dataset. This mechanism can increase the probability of key consistency features being selected, thereby enhancing the model's ability to distinguish consistency anomalies; based on the weighted splitting, a coarse-fine granularity hierarchical strategy is further adopted to construct an adaptive multi-granularity isolation forest, forming a three-layer detection architecture directly corresponding to the physical level of the energy storage battery; and introducing a fully data-driven operating condition adaptive contamination rate mechanism, whereby the contamination rate parameter of the isolation forest is adaptively determined based on the distribution of abnormal scores of historical normal data under each operating condition, without the need for manual setting of operating condition correction coefficients.

[0050] like Figure 2As shown, the three-layer detection architecture consists of: a first layer, a coarse-grained isolation forest at the battery cluster level, which uses the median vector of all module features within each cluster as the aggregation representation of that cluster to quickly filter potential candidate anomalous clusters; a second layer, a medium-grained isolation forest at the module level, which uses the median vector of all cell features within each module as the aggregation representation of that module to filter candidate anomalous modules; and a third layer, a fine-grained isolation forest at the cell level, which directly uses the feature vector of a single cell to accurately locate the anomalous cell for each candidate anomalous module. By aggregating layer by layer, computational resources are concentrated on candidate anomalous regions, reducing the overall detection computational overhead in large-scale energy storage system scenarios.

[0051] The adaptive pollution rate mechanism described is as follows: For three operating conditions—charging, discharging, and resting—the abnormal score distribution of historical normal operating data is statistically analyzed for each condition. The median and median absolute deviation (MAD) of the abnormal score distribution for each condition are calculated. The median plus a preset fixed multiple of MAD is used as the pollution rate threshold for that condition. This preset fixed multiple is a built-in general statistical constant, calculated independently for different operating conditions. The specific value of the pollution rate threshold is entirely adaptively determined by the data distribution under each condition, eliminating the need for manual threshold adjustment for different battery systems. The core advantage of this mechanism is that while the general statistical constant remains constant, the specific value of the pollution rate threshold automatically evolves with the battery system's operating state and cell aging process, fundamentally eliminating the lack of adaptability caused by manually setting the threshold.

[0052] The specific steps for determining hierarchical anomalies in step 7) are as follows: (1) In the cell-level determination, the deviation contribution is the ratio of the abnormal score of the target cell to the highest abnormal score of the cell in the same module other than the target cell. When the ratio exceeds the critical value determined adaptively based on the statistical distribution of the ratio in historical normal data, it is determined to be an independent abnormality of a single cell. Deviation Contribution Ratio (DCR):

[0053] in, This indicates the abnormal score of the target battery cell. This represents the highest abnormal score within the same module, excluding the target cell. The DCR judgment threshold is adaptively determined based on the statistical distribution of DCR values ​​in historical normal data (e.g., taking the median plus twice the MAD). When the DCR exceeds this threshold, it is determined that the target cell has obvious independent abnormal characteristics, indicating that a single cell failure may occur.

[0054] (2) In the module-level determination, the proportion of abnormal cells in the module where the target cell is located is statistically analyzed. When the proportion exceeds the threshold determined adaptively based on the random fluctuation range in historical normal data, it is determined to be a module-level inconsistency. Percentage of abnormal cells within the module containing the target cell :

[0055] Where Nabnormal represents the number of abnormal battery cells, and Ntotal represents the total number of battery cells. The judgment threshold is adaptively determined based on the random fluctuation range of the proportion of abnormal battery cells in the module from historical normal data. If the threshold is exceeded, the module is determined to have module-level inconsistency.

[0056] (3) In the cluster-level determination, the proportion of abnormal modules in the cluster where the target cell is located is statistically analyzed. When the proportion exceeds the threshold determined adaptively based on the random fluctuation range in historical normal data, it is determined to be a cluster-level inconsistency. Distribution ratio of abnormal modules within the cluster where the target cell is located :

[0057] Among them, M abnormal M represents the number of abnormal modules. total The total number of modules; similarly, the judgment threshold is adaptively determined based on the random fluctuation range of historical normal data, when... If the threshold is exceeded, the cluster is determined to have cluster-level inconsistency.

[0058] (4) In the distinction of abnormal types, the coupling propagation index is the proportion of the number of adjacent cells that are strongly coupled with the target cell and have synchronously experienced abnormalities to the total number of its strongly coupled neighbors. When the proportion is lower than the threshold determined adaptively based on the statistical distribution of the index in historical normal data, it is judged as an isolated deviation. When it is not lower than the threshold, it is judged as a systematic inconsistency.

[0059] Coupling Propagation Index (CPI):

[0060] Where, N co-anomalous N represents the number of cells that are strongly coupled to the target cell (defined as having a mutual information value higher than the median of the mutual information distribution within the cluster + 1 MAD) and that exhibit synchronous anomalies. strongly-coupledThis represents the total number of strongly coupled neighbors. When the CPI is lower than the preset threshold, it indicates that the anomaly is mainly concentrated in a single cell, which is judged as an isolated deviation, suggesting possible sensor failure or poor contact. When the CPI is not lower than the preset threshold, it indicates that the anomaly has propagated along the coupling relationship, which is judged as a systemic inconsistency, suggesting possible thermal imbalance or consistency degradation.

[0061] In addition, to avoid transient noise interference, a timing continuity verification mechanism is introduced: the final alarm is triggered only when the candidate cell meets the abnormal conditions within T consecutive sampling windows, where T is a configurable positive integer, preferably 3.

[0062] In step 8), the status of cells, modules and clusters is classified according to the final abnormal score and the hierarchical progressive judgment result, that is, multi-level status classification.

[0063] Cell-level grading is based on the statistical characteristics of the distribution of abnormal scores in historical normal operation data, classifying cell status into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent according to built-in fixed grading criteria. Module-level grading comprehensively considers the proportion of abnormal cells within the module and the overall abnormal score of the module, classifying it into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent. Cluster-level grading combines the spatial distribution characteristics of abnormal modules within the cluster and the cluster-level comprehensive abnormal index to achieve four levels of status determination: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent. The threshold values ​​for each grading level are adaptively determined based on the distribution characteristics of historical data at the corresponding level, eliminating the need for manually setting fixed values.

[0064] Simultaneously, it generates visualization results such as cluster-module layer hierarchical coupling anomaly screening diagram and anomaly score time series heat map, realizing intuitive display of the consistency status of energy storage batteries and auxiliary decision-making.

[0065] The following is a detailed implementation process of the above-mentioned multi-level voltage consistency detection method for energy storage batteries.

[0066] Data acquisition and preprocessing are performed according to the above method, wherein the sliding window length is preferably set to 60 min, the sliding step size is set to 10 min, and the Z-score method is used for standardization. The standardization parameters are saved during the training phase and reused during the online inference phase.

[0067] Feature extraction is performed in a bottom-up, progressive manner at the cell level, module level, and cluster level, providing multi-granularity feature inputs for subsequent top-down (cluster → module → cell) layered detection. At the cell level, the mean voltage, standard deviation, range, coefficient of variation, and mean voltage change rate are extracted. At the module level, relative deviation, coupling following deviation, and mutual information coupling strength are extracted. At the cluster level, based on the coupled weighted network graph, weighted degree centrality, betweenness centrality, local clustering coefficient, and cluster-level aggregation statistics (module voltage dispersion, voltage range extreme values, intra-group correlation coefficient, ICC, etc.) are extracted.

[0068] The construction parameters of the adaptive multi-granularity isolation forest model are as follows: the forest size is set to 300 independent trees, the sampling limit for each tree is 256 samples, and the training set is used to complete the model training. In feature-aware weighted splitting, the feature weights are determined by the variance contribution rate of each feature fitted on the training set. In coarse-grained hierarchical aggregation, the cluster level uses the median value vector of all module features in the cluster as the aggregation representation, the module level uses the median value vector of cell features within the module as the aggregation representation, and the cell level directly uses the single cell feature vector. Figure 3 The coupling topology and anomaly propagation relationships between the cluster and module levels are illustrated. Figure 4 The differences between the voltage trajectories of typical abnormal cells and cell groups are shown. The two figures assist in verifying the location results of abnormal cells from two aspects: hierarchical anomaly screening and cell voltage deviation.

[0069] like Figure 3 As shown in the diagram, the upper layer is the battery cluster level. The median vector of all module features within each cluster is used as the aggregation representation of that cluster. Anomaly scores are calculated using a coarse-grained isolation forest at the cluster level. Candidate anomalous clusters are selected based on a threshold adaptively determined by the anomaly score distribution of historical data under this operating condition. These candidate anomalous clusters are marked with dark fill or thick borders. The lower layer is the module level. Further medium-grained filtering is performed only on modules located within candidate anomalous clusters. The median vector of all cell features within a module is used as the aggregation representation of that module. Anomaly scores are calculated using a module-level isolation forest, and candidate anomalous modules are selected and also marked with dark fill. Solid lines connect cluster nodes and module nodes to represent hierarchical relationships. Lines of varying thickness between module nodes represent coupling strength based on mutual information; thicker lines indicate stronger coupling. This visually presents the topological structure of anomalies propagating along strong coupling paths between the cluster and module levels. Figure 4 As shown in the figure, the horizontal axis represents operating time, the vertical axis represents cell voltage, the gray area represents the 5% to 95% percentile range of the cell group, the median curve represents the overall voltage change trend of the group, and the voltage trajectory of abnormal cells is plotted. As can be seen from the figure, some located cells exhibit varying degrees of voltage deviation from the group's change trend during specific operating phases. By comparing with the group's percentile range and median, the abnormal voltage response of the target cell can be intuitively reflected. Further verification was conducted based on actual BMS data from 157 effective operating days and 1354 analysis windows. After adopting an adaptive threshold based on operating conditions and continuous judgment for three consecutive windows, the final alarm rate was 0.66%, the detection rate of moderate and above abnormal events was 90.0%, and the cell-level location hit rate was 82.2%. These results indicate that the proposed method can effectively suppress false alarms caused by instantaneous anomalies by utilizing multi-level progressive screening and temporal continuity constraints, and achieve the detection and cell-level location of abnormal cells.

[0070] In the adaptive contamination rate mechanism, the median of the abnormal score distribution of historical normal data under each operating condition, plus three times the median absolute deviation (median + 3 × MAD), is taken as the contamination rate threshold for that operating condition. The three operating conditions are calculated independently and automatically updated with the running data. Three times MAD is a robust statistical criterion commonly used in the field of anomaly detection, which can achieve anomaly boundary identification effect similar to the traditional 3σ method, while having better robustness to outliers, without the need for parameter adjustment for different battery systems or application scenarios.

[0071] In the hierarchical anomaly detection, each threshold is adaptively determined based on the statistical distribution of the corresponding indicator in historical normal data, without the need for manually setting fixed values. Specifically: the critical value for deviation contribution rate (DCR) is the median of the historical normal data DCR distribution plus twice the MAD; the threshold for the proportion of abnormal cells within a module is the global pollution rate plus twice the standard deviation of the binomial distribution; the threshold for the proportion of abnormal modules within a cluster is determined similarly; the threshold for the coupling propagation index (CPI) is the median of the historical normal data CPI distribution plus twice the MAD; the time series persistence window number T is set to 3 (T is a built-in fixed parameter of the method), and each threshold is automatically updated as historical data accumulates.

[0072] In the multi-level status classification, the threshold values ​​for each level are adaptively determined based on the statistical distribution of the corresponding level's indicators in historical normal data. At the cell level, the median, median+1×MAD, and median+2×MAD of the abnormal score distribution serve as three-level boundaries, classifying cell status into four levels: normal (below median), slightly inconsistent (between median and median+1×MAD), moderately inconsistent (between median+1×MAD and median+2×MAD), and severely inconsistent (not below median+2×MAD or judged as systematic inconsistent). At the module level, the classification thresholds are determined based on the proportion of abnormal cells within the module and the statistical distribution characteristics of the module's overall abnormal score; the cluster level follows the same principle. The boundary values ​​for each level are automatically updated with the accumulation of historical data, eliminating the need for manually setting fixed values.

[0073] To adapt to data distribution drift caused by cell aging during long-term operation, this embodiment also introduces an incremental update mechanism. As new data continues to accumulate, the standardized parameters are updated using an exponential moving average method, and some older isolation trees are periodically replaced for retraining. This ensures that the detection model always adapts to the current operating state of the battery system, improving the stability and accuracy of long-term online monitoring.

[0074] Example 2 This embodiment provides an adaptive anomaly detection system for multi-level voltage consistency detection of energy storage batteries, such as... Figure 5As shown, the method for detecting the multi-level voltage consistency of energy storage batteries described in Example 1 consists of a data acquisition module, a data preprocessing module, a multi-level feature extraction module, an adaptive multi-granularity isolated forest detection module, a hierarchical progressive judgment module, and a multi-level state classification and visualization report module. The modules communicate with each other through a data interface to realize the above-mentioned energy storage battery voltage consistency detection process.

[0075] The data acquisition module is used to collect voltage time-series data of each cell in each module within multiple battery clusters of the energy storage battery under test, under the same operating conditions. It also acquires the hierarchical topology information of the battery cluster-module-cell structure and can simultaneously receive auxiliary operating information such as current, temperature, state of charge, and module number. The acquired data is organized according to timestamps and cell numbers to form a window-level voltage dataset for subsequent analysis.

[0076] The data preprocessing module is used to perform time alignment, duplicate timestamp processing, missing value completion, outlier sampling point smoothing, and standardization on the raw voltage time series data to obtain characteristic data with a unified time base and a unified data scale, namely voltage time series segments.

[0077] The multi-level feature extraction module is used to extract features progressively at three levels—cell level, module level, and cluster level—within a sliding time window. This module includes: a cell-level statistical feature extraction unit, used to extract the mean voltage, standard deviation, range, coefficient of variation, and mean voltage change rate of each cell; a module-level coupling feature extraction unit, used to extract the relative deviation, coupling following deviation, and mutual information coupling strength between cells within a module; and a cluster-level graph feature extraction unit, used to construct a coupled weighted undirected graph and extract weighted degree centrality, betweenness centrality, local clustering coefficient, and cluster-level aggregation statistics.

[0078] The adaptive multi-granularity isolation forest detection module is used for unsupervised anomaly detection on standardized feature data. This module includes a feature-aware weighted splitting unit, a cluster-level coarse-grained isolation unit, a module-level medium-grained isolation unit, a cell-level fine-grained isolation unit, and an adaptive pollution rate adjustment unit. The feature-aware weighted splitting unit determines feature sampling weights based on the feature variance contribution rate; the cluster-level coarse-grained isolation unit filters candidate anomalous clusters at the battery cluster level; the module-level medium-grained isolation unit filters candidate anomalous modules within candidate anomalous clusters; the cell-level fine-grained isolation unit locates candidate anomalous cells within candidate anomalous modules; and the adaptive pollution rate adjustment unit statistically analyzes the anomaly score distribution of historical normal data for charging, discharging, and resting conditions, and adaptively determines the pollution rate parameters for each condition based on the median of the anomaly score distribution and the MAD statistic.

[0079] The hierarchical progressive judgment module is used to perform hierarchical progressive fine judgment on candidate abnormal cells. The first stage outputs candidate cells whose abnormal scores exceed the dynamic threshold and their corresponding time windows based on an adaptive multi-granularity isolation forest. The second stage sequentially performs cell-level deviation contribution judgment, module-level abnormal cell proportion judgment, and cluster-level abnormal module distribution judgment, and combines the coupling propagation index to distinguish between single cell independent abnormality, module-level inconsistency, cluster-level inconsistency and systemic inconsistency. The final abnormal judgment result is output through the time-series persistence constraint of T consecutive sampling windows.

[0080] The multi-level state classification and visualization report module is used to classify the state of cells, modules, and clusters based on the final anomaly score, deviation contribution, coupling propagation index, and time series persistence determination results, and generate detection reports. The thresholds for each classification level are adaptively determined based on the distribution characteristics of historical data for the corresponding level. The detection report includes the cell number, time window, anomaly score, candidate anomaly modules, state level, determination results of isolated deviations or systematic inconsistencies, as well as a cell coupling topology diagram, anomaly score time series heatmap, and deviation contribution ranking diagram.

[0081] In some implementations, the system can periodically receive new operating data and automatically update window-level samples, standardized parameters, abnormal score distribution statistics for each operating condition, and partial isolation tree models to adapt to data distribution drift caused by cell aging during long-term operation, thereby achieving dynamic monitoring of the voltage consistency status of the energy storage battery.

[0082] The system described in this embodiment is based on multi-level feature extraction and adaptive multi-granularity isolated forest detection, using hierarchical progressive judgment to achieve hierarchical localization and type differentiation of anomalies. Through this system, the voltage consistency status of energy storage batteries can be continuously monitored without relying on fault tags, and multi-level status levels and anomaly type descriptions can be output when cells deviate from normal conditions, thereby improving the stability and interpretability of battery consistency detection in energy storage power stations.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various modifications, substitutions, or variations can be made to the specific embodiments of the present invention without departing from the spirit and substance of the present invention, and all such modifications, substitutions, or variations should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive anomaly detection of multi-level voltage consistency in energy storage batteries, characterized in that, include: The voltage time series data of each cell in each module within multiple battery clusters of the energy storage battery under test are collected under the same operating conditions, and the hierarchical topology information of battery cluster-module-cell is obtained at the same time. Features are extracted from the bottom up in the order of cell level, module level and cluster level. After feature extraction is completed, an adaptive multi-granularity isolation forest model is constructed, and anomaly detection is performed layer by layer from top to bottom in the order of cluster level, module level and cell level. The construction process of the adaptive multi-granularity isolation forest model includes: introducing a feature-aware weighted splitting mechanism during the training of nodes at each level, and determining the feature selection weight based on the variance contribution rate of each feature in the training dataset; based on the weighted splitting, further adopting a coarse-fine granularity hierarchical strategy to construct an adaptive multi-granularity isolation forest, forming a three-layer detection architecture directly corresponding to the physical level of the energy storage battery; introducing a fully data-driven working condition adaptive contamination rate mechanism, whereby the contamination rate parameter of the isolation forest is adaptively determined based on the abnormal score distribution of historical normal data under each working condition.

2. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 1, characterized in that, The voltage time series data is divided into sliding windows and subjected to feature standardization to obtain voltage time series segments.

3. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 2, characterized in that, At the first cell level, statistical features are extracted for the voltage time series segments of each cell within each window. At the second module level, based on the cell-level statistical features and voltage time series segments, coupling relationship features between cells within the module are extracted, taking the module as the unit. At the third cluster level, taking battery clusters as units, the mutual information coupling strength in the coupling relationship features is aggregated by module to construct a coupling strength matrix between modules within the cluster. After normalizing this matrix, it is used as edge weights and modules as nodes to construct a weighted undirected graph. The graph topology features of each cluster are extracted and combined with cluster-level aggregated statistics.

4. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 3, characterized in that, The statistical features include voltage mean, voltage standard deviation, voltage range, voltage coefficient of variation, and voltage rate of change mean; the coupling relationship features include relative deviation, coupling follow-up deviation, and mutual information coupling strength; the graph topology features include weighted degree centrality, betweenness centrality, and local clustering coefficient; and the cluster-level aggregation statistics include intra-cluster module voltage dispersion, intra-cluster voltage range extreme values, and intra-group correlation coefficient.

5. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 1, characterized in that, The three-layer detection architecture is as follows: The first layer is a coarse-grained isolation forest at the battery cluster level, which uses the median value vector of all module features in each cluster as the aggregation representation of the cluster to quickly screen potential candidate abnormal clusters; The second layer, the module-level medium-granularity isolation forest, uses the median vector of all cell features within the module as the aggregate representation of the module to filter candidate abnormal modules for candidate abnormal clusters; the third layer, the cell-level fine-granularity isolation forest, directly uses the feature vector of a single cell to achieve accurate location of abnormal cells for candidate abnormal modules.

6. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 5, characterized in that, For the abnormal cells, a three-level progressive fine judgment is performed sequentially at the cell level, module level, and cluster level: the cell level calculates the deviation contribution to identify independent abnormal cells, the module level calculates the proportion of abnormal cells within the module to determine module-level degradation, and the cluster level calculates the distribution ratio of abnormal modules within the cluster to determine cluster-level degradation. The coupling propagation index is used to distinguish between isolated deviations and precursors of systemic degradation, and the final anomaly determination result is obtained through time-series persistence constraints.

7. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 1, characterized in that, The aforementioned adaptive pollution rate mechanism is as follows: for three operating conditions—charging, discharging, and standing—the abnormal score distribution of historical normal operating data under each condition is statistically analyzed, and the median and median absolute deviation of the abnormal score distribution under each condition are calculated. The median plus the median absolute deviation by a preset fixed multiple is used as the pollution rate threshold for that operating condition.

8. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 1, characterized in that, Also includes: Based on the anomaly scores and anomaly detection results, the status of clusters, modules, and cells is classified into different levels, forming a top-down hierarchical diagnostic conclusion and generating a test report containing visual charts.

9. The method for detecting the voltage consistency of multi-level energy storage batteries according to claim 8, characterized in that, The specific steps for classifying the status of clusters, modules, and cells are as follows: Based on the statistical characteristics of the distribution of abnormal scores in historical normal operation data, the cell status is divided into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent according to the preset four-level classification framework. The classification threshold is adaptively determined by the distribution characteristics of the historical data of the corresponding level. Based on the percentage of abnormal cells within the module and the overall abnormality score of the module, the module is classified into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent. Based on the spatial distribution of abnormal modules within the cluster and the corresponding comprehensive abnormality index of the cluster, the cluster level is divided into four levels: normal, slightly inconsistent, moderately inconsistent, and severely inconsistent.

10. An adaptive anomaly detection system for multi-level voltage consistency testing of energy storage batteries, used to implement the multi-level voltage consistency testing method for energy storage batteries according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect voltage time series data of each cell in each module within multiple battery clusters of the energy storage battery under test under the same operating conditions, and at the same time obtain the hierarchical topology information of battery cluster-module-cell. The data preprocessing module performs sliding window partitioning and feature standardization on the voltage time series data to obtain voltage time series segments; The multi-level feature extraction module includes: a cell-level statistical feature extraction unit, used to extract the mean voltage, standard deviation, range, coefficient of variation, and mean voltage change rate of each cell; The module-level coupling feature extraction unit is used to extract the relative deviation, coupling following deviation, and mutual information coupling strength between cells within the module; the cluster-level graph feature extraction unit is used to construct a coupled weighted undirected graph and extract the weighted degree centrality, betweenness centrality, local clustering coefficient, and cluster-level aggregation statistics. The adaptive multi-granularity isolation forest detection module integrates a feature-aware weighted splitting unit, a cluster-level coarse-grained isolation unit, a module-level medium-grained isolation unit, a cell-level fine-grained isolation unit, and an operating condition adaptive pollution rate adjustment unit. The pollution rate parameter is adaptively determined based on the abnormal score distribution of historical normal data under each operating condition. The hierarchical progressive judgment module is used to sequentially perform cell-level deviation contribution judgment, module-level abnormal cell proportion judgment, and cluster-level abnormal module distribution judgment. It also combines the coupling propagation index to distinguish between isolated deviations and systematic inconsistencies, and outputs the final anomaly judgment result through time-series persistence constraints.

Images