A battery fault diagnosis method and system based on a CCS module

Abstract

The application relates to the technical field of battery management, and discloses a battery fault diagnosis method and system of a CCS module, the method comprising the following steps: acquiring a multi-dimensional data sequence of battery pack voltage, current and temperature, and performing abnormality detection to obtain a potential abnormality time period and a multi-dimensional data subset; performing clustering analysis on the multi-dimensional data subset to obtain an abnormal feature clustering center, and performing dynamic correlation analysis to obtain a correlation mode vector set; calculating the similarity of each single-body data and a fault propagation path, matching the propagation path with a preset mode, and obtaining a fault propagation score; fusing the clustering center and the propagation score, positioning a fault single body through a support vector machine regression model, and determining a fault single body identifier; and segmenting and clustering data corresponding to the identifier and performing logical judgment to determine a fault type. The application can realize accurate positioning and type determination of a fault single body.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to a battery fault diagnosis method and system based on a CCS module. Background Technology

[0002] Battery management systems (BMS) are a core technology in the fields of new energy vehicles and energy storage, and their performance directly determines the safety, lifespan, and reliability of batteries. With the rapid development of the industry, accurate condition monitoring and fault diagnosis of battery systems have become a key link in the fault prediction and health management (PHM) system, which is of vital importance for ensuring system safety and improving operational efficiency.

[0003] In a representative existing technology, battery fault diagnosis is mainly achieved through a monitoring method based on fixed thresholds. Specifically, this method uses sensors to monitor the voltage data of each cell in the battery pack in real time, and pre-sets a normal operating voltage range (e.g., 3.2V to 4.2V). During system operation, once the voltage of a cell exceeds this preset upper or lower threshold, the method determines that the cell has failed and issues an alarm. Due to its simplicity and low computational cost, this method is widely used in basic overcharge or over-discharge protection scenarios.

[0004] However, the diagnostic method described above, which relies solely on a single-dimensional (voltage) static threshold, has serious limitations when faced with complex and ever-changing real-world operating conditions. This method completely ignores the strong dynamic correlation between multiple parameters of the battery, such as voltage, current, and temperature, during actual operation. For example, under high-load conditions such as rapid vehicle acceleration or fast charging, the voltage of a healthy cell may instantly drop below the normal threshold, causing the system to misjudge it as a fault. Conversely, a cell experiencing an internal micro-short circuit may still have a voltage within the normal threshold range, but its temperature may be abnormally rising—a potentially serious risk that this method completely ignores.

[0005] Therefore, existing technologies suffer from the technical problem of low fault diagnosis accuracy under complex working conditions due to the inability to comprehensively analyze the dynamic correlation of multidimensional data. Summary of the Invention

[0006] This invention provides a battery fault diagnosis method and system based on CCS modules to improve the accuracy of fault diagnosis under complex operating conditions.

[0007] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a battery fault diagnosis method based on a CCS module, comprising: Acquire a multidimensional data sequence of the battery pack, the multidimensional data sequence including time series of voltage parameters, current parameters and temperature parameters; Anomaly detection is performed on the multidimensional data sequence to determine potential abnormal time periods, thereby obtaining a multidimensional data subset of the potential abnormal time periods; Cluster analysis was used to obtain cluster centers for abnormal features from the multidimensional data subset; Dynamic correlation analysis is performed on the multidimensional data sequences corresponding to the cluster centers of the abnormal features to obtain a set of correlation pattern vectors; Calculate the similarity between data of each individual in the associated pattern vector set and the fault propagation path, and match the fault propagation path with the preset fault pattern to obtain the fault propagation score; The cluster centers of the abnormal features and the fault propagation scores of each individual are fused as input features. The physical location coordinates of the faulty individual are predicted by a preset support vector machine regression model to determine a unique faulty individual identifier. Obtain the multidimensional data sequence corresponding to the faulty unit identifier within the potential abnormal time period, and perform segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type.

[0008] Preferably, the step of performing anomaly detection on the multidimensional data sequence, determining potential anomalous time periods, and obtaining a multidimensional data subset of potential anomalous time periods includes: The multidimensional data sequence is denoised to obtain a cleaned data sequence; The cleaned data sequence is classified using a preset anomaly detection model, and data points that exceed the preset classification boundary of the model are identified as anomalies. The time period containing one or more of the aforementioned anomalies is marked as the potential anomaly time period; From the multidimensional data sequence, data corresponding to the potential abnormal time period is extracted to obtain a multidimensional data subset of the potential abnormal time period.

[0009] Preferably, obtaining anomalous feature cluster centers from the multidimensional data subset through cluster analysis includes: The multidimensional data subset is cleaned and standardized to obtain a standardized data sequence; The standardized data sequences are clustered and grouped, and the cluster centers of each data group are determined. The deviation between the cluster centers of each data group and the preset normal pattern is calculated. When the deviation value exceeds the preset normal mode threshold, the corresponding cluster center is determined as an abnormal feature cluster center.

[0010] Preferably, the step of performing dynamic correlation analysis on the multidimensional data sequences corresponding to the cluster centers of the abnormal features to obtain a set of correlation pattern vectors includes: Based on the individual cell identifier, separate the independent time series data belonging to each individual cell from the multidimensional data subset; The core statistical features of each individual cell are calculated on its independent time series data to generate the feature vector corresponding to each individual cell; The feature vectors of all individual cells are combined to form a set of associated pattern vectors that contain the states of each individual cell.

[0011] Preferably, the step of calculating the similarity between data points and fault propagation paths of individuals in the associated pattern vector set, and matching the fault propagation paths with preset fault patterns to obtain a fault propagation score, includes: Calculate the vector cosine similarity between the feature vectors of each individual data in the association pattern vector to obtain the similarity score between the individuals; When the similarity score is lower than a preset similarity threshold, the corresponding individual is determined to be an abnormal individual, and an abnormal individual set is formed. The propagation path of the abnormal entity set is traced, the correlation and impact of the abnormal entity on the data of other entities are analyzed, and the propagation path sequence of the fault is obtained. Calculate the matching degree between the propagation path sequence and the preset fault mode, and use the matching degree as the fault propagation score.

[0012] Preferably, the step of fusing the cluster centers of the abnormal features and the fault propagation scores of each individual as input features, and predicting the physical location coordinates of the faulty individual through a preset support vector machine regression model to determine a unique faulty individual identifier includes: The abnormal feature cluster centers are fused with the fault propagation scores to construct a fused feature vector for location prediction. The fused feature vector is input into a preset support vector machine regression model, and the model is used to predict and output the physical location coordinates of the faulty cell within the battery pack. The physical location coordinates are used as the identifier of the faulty cell.

[0013] Preferably, the step of obtaining the multidimensional data sequence corresponding to the faulty unit identifier within the potential abnormal time period, and performing segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type includes: Based on the faulty unit identifier, its corresponding multidimensional data sequence is obtained, and time series analysis is performed to obtain multiple initial trajectory segment sequences; The initial trajectory segment sequence is clustered to form trajectory segment groups, and the feature vector of each trajectory segment group is calculated; Calculate the cosine similarity between the feature vector of the trajectory segment group and the preset fault mode library, so as to match the trajectory segment group to the corresponding fault mode label; By combining the fault mode labels with the dynamic trend analysis of the trajectory segment sequence, a decision tree is used for classification to determine and output the fault type.

[0014] Secondly, the present invention provides a battery fault diagnosis system based on a CCS module, comprising: The data acquisition module is used to acquire a multi-dimensional data sequence of the battery pack, which includes a time series of voltage parameters, current parameters, and temperature parameters. An anomaly detection module is used to perform anomaly detection on the multidimensional data sequence, determine potential abnormal time periods, and obtain a multidimensional data subset of potential abnormal time periods; The feature extraction module is used to perform cluster analysis on the multidimensional data subset to obtain cluster centers of abnormal features; The association analysis module is used to perform dynamic association analysis on the multidimensional data sequences corresponding to the cluster centers of the abnormal features to obtain a set of association pattern vectors. The propagation score calculation module is used to calculate the similarity between the data of each individual in the associated pattern vector set and the fault propagation path, and to perform matching calculations with the preset fault patterns based on the fault propagation path to obtain the fault propagation score. The single-unit localization module is used to fuse the cluster center of the abnormal features and the fault propagation score of each single unit as input features, and predict the physical location coordinates of the faulty single unit through a preset support vector machine regression model to determine a unique faulty single unit identifier. The fault diagnosis module is used to obtain the multidimensional data sequence corresponding to the faulty individual identifier within the potential abnormal time period, and to perform segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention first performs anomaly detection on a multidimensional data sequence containing voltage, current, and temperature to pinpoint potential abnormal time periods. Then, through cluster analysis, it aggregates discrete data points within this time period into a multidimensional anomaly feature cluster center that can characterize the core state of the anomaly. This aggregation analysis strategy, from discrete data to feature centers, overcomes the limitations of existing technologies that rely solely on a single static threshold. It can comprehensively evaluate the overall distribution and dynamic coupling relationship of each parameter within a multidimensional feature space. Based on this, this invention can accurately distinguish between real faults caused by the combined effect of multiple parameters and single-parameter fluctuations under normal operating conditions, thereby achieving high-precision capture and characterization of potential abnormal events.

[0016] (2) After extracting the feature vectors of each individual cell, this invention identifies the set of abnormal individual cells through vector similarity calculation, and further employs a propagation path tracing method to analyze the transmission path and temporal relationship of the anomaly among the individual cells, ultimately quantifying the fault propagation score. This analytical paradigm, from individual identification to group association, can effectively address the non-locality of battery fault characteristics. By transforming static individual cell anomalies into dynamic propagation paths for analysis, this invention can reveal the diffusion mechanism and impact range of faults within the battery pack, solving the problem of existing technologies struggling to determine the source of anomalies and assess systemic risks, and providing crucial system-level insights for subsequent precise localization.

[0017] (3) This invention constructs a cascaded intelligent diagnostic process from localization to qualitative analysis. First, it innovatively integrates the cluster centers of abnormal features representing macroscopic abnormal states with the fault propagation scores that quantify microscopic propagation paths, and uses a support vector machine regression model to predict the physical coordinates of the faulty individual. Subsequently, for the located individual, its final fault type is determined by segmenting and logically judging its historical data trajectory. This diagnostic strategy, which combines multi-source feature fusion with in-depth historical data mining, achieves a comprehensive characterization of the fault. It can not only answer "where is broken" with high precision through machine learning models, but also delve into the root causes through subsequent qualitative analysis, reliably distinguishing "what is broken," thereby constructing a complete diagnostic closed loop and providing a comprehensive and reliable decision-making basis for the precise maintenance and health management of battery systems. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of a battery fault diagnosis method for a CCS module provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a battery fault diagnosis system for a CCS module provided in the second embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 The first embodiment of the present invention provides a battery fault diagnosis method for a CCS module, including the following steps: S11, Obtain the multi-dimensional data sequence of the battery pack, the multi-dimensional data sequence including the time series of voltage parameters, current parameters and temperature parameters; S12, perform anomaly detection on the multidimensional data sequence, determine potential abnormal time periods, and obtain a multidimensional data subset of potential abnormal time periods; S13, cluster centers of abnormal features are obtained by cluster analysis on the multidimensional data subset; S14, Perform dynamic correlation analysis on the multidimensional data sequence corresponding to the cluster centers of the abnormal features to obtain a set of correlation pattern vectors; S15, calculate the similarity between the data of each individual in the associated pattern vector set and the fault propagation path, and perform matching calculation with the preset fault mode according to the fault propagation path to obtain the fault propagation score; S16, the abnormal feature cluster center and the fault propagation score of each individual are fused as input features, and the physical location coordinates of the faulty individual are predicted by a preset support vector machine regression model to determine a unique faulty individual identifier. S17, obtain the multidimensional data sequence corresponding to the faulty unit identifier within the potential abnormal time period, and perform segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type.

[0021] In step S11, a multidimensional data sequence of the battery pack is obtained, which includes time series of voltage parameters, current parameters, and temperature parameters.

[0022] Specifically, this step is performed via a Controller Area Network (CAN) bus interface integrated with the Battery Management System (BMS). This interface is responsible for real-time acquisition and integration of data from multiple sensors deployed within the battery pack, including voltage sensors for monitoring the voltage of individual cells, thermistor (NTC) sensors for measuring the temperature at critical locations, and Hall effect sensors for measuring the total charge and discharge current of the entire battery pack circuit.

[0023] The core working principle of this step lies in converting the continuously changing physical quantities sensed by the sensor into discrete digital signals through a multi-channel, 24-bit high-precision analog-to-digital converter (ADC), and attaching a timestamp accurate to the millisecond level to each frame of synchronously acquired data, thereby forming a structured multidimensional data sequence. For example, the preset sampling frequency for data acquisition can be set to 1 Hz. It is worth noting that this preset sampling frequency is determined based on a trade-off between the battery's electrochemical reaction kinetics and the system's computational resource costs. On the one hand, the battery's fault evolution process (such as a slow voltage decrease and a gradual temperature increase) is typically a slow process on the order of minutes or even hours. A sampling frequency of 1 Hz is sufficient to capture these dynamic trends that characterize fault precursors without losing crucial information. On the other hand, excessively high sampling frequencies (e.g., above 100 Hz) generate massive amounts of redundant data, greatly increasing the burden of data storage and the computational complexity of subsequent analysis steps, while providing very limited information gain for the fault modes of interest in this invention. Therefore, 1 Hz is an optimal engineering practice value that balances system economy and real-time performance while ensuring diagnostic accuracy.

[0024] In a specific application scenario, the collected multidimensional data sequence is organized into a standard time series format, for example, stored as a comma-separated value (CSV) file. Each row contains fields such as timestamp, cell ID, voltage value (e.g., 3.0V to 4.2V), temperature value (e.g., 20°C to 60°C), and the corresponding total battery pack current value (e.g., -100A to 100A), providing a high-quality, high-dimensional raw data foundation for subsequent steps such as anomaly detection and feature analysis.

[0025] In step S12, anomaly detection is performed on the multidimensional data sequence to determine potential abnormal time periods, thereby obtaining a multidimensional data subset of potential abnormal time periods.

[0026] It should be noted that, in the first embodiment of the present invention, this processing step specifically includes steps S121 to S124: S121, Denoise the multidimensional data sequence to obtain a cleaned data sequence; S122, The cleaned data sequence is classified using a preset anomaly detection model, and data points that exceed the preset classification boundary of the model are identified as anomalies; S123, the time period containing one or more of the above anomalies is marked as the potential abnormal time period; S124, extract the data corresponding to the potential abnormal time period from the multidimensional data sequence to obtain a multidimensional data subset of the potential abnormal time period.

[0027] In step S121, the multidimensional data sequence is denoised to obtain a cleaned data sequence. Specifically, this step employs a two-stage cleaning strategy to preprocess the data obtained in the previous step. In the first stage, a sliding window outlier removal method based on the three-sigma criterion is applied. This method calculates the mean and standard deviation within a fixed-size window (e.g., containing 60 consecutive sampling points) and identifies any data points exceeding the range of "mean ± 3 times the standard deviation" as statistical outliers and removes them. The 60 sampling points (corresponding to 60 seconds at a 1Hz sampling rate) are empirical values ​​determined through experimental statistics on a large amount of historical data, balancing statistical stability and dynamic adaptability. In the second stage, the pre-processed sequence is smoothed using a median filter algorithm. This algorithm effectively filters out transient impulse noise by replacing the value of the center point with the median of the data points within a small sliding window (e.g., 5 sampling points), ultimately obtaining a cleaned data sequence that accurately reflects changes in battery state. The window size setting is based on common knowledge in the field: transient impulse noise caused by electromagnetic interference usually lasts only 1-2 sampling points. Selecting an odd number of windows that are slightly larger than the noise width (5 sampling points) can effectively filter out such spike noise and minimize the loss of detail due to over-smoothing of the real, relatively gradual battery status signal.

[0028] In step S122, a preset anomaly detection model is used to classify the cleaned data sequence, and data points that exceed the preset classification boundary of the model are identified as anomalies. Specifically, this step uses a pre-trained support vector machine (SVM) classifier for binary classification (normal / abnormal) as the anomaly detection model.

[0029] The core structure of this model lies in its use of the Radial Basis Function (RBF) kernel. The advantage of this kernel is its ability to project the original, linearly inseparable feature data (e.g., a three-dimensional feature vector composed of voltage, current, and temperature) into a higher-dimensional feature space through nonlinear mapping. Within this higher-dimensional space, an optimal classification hyperplane is constructed to achieve accurate differentiation between normal and abnormal states under complex operating conditions. The model's performance is determined by two key hyperparameters: the penalty coefficient C and the kernel function coefficient gamma. The penalty coefficient C balances maximizing the classification margin with the tolerance for sample classification errors; the kernel function coefficient gamma defines the size of the influence range of a single training sample, together determining the complexity of the final classification boundary and the generalization ability.

[0030] The training data for the model comes from the collection and expert annotation of a large amount of historical battery operation data. For example, a labeled dataset containing 10,000 records includes 9,000 normal operation data and 1,000 abnormal data covering typical scenarios such as overcharging, over-discharging, and internal micro-short circuits. Each record is a feature vector containing key parameters such as voltage, current, and temperature, and is labeled as "normal" or "abnormal".

[0031] The model is trained using supervised learning, with the goal of finding an optimal combination of hyperparameters (C, gamma). Specifically, it employs a combination of k-fold cross-validation (e.g., k=5) and grid search for hyperparameter optimization. First, candidate value ranges are defined for C and gamma (e.g., candidate values ​​for C are [0.1, 1, 10], and for gamma are [0.01, 0.1, 1]). Then, for each (C, gamma) combination on the training set, training and evaluation are performed using 5-fold cross-validation, with the F1 score used as the performance metric. Finally, the hyperparameter combination that yields the highest average F1 score is selected as the final model configuration. The model obtained using this method can accurately identify data points exceeding their classification boundaries as outliers with optimal classification performance.

[0032] In step S123, the time period containing one or more of the anomalies is marked as the potential anomaly time period. Specifically, this step uses a two-stage processing flow to transform the discrete anomalies identified in S122 into meaningful, continuous potential anomaly time periods: The first stage is the aggregation of outliers. This process first scans the discrete outlier sequence output by S122, sorted by timestamp, and applies a preset "time merging threshold" (e.g., 10 seconds) to determine the continuity of outliers. Specifically, starting from the first outlier, if the time interval between it and the next outlier is less than the threshold, they are grouped into the same temporary outlier event segment. The system continues to aggregate until the time interval between two consecutive outliers exceeds the threshold, at which point the current event segment ends, and a new event segment is constructed starting from the next outlier. The threshold is set based on experimental statistics of real fault events in historical data, and its purpose is to bridge the small data gaps caused by instantaneous signal loss and ensure the integrity of individual fault events.

[0033] The second stage is the screening of event segments. This process calculates the total duration (i.e., the end timestamp minus the start timestamp) of each temporary anomalous event segment generated in the first stage. This duration is then compared to a preset "minimum duration threshold" (e.g., 5 minutes). Only when the total duration of an event segment is greater than or equal to this threshold is the segment deemed analytically valuable and formally marked as a "potentially anomalous time period," and its start and end time ranges are passed to subsequent steps. Conversely, if the duration is less than the threshold, the event segment is considered a benign transient fluctuation and filtered out. This minimum duration threshold is set based on well-known common knowledge in the field of battery electrochemistry and thermodynamics, aiming to concentrate computational resources on analyzing persistent, potentially anomalous events that are more likely to evolve into real failures.

[0034] In step S124, data corresponding to the potential abnormal time periods are extracted from the multidimensional data sequence to obtain a multidimensional data subset of the potential abnormal time periods. Specifically, this step performs a timestamp-based index slicing operation. Using the start and end timestamps of each potential abnormal time period determined in the previous sub-step as query boundaries, it precisely extracts all data points within that time range from the original multidimensional data sequence obtained in step S11. For example, a potential abnormal time period lasting 10 minutes with a sampling frequency of 1 Hz will yield a multidimensional data subset containing 600 records. This step ensures that subsequent analyses are performed on high-quality data that strictly corresponds to the abnormal event, guaranteeing the relevance and accuracy of the analysis.

[0035] In step S13, cluster centers of abnormal features are obtained from the multidimensional data subset through cluster analysis.

[0036] It should be noted that, in the first embodiment of the present invention, this processing step specifically includes steps S131 to S134: S131, perform data cleaning and standardization on the multidimensional data subset to obtain a standardized data sequence; S132, cluster the standardized data sequence and determine the cluster center of each data group; S133, calculate the deviation between the cluster centers of each data group and the preset normal pattern; S134, when the deviation value exceeds the preset normal mode threshold, the corresponding cluster center is determined as the abnormal feature cluster center.

[0037] In step S131, the multidimensional data subset is cleaned and standardized to obtain a standardized data sequence. Specifically, this step first uses a median filter algorithm to clean the multidimensional data subset obtained in the previous step to further eliminate residual impulse noise. First, a fixed-size sliding window is set (e.g., containing 5 consecutive sampling points); second, the window is slid point by point along the time axis, traversing the entire time series; third, at each position of the sliding window, all data points (e.g., 5 consecutive voltage values) within the window are extracted, numerically sorted, and the median is selected; finally, the calculated median is used to replace the original value at the center point of the window. Since instantaneous impulse noise usually manifests as isolated extreme points, it will be placed at both ends of the sequence rather than in the middle after sorting. Therefore, this method can effectively remove abnormal peak values ​​and achieve smoothing of the sequence. It is worth noting that the setting of the window size (e.g., containing 5 consecutive sampling points) is based on the need to achieve a balance between noise suppression capability and signal fidelity. Based on common knowledge in the field, transient impulse noise in battery data typically lasts only 1 to 2 sampling points. Choosing a window size of 5 serves several purposes: firstly, it is an odd number, ensuring a unique center point within the window for replacement; secondly, the size is large enough to effectively cover and eliminate noise spikes of this width, yet small enough to avoid excessive smoothing or delaying the relatively flat signal curve that characterizes the actual fault evolution, thus preserving the critical details of the fault occurrence.

[0038] Subsequently, to eliminate the dimensional differences caused by different physical units among the data dimensions, this step uses the Z-score normalization method to process the cleaned data sequence. The core principle of this method is to independently calculate the mean for each data dimension (voltage, current, temperature). with standard deviation And apply the transformation formula The original numerical value x is converted into a standardized z-score value. After this processing, the values ​​of all dimensions are mapped to the same scale with a mean of 0 and a standard deviation of 1, thus avoiding the problem that dimensions with larger numerical ranges disproportionately dominate distance calculations in subsequent clustering calculations.

[0039] In step S132, the standardized data sequence is clustered and grouped, and the cluster centers of each data group are determined. Specifically, this step uses the K-means clustering algorithm to perform unsupervised grouping of the standardized data sequence obtained in S131. First, the optimal number of clusters K (e.g., K=3) is determined based on the Elbow Method and the actual physical meaning of the battery system operation. Then, K cluster centers are initialized. Typically, the K-means algorithm is used to select initial centroids that are far apart to avoid convergence to a local optimum. Next, the Euclidean distance between each data point and each cluster center is calculated, and it is assigned to the nearest cluster. Then, the mean of all data points in each cluster is recalculated as the new cluster center. This process is iterated repeatedly until the change in the cluster centers is less than the preset convergence threshold or the maximum number of iterations (e.g., 100 times) is reached. At this point, the algorithm converges, and the final K cluster groups and the corresponding cluster center coordinates are output. To improve the stability of the results, the algorithm can be run multiple times and the result with the smallest sum of squares (SSE) within a cluster can be selected as the final clustering result. At the same time, the silhouette coefficient of each cluster can be recorded to evaluate the clustering quality and ensure that the divided state categories (such as normal, high temperature abnormality, voltage fluctuation, etc.) have good separability and internal consistency in the feature space.

[0040] In step S133, the deviation between the cluster centers of each data group and the preset normal mode is calculated. Specifically, this step uses Euclidean distance to quantify the degree of deviation between each cluster center obtained in S132 and the "preset normal mode". It is worth noting that the "preset normal mode" is a benchmark feature vector representing the healthy operating state of the battery. Its setting is based on first constructing a benchmark historical database to define "normal operating conditions". This database is preferably composed of the following two types of data: first, bench test data from the laboratory, that is, data collected by performing standardized test conditions (e.g., WLTC cycle conditions) on the battery pack in a controlled environment; second, early operating data from actual vehicles or energy storage power stations, that is, through data cleaning, only those data that are confirmed to be fault-free and have stable operating conditions are retained. Subsequently, by performing statistical analysis on all data points in the benchmark database, the centroid of each data point in the standardized feature space (i.e., the mean vector of all normal data points) is calculated and solidified into the preset normal mode.

[0041] In step S134, when the deviation value exceeds a preset normal mode threshold, the corresponding cluster center is determined as an abnormal feature cluster center. Specifically, this step compares each deviation value calculated in S133 with the preset normal mode threshold using direct comparison logic. The threshold is set based on receiver operating characteristic (ROC) curve analysis on a validation dataset containing labeled normal and abnormal samples. By testing a series of candidate thresholds and evaluating the true positive rate (the ability to detect real anomalies) and false positive rate (the probability of misreporting normal as anomaly) at each threshold, the threshold point that maximizes the Youden's Index (e.g., 0.5) is finally selected. This threshold represents the optimal balance between maximizing detection sensitivity and minimizing the false positive rate. Any cluster center with a deviation value exceeding this threshold is ultimately determined as an "abnormal feature cluster center" representing the core characteristics of the current fault.

[0042] In step S14, dynamic correlation analysis is performed on the multidimensional data sequence corresponding to the abnormal feature cluster center to obtain a set of correlation pattern vectors.

[0043] It should be noted that, in the first embodiment of the present invention, this processing step specifically includes steps S141 to S143: S141, Based on the individual cell identifier, separate the independent time series data belonging to each individual cell from the multidimensional data subset; S142, calculate the core statistical features of each individual cell on its independent time series data to generate the feature vector corresponding to each individual cell; S143, combine the feature vectors of all individual cells to form a set of associated pattern vectors containing the states of each individual cell.

[0044] In step S141, based on the cell identifier, independent time-series data belonging to each individual cell are separated from the multidimensional data subset. Specifically, this step performs a data grouping operation on the "multidimensional data subset" obtained in S13. This operation uses the unique identifier (Cell ID) of the cell as the key to logically divide the mixed data stream into multiple parallel time-series data, each containing only the voltage, temperature, and associated total current of a specific cell (e.g., cell 1, cell 2, etc.). This step is a mandatory prerequisite for achieving refined cell-level diagnostics; it deconstructs the macroscopic battery pack data into microscopic cell data, laying the foundation for subsequent independent feature engineering calculations for each cell.

[0045] In step S142, the core statistical features of each individual battery cell are calculated on its independent time-series data to generate a feature vector corresponding to each individual battery cell. Specifically, this step performs feature engineering on the independent time-series data of each individual cell separated in S141, aiming to transform the dynamic time-series data into a feature vector that can statically describe its core behavior during abnormal periods. This feature vector is a high-dimensional vector containing multi-dimensional information such as voltage, current, and temperature, used to accurately describe the comprehensive operating state of the individual cell during abnormal periods. Preferably, this feature vector is constructed by calculating the following three types of core statistical features: Steady-state statistical characteristics include the average, standard deviation, maximum and minimum values ​​of each physical quantity (voltage, temperature, etc.), which are used to characterize the average operating point, fluctuation stability and extreme stress that the unit can withstand during abnormal periods.

[0046] Dynamic trend characteristics: For example, by fitting the linear regression slope of temperature change over time using the least squares method, the slope value can robustly quantify the continuous upward or downward trend of temperature, which is a key dynamic indicator for judging the risk of thermal runaway.

[0047] Data distribution morphology characteristics: For example, calculating the skewness and kurtosis of voltage sequences to describe the symmetry and dispersion of their data distribution can reveal potential nonlinear degradation patterns.

[0048] In step S143, the feature vectors of all individual battery cells are combined to form a set of association pattern vectors containing the states of each individual cell. Specifically, this step performs a data aggregation operation. It stacks the feature vectors generated independently for each individual cell in S142 row-wise into a unified data structure, such as an N×M matrix, while maintaining the order of the individual cells. Here, N is the total number of cells in the battery pack, and M is the dimension of each feature vector constructed in S142 (e.g., a 12-dimensional vector containing 12 statistical features). This final matrix is ​​formally defined as the "set of association pattern vectors." Its well-structured format and complete information make it the standard input format for matrix operations such as vector cosine similarity calculation in subsequent steps, greatly improving analysis efficiency.

[0049] In step S15, the similarity between the data of each individual in the associated pattern vector set and the fault propagation path are calculated, and the fault propagation score is obtained by matching the fault propagation path with the preset fault mode.

[0050] It should be noted that, in the first embodiment of the present invention, this processing step specifically includes steps S151 to S154: S151, Calculate the vector cosine similarity between the feature vectors of each individual entity in the set of associated pattern vectors to obtain the similarity score between the individual entities; S152, when the similarity score is lower than a preset similarity threshold, the corresponding individual is determined to be an abnormal individual, and an abnormal individual set is formed; S153, perform propagation path tracing on the abnormal unit set, analyze the correlation and impact of abnormal units on other unit data, and obtain the propagation path sequence of the fault. S154, calculate the matching degree between the propagation path sequence and the preset fault mode, and use the matching degree as the fault propagation score.

[0051] In step S151, the vector cosine similarity between the individual feature vectors in the association pattern vector set is calculated to obtain the similarity score between the individuals. Specifically, this step uses the Vector Cosine Similarity algorithm to process the association pattern vector set (N×M matrix) obtained in S14, which contains N individual feature vectors. The core principle of this algorithm is to measure the similarity of their operating states by calculating the cosine value of the angle between any two individual feature vectors in the M-dimensional feature space. The calculation is achieved through a normalized vector dot product, with a value range between [-1, 1]. The closer the value is to 1, the more similar the operating state patterns of the two individuals are. This method is not sensitive to the absolute value of the vectors and is particularly suitable for comparing the differences between high-dimensional feature vectors after standardization. Finally, this step outputs an N×N similarity matrix, where each element represents the similarity score between the corresponding two individuals (for example, the similarity score between individual 1 and individual 2 is 0.75).

[0052] In step S152, when the similarity score is lower than a preset similarity threshold, the corresponding individual is identified as an abnormal individual, and an abnormal individual set is formed. Specifically, this step first analyzes the similarity matrix generated in S151, for example, by calculating the average similarity score of each individual with all other individuals. Then, the average similarity score of each individual is compared with a "preset similarity threshold." The threshold is set based on statistical analysis of historical healthy battery datasets. By calculating the distribution of similarity scores among individuals in a large number of healthy battery packs, a lower percentile (e.g., the 5th percentile) is selected as the threshold (e.g., 0.8). Any individual with an average similarity score lower than this threshold is determined to be an abnormal individual due to its significantly deviated operating state from the group's normal state and is included in an "abnormal individual set" for subsequent propagation path analysis.

[0053] In step S153, the propagation path of the abnormal unit set is traced, and the correlation influence of the abnormal unit on the data of other units is analyzed to obtain the propagation path sequence of the fault. Specifically, this step preferably uses cross-correlation analysis to accurately trace the propagation path of the fault. For each source unit in the abnormal unit set determined in S152, the system first determines the key feature dimension for analysis. This dimension is preferably the feature that exhibits the largest statistical deviation in the feature vector obtained in step S142 (for example, if the standard deviation of temperature deviates the most from the normal value, then temperature is selected). Subsequently, the system extracts the key feature time series of the source unit and calculates the cross-correlation function between this series and the corresponding feature time series of all adjacent target units in the preset physical topology of the battery pack. This function can reveal the degree of correlation between the two time series at different time delays. If the cross-correlation peak that appears at a certain positive time delay (for example, 5 minutes) has a correlation score that exceeds a preset correlation threshold (for example, 0.7), then it is determined that there is an effective propagation path from the source unit to the target unit. It is worth noting that the correlation threshold is set so that it is higher than the spurious correlation level caused by random noise. This value is typically determined by statistically analyzing the cross-correlation coefficients between different units under a large number of normal operating conditions and selecting the upper bound of its 95% confidence interval to ensure that the identified propagation paths are statistically significant. Finally, by connecting all detected valid propagation paths, the system constructs one or more fault propagation path sequences that include the unit number, propagation direction, and time delay of the anomalous impact.

[0054] In step S154, the matching degree between the propagation path sequence and the preset fault mode is calculated, and the matching degree is used as the fault propagation score. Specifically, this step employs a sequence matching algorithm based on Dynamic Time Warping (DTW). The "preset fault mode" is stored in a fault propagation mode library, where each fault (such as "thermal runaway" or "internal short circuit") is defined by its typical, verified propagation path sequence as a template. It is worth noting that this fault propagation mode library is constructed based on a large number of fault injection experiments and the analysis of real historical failure data. The construction method and data structure of this fault propagation mode library are as follows: First, through a large number of fault injection experiments (e.g., overcharging or nail penetration tests on the battery pack on a test bench) and the analysis of real historical failure data, a complete multidimensional dataset of various known fault types is obtained. Second, for each set of fault data, the propagation path tracing method described in step S153 is applied to extract its empirical propagation path sequence. Next, for all empirical propagation path sequences of the same fault type (e.g., "thermal runaway"), a sequence averaging algorithm (e.g., Barycentric Averaging based on DTW) is used to calculate an "average path" that represents the typical propagation characteristics of this type of fault, serving as the propagation mode template for that fault. Finally, each fault mode template in the fault propagation mode library is represented as a standardized multidimensional numerical sequence. For example, the template sequence for the "thermal runaway" mode can be defined as a three-node multidimensional numerical sequence: node 1 = [120 seconds, 0.90], node 2 = [210 seconds, 0.88], node 3 = [300 seconds, 0.85], where each dimension represents the "cumulative time delay to reach the node" and the "standardized correlation score," respectively.

[0055] During matching, the propagation path sequence obtained in S153 (e.g., [(Source: #5, Target: #6, Delay: 240s, Correlation: 0.91)]) is first transformed into a multidimensional numerical sequence consistent with the template format. The WDTW algorithm finds the optimal alignment path between the observed sequence and each template sequence in the fault propagation pattern library by non-linearly "bending" the time axis, and calculates its weighted cumulative distance (d). The weight (w) in this "weighting" step is dynamically set based on the correlation score of each node in the observed path. Specifically, the higher the correlation score of a node, the larger its corresponding weight w, which allows the matching process to focus more on propagation links with high confidence, thereby improving matching accuracy. This cumulative distance is transformed into the final matching degree through a normalized exponential function, calculated as follows: Matching degree = Where d is the cumulative distance calculated by WDTW. A preset scaling parameter, determined through tuning on a validation set, is used to adjust the distance sensitivity, ensuring that the final matching score can effectively distinguish different fault propagation modes. Finally, the system matches the observed sequences with all templates in the fault propagation mode library and determines the highest matching score (e.g., 0.9 for the "thermal runaway" mode) as the final fault propagation score.

[0056] In step S16, the abnormal feature cluster center and the fault propagation score of each individual are fused as input features. The physical location coordinates of the faulty individual are predicted by a preset support vector machine regression model to determine a unique faulty individual identifier.

[0057] It should be noted that, in the first embodiment of the present invention, this processing step specifically includes steps S161 to S162: S161, the abnormal feature cluster centers are fused with the fault propagation scores to construct a fused feature vector for location prediction; S162, the fused feature vector is input into a preset support vector machine regression model, and the model is used to predict and output the physical location coordinates of the faulty cell within the battery pack, and the physical location coordinates are used as the identifier of the faulty cell.

[0058] In step S161, the abnormal feature cluster centers and the fault propagation score are fused to construct a fused feature vector for location prediction. Specifically, this step performs a feature fusion operation. This operation concatenates two features of different dimensions but complementary physical meanings: one is the abnormal feature cluster centers obtained in S13, representing the "macroscopic abnormal state" (e.g., a 10-dimensional vector); the other is the fault propagation score obtained in S15, quantifying the "microscopic propagation pattern" (a scalar). By concatenating these two vectors, the system constructs a "fused feature vector" (e.g., a 10+1=11-dimensional vector) with higher information density and a more comprehensive characterization of the fault, providing a more discriminative input for subsequent machine learning models.

[0059] In step S162, the fused feature vector is input into a preset support vector machine regression model, and the model is used to predict and output the physical location coordinates of the faulty cell within the battery pack. These physical location coordinates are used as the identifier of the faulty cell. Specifically, this step uses a pre-trained support vector regression (SVR) model.

[0060] The core of this model lies in its use of the Radial Basis Function (RBF Kernel), which enables it to learn and fit the complex nonlinear mapping relationship between the fused feature vectors in S161 and the physical location coordinates of the faulty individual. Its regression principle involves finding a regression hyperplane in a high-dimensional feature space such that the distance deviation of all training sample points from this hyperplane falls within a preset tolerance. Within this range, the spacing band of the hyperplane is maximized.

[0061] The training data for the model comes from a large database of historical failure cases or a dataset generated through failure injection experiments. In this dataset, the input of each sample is the "fusion feature vector" constructed in S161, and its corresponding target output is the known and real physical location coordinates of the faulty individual in the failure case (e.g., a two-dimensional vector [row, column] representing its row and column position in the battery module).

[0062] The model is trained using supervised learning. A combination of k-fold cross-validation and grid search is used to systematically optimize the model's three key hyperparameters: the penalty coefficient C, the kernel coefficient gamma, and the tolerance epsilon. The optimization objective is to minimize the root mean square error (RMSE) between the predicted and true coordinates. Finally, the optimal combination of hyperparameters is selected to construct and solidify the pre-defined SVR model.

[0063] During the diagnostic phase, the model receives a new fused feature vector and directly outputs a continuous physical location coordinate (e.g., [2.1, 5.3]). Subsequently, the system performs a nearest neighbor lookup based on Euclidean distance to map the coordinates to the individual cells. Specifically, the system iterates through a pre-defined "cell location layout map," which stores the unique engineering number of each cell within the battery pack and its corresponding discrete physical coordinates (e.g., cell #28 corresponds to coordinates [2, 5]). The system calculates the Euclidean distance between the predicted coordinates output by the model and the discrete coordinates of each cell in the layout map, and finds the cell with the smallest distance. The unique engineering number corresponding to this cell with the smallest distance (e.g., "cell #28") is ultimately identified as the faulty cell in this diagnostic process.

[0064] In step S17, the multidimensional data sequence corresponding to the faulty unit identifier within the potential abnormal time period is obtained, and the multidimensional data sequence is segmented, clustered, and logically judged to determine the final fault type.

[0065] It should be noted that, in the first embodiment of the present invention, this processing step specifically includes steps S171 to S174: S171, Based on the faulty unit identifier, obtain its corresponding multidimensional data sequence, and perform time series segmentation to obtain multiple initial trajectory segment sequences; S172, cluster the initial trajectory segment sequence to form trajectory segment groups, and calculate the feature vector of each trajectory segment group; S173, calculate the cosine similarity between the feature vector of the trajectory segment group and a preset fault mode library, so as to match the trajectory segment group to the corresponding fault mode label; S174, Combining the fault mode labels with the dynamic change trend analysis of the trajectory segment sequence, classification is performed using a decision tree to determine and output the fault type.

[0066] In step S171, based on the faulty individual identifier, its corresponding multidimensional data sequence is obtained and time-series segmented to obtain multiple initial trajectory segment sequences. Specifically, this step first retrieves the complete multidimensional data sequence of the faulty individual within the potential abnormal time period from the historical database based on the faulty individual identifier determined in S16. Subsequently, a fixed-size sliding window segmentation method is used to segment this continuous data trajectory into multiple standardized initial trajectory segment sequences. The window size (e.g., 5 minutes) is set based on the need to strike a balance between capturing a complete and meaningful operating pattern and finely distinguishing different fault evolution stages; this value is typically determined based on the characteristic time constant of the battery electrochemical process.

[0067] In step S172, the initial trajectory segment sequence is clustered to form trajectory segment groups, and the feature vector of each trajectory segment group is calculated. Specifically, this step first calculates a feature vector for each trajectory segment obtained in S171 (e.g., containing statistics such as the mean and standard deviation of voltage, current, and temperature within the segment). Then, the K-means clustering algorithm is used to perform unsupervised grouping of the feature vectors of all trajectory segments. The detailed implementation process of this algorithm is as follows: First, the optimal number of clusters K is determined based on the Elbow Method combined with domain knowledge of the typical behavior patterns of faulty cells (e.g., K=3 may correspond to three different behavior patterns: "stable overvoltage", "intermittent fluctuation", and "normal"). Subsequently, the K-means++ algorithm is used to initialize K cluster centers to select initial centroids that are far apart to improve the globality of convergence. Next, the algorithm enters an iterative process: In the allocation step, the Euclidean distance between the feature vector of each trajectory segment and the current K cluster centers is calculated, and the segment is assigned to the cluster with the closest distance. In the update step, the mean vector of the feature vectors of all trajectory segments within each cluster is recalculated, and this mean vector is used as the new cluster center for that cluster. This allocation-update process iterates repeatedly until the change in the position of the cluster center is less than a preset convergence threshold or the maximum number of iterations is reached. This algorithm aggregates trajectory segments with similar behavioral patterns (e.g., all segments exhibiting "high voltage, stable current") into the same trajectory segment group. The centroid of each group is the feature vector of that group, representing a unique operating pattern exhibited by the faulty individual.

[0068] In step S173, the cosine similarity between the feature vectors of the trajectory segment groups and a preset fault mode library is calculated to match the trajectory segment groups to their corresponding fault mode labels. Specifically, this step calculates the cosine similarity between the feature vectors of each trajectory segment group obtained in S172 and a template vector in a "preset fault mode library". It is worth noting that the construction method and data structure of this "preset fault mode library" are as follows: Construction Method: First, through extensive accelerated aging experiments and analysis of real historical failure data with confirmed root causes, a large number of multidimensional data time segments of various known fault types (such as "overvoltage fault," "internal short circuit," etc.) at different stages are obtained. Second, for each time segment, the feature vector is calculated using the method described in step S172. Third, for all feature vectors of the same fault type (e.g., "overvoltage fault"), a central feature vector that can represent the typical state of this type of fault is calculated using methods such as Gaussian Mixture Model (GMM) or mean calculation, serving as a template for the fault.

[0069] Data Structure: This library is a collection of key-value pairs. The key is a text label for the fault type (e.g., "overvoltage fault"); the value is a template feature vector representing the core state of the fault mode (e.g., a 12-dimensional vector containing statistics such as mean voltage and standard deviation of temperature).

[0070] During matching, the similarity score (e.g., 0.9) between the trajectory segment grouping feature vector output by S172 and each template vector in the preset fault mode library is calculated and compared with a preset "similarity threshold" (e.g., 0.85).

[0071] It is worth noting that the similarity threshold is set based on receiver operating characteristic (ROC) curve analysis on a validation dataset containing samples of various labeled fault types. By testing a series of candidate thresholds (e.g., from 0.7 to 0.95) on this validation set and evaluating the true positive rate (the ability to match correctly) and false positive rate (the probability of mismatches) at each threshold, the threshold point that optimizes the classification performance metric (such as the Youden index) is ultimately selected as the fixed preset value. This threshold setting ensures that a behavioral pattern group is only assigned a preliminary fault pattern label when it shows a sufficiently high similarity to a preset fault pattern, thus guaranteeing the confidence level of the initial qualitative assessment. This step assigns a most probable, preliminary fault pattern label to each behavioral pattern group.

[0072] In step S174, the fault mode label is combined with the dynamic trend analysis of the trajectory segment sequence, and classification is performed using a decision tree to determine and output the fault type. Specifically, this step achieves a final accurate qualitative assessment of the fault type through a two-stage process that combines dynamic feature extraction and logical decision-making.

[0073] The first stage is dynamic feature extraction. This stage applies an autoregressive (AR) model to each trajectory segment in S173 that has been initially labeled, performing dynamic trend analysis. In this step, by fitting an AR model (e.g., a 5th-order AR(5) model) to the time series of a key dimension (such as temperature) of the trajectory segment, a set of model coefficients that quantify its internal dynamic characteristics can be obtained. These coefficients themselves constitute a "dynamic trend feature vector," reflecting whether the sequence exhibits continuous linear growth or periodic oscillations. The order p of the model (e.g., p=5) is determined by applying the Akaike Information Criterion on the validation set, aiming to achieve the best balance between model fit and complexity.

[0074] The second stage is logical decision-making. This stage uses a pre-trained decision tree classifier as the final classification model. The model's structure consists of a series of "if-then" judgments with clear physical meaning, generated by learning from historical fault data. In the model's training data, the input of each sample is a fused feature vector, which is formed by concatenating the "preliminary fault mode label" obtained from S173 (after one-hot encoding) with the "dynamic trend feature vector" output from the previous stage's AR model; its target output is the final fault type label corresponding to the sample, verified by experts. The model is trained using supervised learning, preferably employing the C4.5 or CART (Classification and Regression Tree) algorithm. This algorithm constructs the optimal tree-like decision structure by recursively selecting features that provide the maximum information gain to partition nodes, and prevents overfitting through pruning operations. During diagnosis, the trained decision tree receives the fused feature vector and, through a series of logical judgments (e.g., "IF initial label = 'overpressure fault' AND temperature AR coefficient indicates continuous increase, THEN final type = 'overpressure triggers thermal runaway risk'"), finally outputs a high-confidence, specific fault type classification label.

[0075] Reference Figure 2 The second embodiment of the present invention provides a battery fault diagnosis system based on a CCS module, comprising: The data acquisition module is used to acquire a multi-dimensional data sequence of the battery pack, which includes a time series of voltage parameters, current parameters, and temperature parameters. An anomaly detection module is used to perform anomaly detection on the multidimensional data sequence, determine potential abnormal time periods, and obtain a multidimensional data subset of potential abnormal time periods; The feature extraction module is used to perform cluster analysis on the multidimensional data subset to obtain cluster centers of abnormal features; The association analysis module is used to perform dynamic association analysis on the multidimensional data sequences corresponding to the cluster centers of the abnormal features to obtain a set of association pattern vectors. The propagation score calculation module is used to calculate the similarity between the data of each individual in the associated pattern vector set and the fault propagation path, and to perform matching calculations with the preset fault patterns based on the fault propagation path to obtain the fault propagation score. The single-unit localization module is used to fuse the cluster center of the abnormal features and the fault propagation score of each single unit as input features, and predict the physical location coordinates of the faulty single unit through a preset support vector machine regression model to determine a unique faulty single unit identifier. The fault diagnosis module is used to obtain the multidimensional data sequence corresponding to the faulty individual identifier within the potential abnormal time period, and to perform segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type.

[0076] It should be noted that the battery fault diagnosis system based on CCS module provided in this embodiment of the invention is used to execute all the process steps of the battery fault diagnosis method based on CCS module in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0077] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a correlation analysis program. When the processor executes the computer program, it implements the steps in the various battery fault diagnosis method embodiments based on CCS modules described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as a single-unit positioning module.

[0078] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0079] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0080] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0081] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0082] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0083] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0084] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A battery fault diagnosis method based on a CCS module, characterized in that, include: Acquire a multidimensional data sequence of the battery pack, the multidimensional data sequence including time series of voltage parameters, current parameters and temperature parameters; Anomaly detection is performed on the multidimensional data sequence to determine potential abnormal time periods, thereby obtaining a multidimensional data subset of the potential abnormal time periods; Cluster analysis was used to obtain cluster centers for abnormal features from the multidimensional data subset; Dynamic correlation analysis is performed on the multidimensional data sequences corresponding to the cluster centers of the abnormal features to obtain a set of correlation pattern vectors; The similarity between data of each individual entity in the association pattern vector set and the fault propagation path are calculated. The fault propagation path is matched with a preset fault pattern to obtain a fault propagation score. This includes calculating the vector cosine similarity between the feature vectors of each individual entity in the association pattern vector to obtain a similarity score between entities. When the similarity score is lower than a preset similarity threshold, the corresponding entity is determined to be an abnormal entity, and an abnormal entity set is formed. The propagation path of the abnormal entity set is traced, the correlation and impact of the abnormal entity on the data of other entities are analyzed, and the propagation path sequence of the fault is obtained. Calculate the matching degree between the propagation path sequence and the preset fault mode, and use the matching degree as the fault propagation score; The cluster centers of the abnormal features and the fault propagation scores of each individual are fused as input features. The physical location coordinates of the faulty individual are predicted by a preset support vector machine regression model to determine a unique faulty individual identifier. Obtain the multidimensional data sequence corresponding to the faulty unit identifier within the potential abnormal time period, and perform segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type.

2. The battery fault diagnosis method based on CCS module as described in claim 1, characterized in that, The step of performing anomaly detection on the multidimensional data sequence, determining potential anomalous time periods, and obtaining a multidimensional data subset of potential anomalous time periods includes: The multidimensional data sequence is denoised to obtain a cleaned data sequence; The cleaned data sequence is classified using a preset anomaly detection model, and data points that exceed the preset classification boundary of the model are identified as anomalies. The time period containing one or more of the aforementioned anomalies is marked as the potential anomaly time period; From the multidimensional data sequence, data corresponding to the potential abnormal time period is extracted to obtain a multidimensional data subset of the potential abnormal time period.

3. The battery fault diagnosis method based on CCS module as described in claim 1, characterized in that, The process of obtaining anomalous feature cluster centers from the multidimensional data subset through cluster analysis includes: The multidimensional data subset is cleaned and standardized to obtain a standardized data sequence; The standardized data sequences are clustered and grouped, and the cluster centers of each data group are determined. The deviation between the cluster centers of each data group and the preset normal pattern is calculated. When the deviation value exceeds the preset normal mode threshold, the corresponding cluster center is determined as an abnormal feature cluster center.

4. The battery fault diagnosis method based on CCS module as described in claim 1, characterized in that, The dynamic correlation analysis of the multidimensional data sequences corresponding to the cluster centers of the abnormal features yields a set of correlation pattern vectors, including: Based on the individual cell identifier, separate the independent time series data belonging to each individual cell from the multidimensional data subset; The core statistical features of each individual cell are calculated on its independent time series data to generate the feature vector corresponding to each individual cell; The feature vectors of all individual cells are combined to form a set of associated pattern vectors that contain the states of each individual cell.

5. The battery fault diagnosis method based on CCS module as described in claim 1, characterized in that, The step of fusing the cluster centers of the abnormal features and the fault propagation scores of each individual as input features, and predicting the physical location coordinates of the faulty individual through a preset support vector machine regression model to determine a unique faulty individual identifier includes: The abnormal feature cluster centers are fused with the fault propagation scores to construct a fused feature vector for location prediction. The fused feature vector is input into a preset support vector machine regression model, and the model is used to predict and output the physical location coordinates of the faulty cell within the battery pack. The physical location coordinates are used as the identifier of the faulty cell.

6. The battery fault diagnosis method based on CCS module as described in claim 1, characterized in that, The process of obtaining the multidimensional data sequence corresponding to the faulty unit identifier within the potential abnormal time period, and performing segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type includes: Based on the faulty unit identifier, its corresponding multidimensional data sequence is obtained, and time series analysis is performed to obtain multiple initial trajectory segment sequences; The initial trajectory segment sequence is clustered to form trajectory segment groups, and the feature vector of each trajectory segment group is calculated; Calculate the cosine similarity between the feature vector of the trajectory segment group and the preset fault mode library, so as to match the trajectory segment group to the corresponding fault mode label; By combining the fault mode labels with the dynamic trend analysis of the trajectory segment sequence, a decision tree is used for classification to determine and output the fault type.

7. A battery fault diagnosis system based on a CCS module, characterized in that, include: The data acquisition module is used to acquire a multi-dimensional data sequence of the battery pack, which includes a time series of voltage parameters, current parameters, and temperature parameters. An anomaly detection module is used to perform anomaly detection on the multidimensional data sequence, determine potential abnormal time periods, and obtain a multidimensional data subset of potential abnormal time periods; The feature extraction module is used to perform cluster analysis on the multidimensional data subset to obtain cluster centers of abnormal features; The association analysis module is used to perform dynamic association analysis on the multidimensional data sequences corresponding to the cluster centers of the abnormal features to obtain a set of association pattern vectors. The propagation score calculation module is used to calculate the similarity between the data of each individual entity in the association pattern vector set and the fault propagation path, and to perform matching calculations with the fault propagation path and preset fault modes to obtain the fault propagation score. Specifically, it includes calculating the vector cosine similarity between the feature vectors of each individual entity data in the association pattern vector to obtain the similarity score between entities; when the similarity score is lower than the preset similarity threshold, the corresponding entity is determined to be an abnormal entity, and an abnormal entity set is formed. The propagation path of the abnormal entity set is traced, the correlation and influence of the abnormal entity on the data of other entities are analyzed, and the propagation path sequence of the fault is obtained; the matching degree between the propagation path sequence and the preset fault mode is calculated, and the matching degree is used as the fault propagation score. The single-unit localization module is used to fuse the cluster center of the abnormal features and the fault propagation score of each single unit as input features, and predict the physical location coordinates of the faulty single unit through a preset support vector machine regression model to determine a unique faulty single unit identifier. The fault diagnosis module is used to obtain the multidimensional data sequence corresponding to the faulty individual identifier within the potential abnormal time period, and to perform segmented clustering and logical judgment on the multidimensional data sequence to determine the final fault type.

Images