Power supply power distribution fault prediction method and system based on BMS data analysis driving

Abstract

The invention relates to the field of power supply fault prediction, in particular to a power supply power distribution fault prediction method and system based on BMS data analysis driving. The method comprises the following steps: calling a BMS (Battery Management System) to collect a historical monitoring log of a power supply, carrying out periodic power disturbance change analysis and normalized power perturbation evolution, and constructing a normalized power perturbation state diagram of a plurality of battery cells; state evolution mining is carried out on the battery cells one by one based on a power supply historical monitoring log, multi-dimensional personalized state perception is carried out, and a personalized state portrait of each battery cell is constructed; calculating request power deviation of the battery cells, performing power distribution coordination analysis among the battery cells, performing power distribution game logic mining based on the personalized state portrait, and identifying a power distribution game rule of the battery cells; and performing power distribution prediction and power distribution track deviation comparison according to the power distribution game rule, and extracting a continuous deviation accumulation region. According to the method, the accuracy and the real-time performance of power distribution fault prediction of the power supply are improved.

Description

Technical Field

[0001] The present invention relates to the field of power supply fault prediction, and in particular to a power supply distribution fault prediction method and system based on BMS data analysis and driving. Background Art

[0002] With the rapid development of new energy vehicles, energy storage systems, and smart grids, the performance and reliability of power battery systems, as core components for energy conversion and output, have become critical for the stable operation of power systems. The Battery Management System (BMS), a key control unit that monitors and manages the operating status of battery packs, plays a crucial role in estimating voltage, current, temperature, SOC, and fault detection. Especially with the increasing prevalence of high power density and complex operating conditions, achieving reasonable distribution and precise scheduling of battery power has become a critical factor affecting vehicle operating efficiency and system lifespan.

[0003] In actual operation, power battery systems often face numerous challenges. Firstly, due to variations in consistency between individual cells (such as internal resistance, capacity aging, and thermal response), parallel or series operation of multiple cells can easily lead to problems such as uneven power distribution, local overload, and premature cell degradation. Secondly, external environmental factors such as extreme temperatures, rapid charge and discharge conditions, and sudden load disturbances can exacerbate power distribution fluctuations, leading to the accumulation of scheduling errors and the hidden evolution of potential faults. Furthermore, because traditional BMS systems primarily rely on static models and threshold judgments, they lack the ability to deeply identify the evolving characteristics of cell behavior, resulting in a significant inability to detect power distribution anomalies and early faults. Current methods for detecting battery power distribution anomalies rely primarily on preset rules, empirical models, and fixed alarm logic, such as those based on maximum / minimum voltage deviations, overtemperature alarms, or SOC inconsistency. While these methods can provide some protection under specific conditions, they cannot dynamically adapt to the multi-dimensional evolution of cell conditions and lack the comprehensive ability to identify complex power disturbances, power distribution imbalances, and state degradation. In addition, existing systems mainly rely on post-diagnosis, making it difficult to provide early warning of abnormal power trends and intelligently adjust scheduling paths. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a power distribution fault prediction method and system based on BMS data analysis and driving to solve at least one of the above technical problems.

[0005] To achieve the above objectives, the present invention provides a power distribution fault prediction method based on BMS data analysis and drive, comprising the following steps: Call the BMS to collect historical power monitoring logs, analyze periodic power disturbance changes and normalized power perturbation evolution, and construct normalized power perturbation state diagrams for multiple battery cells; Based on the historical power monitoring log, the battery cell status evolution is mined and multi-dimensional personalized status perception is performed to build a personalized status profile for each battery cell. Calculate the requested power deviation of the battery cells, perform power allocation coordination analysis between the battery cells, and mine the power allocation game logic based on the personalized status profile to identify the power allocation game rules of the battery cells; According to the power allocation game rule, power allocation prediction and power allocation trajectory deviation comparison are performed to extract the continuous deviation accumulation area; Based on the normalized power perturbation state diagram, power distribution paths are mined in the continuous deviation accumulation area, and operation state simulation self-learning optimization is performed to build a fault prediction optimization model.

[0006] In this specification, a power distribution fault prediction system based on BMS data analysis and driving is provided, which is used to execute the power distribution fault prediction method based on BMS data analysis and driving as described above, including: The normalized power state module is used to call the BMS to collect power supply history monitoring logs, perform periodic power disturbance change analysis and normalized power perturbation evolution, and construct normalized power perturbation state diagrams for multiple battery cells; The personalized status module is used to mine the evolution of each battery cell's status based on the power supply historical monitoring log, and perform multi-dimensional personalized status perception to build a personalized status portrait for each battery cell; The power allocation rule module is used to calculate the requested power deviation of the battery cell, perform power allocation coordination analysis between the battery cells, and mine the power allocation game logic based on personalized status profiles to identify the power allocation game rules of the battery cells; An allocation trajectory deviation module is used to perform power allocation prediction and power allocation trajectory deviation comparison according to the power allocation game rule, and extract continuous deviation accumulation areas; The fault prediction optimization module is used to mine the power distribution path of the continuous deviation accumulation area based on the normalized power perturbation state diagram, and perform operation state simulation self-learning optimization to build a fault prediction optimization model.

[0007] The beneficial effects of the present invention include: By modeling periodic power disturbances in historical monitoring logs, the perturbation response behavior of battery cells under normal operation can be restored, forming a "power fluctuation fingerprint." Constructing a normalized perturbation state diagram helps establish the power perturbation boundary of each battery cell in its healthy state, providing a stable reference baseline for subsequent identification of power deviations. The perturbation state diagram incorporates full-cycle, low-amplitude fluctuation information, effectively enhancing the model's insight into long-term stability evolution and avoiding reliance on abnormal peak characteristics. By mining the historical state evolution paths of battery cells, implicit behavioral characteristics such as degradation patterns and thermal response differences during long-term operation can be identified. Personalized state profiling provides a power allocation basis tailored to battery cell characteristics, enabling differentiated management. By identifying "redundant healthy cells" and "marginally degraded cells," more rational battery cell scheduling strategies are supported, improving overall system energy efficiency. Game modeling can characterize the nonlinear responses of multiple battery cells under resource scheduling and identify power coordination conflicts caused by differences in health status between cells. Identifying the game laws of power allocation can be used to modify current static allocation mechanisms and achieve dynamic optimization of scheduling rules. Introducing a game theory mechanism into the scheduling model helps explain the internal logic of allocation strategy changes, enhancing the interpretability and trustworthiness of the prediction model. The continuous deviation interval between the predicted and actual trajectories provides early detection of imbalances in system power scheduling. Continuously accumulating deviation areas reflect the inadequacy of the prediction model in actual operation, providing feedback on the model's effectiveness and areas requiring optimization. Continuous deviation accumulation areas are often a typical sign of critical system instability, helping to identify potential pre-fault windows and facilitate the development of early intervention strategies. Path matching of deviation areas with the normal power perturbation state graph accurately reconstructs the trajectory of power anomaly evolution, improving fault identification accuracy. Operational simulation optimization allows real-world operational deviation feedback to be incorporated into the model's self-update process, continuously improving the model's adaptability to different battery cells and operating conditions. Through the self-learning path of "prediction-verification-feedback-adjustment," a stable closed-loop optimization system is formed, significantly improving the stability and real-time performance of fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a schematic flow chart of the steps of a power distribution fault prediction method based on BMS data analysis and driving according to the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Schematic diagram of the detailed implementation steps of step S3. DETAILED DESCRIPTION

[0009] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0010] This application provides a power distribution fault prediction method and system based on BMS data analysis. The execution entities of this power distribution fault prediction method and system based on BMS data analysis include, but are not limited to, the following: mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc. equipped with the system, which can be regarded as general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio and image management system, an information management system, and a cloud data management system.

[0011] See also Figures 1 to 4 The present invention provides a power distribution fault prediction method based on BMS data analysis and driving, comprising the following steps: Call the BMS to collect historical power monitoring logs, analyze periodic power disturbance changes and normalized power perturbation evolution, and construct normalized power perturbation state diagrams for multiple battery cells; Based on the historical power monitoring log, the battery cell status evolution is mined and multi-dimensional personalized status perception is performed to build a personalized status profile for each battery cell. Calculate the requested power deviation of the battery cells, perform power allocation coordination analysis between the battery cells, and mine the power allocation game logic based on the personalized status profile to identify the power allocation game rules of the battery cells; According to the power allocation game rule, power allocation prediction and power allocation trajectory deviation comparison are performed to extract the continuous deviation accumulation area; Based on the normalized power perturbation state diagram, power distribution paths are mined in the continuous deviation accumulation area, and operation state simulation self-learning optimization is performed to build a fault prediction optimization model.

[0012] In the embodiment of the present invention, see Figure 1 , is a schematic flow chart of the steps of a power distribution fault prediction method based on BMS data analysis and driving according to the present invention. In this example, the steps of the power distribution fault prediction method based on BMS data analysis and driving according to the present invention include: Call the BMS to collect historical power monitoring logs, analyze periodic power disturbance changes and normalized power perturbation evolution, and construct normalized power perturbation state diagrams for multiple battery cells; In this embodiment, by performing periodic perturbation modeling and perturbation state identification on the power fluctuation data at the cell level in the power operation log recorded by the BMS, a "normalized power perturbation state diagram" is constructed to characterize the stability and trend of the long-term power behavior of multiple cells, providing a reference benchmark for subsequent fault warning and power coordination analysis. First, the system retrieves the complete cell-level monitoring data covering the past 90 days from the BMS. The data sampling frequency is 1 Hz, and the fields include the instantaneous voltage, current, temperature and time scale of each cell. By calculating the instantaneous power of the cell ; Obtain the cell time-series power sequence. To remove low-frequency trend interference, a sliding window (60s) mean filter is used as the baseline, and then the baseline is subtracted from the original power sequence to retain the small disturbance components. The system performs periodic disturbance detection on the power perturbation sequence of each cell. Methodologically, the multi-scale wavelet transform (MSWT) is used to perform spectral decomposition of the power perturbation sequence and extract high-frequency wavelet coefficients under typical power perturbation periods (such as 5s, 15s, 30s, and 60s). Then, the dominant perturbation frequency is identified through Fourier envelope spectrum analysis, and periodic perturbation indicators such as perturbation amplitude, perturbation frequency stability, and perturbation duration are calculated to obtain the periodic perturbation feature vector of each cell. To eliminate transient perturbation interference caused by load changes, the system introduces a load state marker and performs perturbation feature normalization. Based on the periodic perturbation characteristics, the system further constructs a "normalized power perturbation state diagram." The graph represents each battery cell as a node, and edge weights represent the correlation between power perturbation behaviors across different cells within the same time period (calculated using the Pearson correlation coefficient or dynamic time warping (DTW) distance). The system categorizes nodes into three categories: stable, warning, and volatile zones based on perturbation amplitude (low, medium, and high) and frequency fluctuation. A power perturbation density map is introduced for visualization. This state diagram serves as a long-term baseline for the cell's power health, providing a reference for abnormal behavior in subsequent diagnosis and enabling path suppression and priority allocation in multi-cell power coordination. In experimental validation, the system was deployed on a new energy vehicle fleet, analyzing the historical power fluctuations of nearly 80,000 battery cells. Statistics show that the system's constructed state diagram accurately reflects the stability differences between individual cells under different load cycles, achieving 91.3% accuracy in identifying periodic perturbations. It also successfully captured the transition of some cells from the warning zone to the volatile zone 21 days in advance, providing critical prior information for triggering subsequent fault prediction models. It can be seen that this step is of great significance in realizing data-driven power perturbation behavior modeling and is the key foundation layer for the entire cell state perception and allocation game analysis.

[0013] Based on the historical power monitoring log, the battery cell status evolution is mined and multi-dimensional personalized status perception is performed to build a personalized status profile for each battery cell. In this example, historical monitoring log data collected over a long period of time is retrieved from the BMS system. The data covers key indicators such as cell voltage, current, temperature, SOC, SOH, charge / discharge rate, power request and response, and internal resistance. A recommended collection window of at least seven days and a sampling frequency of at least 1Hz are recommended to accurately capture the evolutionary trends of the cell's operating state. The first stage is "cell state evolution mining," the core of which is to establish a continuous state trajectory for each cell over time. This method converts the raw multi-dimensional operating data into a time series trajectory vector. Common methods include sliding window feature extraction, time series clustering, and Markov chain modeling. Taking the sliding window as an example, from the 72-hour charge / discharge log of cell #06, a state segment is extracted every 10 minutes. Each segment includes dimensions such as the SOC curve, current fluctuation, temperature rise rate, and internal resistance change. Dynamic Time Warping (DTW) is then used to calculate the similarity between these state segments to construct a time series graph of the cell's state evolution. The second stage is "multi-dimensional personalized state perception." Its goal is to combine the cell's state trajectory with its specific usage context, structural characteristics, and behavioral response model to form a multi-faceted state perception. In this stage, the system not only considers the cell's current operating values ​​but also models the individual characteristics of the cell based on its historical operating intensity (such as total charge and discharge cycles and average rate), usage environment (temperature profile), and system scheduling role (whether it frequently experiences high-load scheduling links). Dimensionality reduction methods such as principal component analysis (PCA) or t-SNE can be used to visualize the state characteristics, and clustering algorithms (such as DBSCAN) can be used to identify abnormal trajectories or drift trends in the state evolution distribution. For example, in a comparison experiment involving multiple cell groups, cell #03 exhibited a sustained increase in temperature rise rate (>1.5°C / min) and increased unit power fluctuation (>0.8W / s) within the high SOC range (70-90%). While the temperature rise rates of other healthy cells were mostly below 1°C / min, this indicates a personalized deviation between its thermal characteristics and power response. Through the state perception mechanism, the cell was classified as having an "abnormal power thermal response" profile. The third stage involves "personalized state profile construction," which involves structured integration of the multi-dimensional behavioral parameters obtained above to generate a "state profile" for each cell. This profile includes the following key components: ① Basic attributes (such as capacity and internal resistance calibration values); ② Summary of state evolution trajectories (such as maximum disturbance range and state transition rate); ③ Typical operating modes (such as the power response curve at high charge rates and the discharge temperature rise curve); ④ Risk indices (such as fluctuation amplitude index and abnormal frequency index); and ⑤ Operational stability labels (such as "thermal susceptibility," "rate hysteresis," and "SOH drift"). This state profile not only enables subsequent adaptive adjustment of power allocation strategies but also serves as a key input dimension for fault warning models.

[0014] Calculate the requested power deviation of the battery cells, perform power allocation coordination analysis between the battery cells, and mine the power allocation game logic based on the personalized status profile to identify the power allocation game rules of the battery cells; In this embodiment, the difference between the "target allocation" and "actual execution" of each battery cell during the power scheduling process is quantified, thereby revealing the response error between the scheduling instructions and the physical execution level. The specific operation process is: first, extract the requested power (allocated by the scheduling controller) and the actual allocated power (calculated by the product of the actual voltage and current) of each battery cell at a certain moment from the BMS log. The difference between the two is the requested power deviation of the battery cell: During the experiment, deviation data was collected for 30 consecutive minutes at a sampling rate of 1Hz during a highly dynamic operating cycle (e.g., mixed uphill and downhill conditions, with an SOC range of 30%-60%). Normally, the deviation for stable and healthy cells remains within ±1W. However, cells experiencing response lag, degradation, or temperature control issues often experience deviations exceeding ±3W, and can even experience short-term sudden changes of ±5-6W. Cell #03 exhibited continuous deviations exceeding 45W·s during multiple high-speed acceleration sections, indicating a decrease in its dispatch responsiveness under these conditions.

[0015] By comparing the power request deviations across all cells in a time series and identifying coordination relationships, it is possible to determine whether the system as a whole is experiencing scheduling imbalances, local overloads, or power conflicts. Analysis methods include covariance matrix analysis, synchronization volatility calculation, and scheduling inconsistency index construction. For example, a power deviation coordination matrix is ​​constructed across all cells. If the deviation directions between certain cells are consistent over a long period of time or compensate for each other, this indicates scheduling dependency. Furthermore, inter-cell power scheduling consistency metrics (such as average synchronization rate or maximum deviation direction angle) can be calculated to assess the overall power allocation coordination within the cell cluster.

[0016] In an urban traffic data segment with a SOC of 50%-80% and an ambient temperature of 25°C, cells #04, #06, and #09 exhibit significant non-cooperative fluctuations, with a synchronization ratio of only 0.18 (less than the overall average of 0.35), suggesting that these cells' performance differences may affect the stability of the scheduling process. The previously constructed personalized cell state profiles serve as the behavioral foundation for scheduling decision analysis, identifying each cell's willingness, ability, and competitiveness in power scheduling. The method draws on the mixed strategy Nash equilibrium concept from game theory, treating power allocation as a multi-agent, resource-constrained, non-cooperative game problem. Each cell is assigned a "power preference vector" based on its state profile parameters (such as temperature rise capability, rate tolerance, and historical power response capability). The system dynamically adjusts power allocation priorities based on the preference vectors of multiple cells and the current scheduling demand distribution. For example, battery cell #05 is of the "rate-stable type" and can maintain low disturbance even under high SOC + high-rate conditions, so its "gambling willingness" is strong; while battery cell #08 is of the "temperature control sensitive type" and actively lowers the allocation target when the temperature exceeds 35°C, and its "gambling concession behavior" is obvious.

[0017] According to the power allocation game rule, power allocation prediction and power allocation trajectory deviation comparison are performed to extract the continuous deviation accumulation area; In this example, the previously constructed power allocation game rules for each cell are transformed into a dynamic prediction model to predict the power allocation that each cell will receive over multiple future time steps. This process relies on two key inputs: one is external conditions such as the current system load demand, SOC distribution, and temperature field changes; the other is a "power willingness vector" derived for each cell based on its personalized state profile and the game logic. Specifically, a sequence prediction model, such as an LSTM (Long Short-Term Memory) network or a Transformer-based architecture, is often used. This model takes as input the actual power allocation values, request deviation values, environmental variables, and state profile indicators for each cell over the past N time steps, and outputs a power prediction sequence for the next T time steps. For example, during a high-speed driving cycle (ambient temperature 30°C, total load current stable at 95A), cell #06 has exhibited stable discharge over the past five minutes. Based on its "rate-responsive" profile label, the prediction model predicts that it will consistently receive above-average power (approximately 11.2W to 12.6W) for the next 60 seconds. Simultaneous predictions for multiple cells construct a "power allocation prediction trajectory matrix." The actual power allocation trajectory is aligned with the predicted trajectory, and the deviation trajectory between the two is calculated. To improve accuracy, the predicted and actual trajectories must first be timestamp-aligned and filtered to eliminate the effects of system sampling jitter and high-frequency noise. The difference calculation uses a sliding window mean square error (MSE) or dynamic time warping (DTW) distance metric to identify short-term and medium-term deviations. Within the forecast deviation data for all cells and the entire cycle, regions where deviations show a persistent accumulation effect are identified. These regions often indicate systemic risks such as potential scheduling imbalances, delayed cell status response, or uncontrolled evolution of conflict. Key methods for extracting these deviations include the cumulative deviation integration method, the fluctuation trend slope method, and local extreme value statistics. Using the cumulative deviation integration method as an example, the deviation threshold for each cell is set at ±2W·s. If the absolute value integral of the deviation exceeds 20W·s continuously within a 120-second analysis window and shows no signs of self-recovery (the average deviation value is consistently positive or negative), the segment is marked as a "continuous deviation accumulation region." During a downhill braking event with the SOC dropping from 65% to 40%, cell #09 continuously experienced unexpectedly high power allocations (average deviation +1.8W), with a cumulative deviation of 28.9W·s over 135 seconds, marking it as a "high-risk allocation drift zone." This extracted region of accumulated deviations will ultimately serve as an important input to the fault prediction system, used in subsequent steps to trace the cause of the deviations, quantitatively assess cell risks, and design power allocation optimization strategies. This process completes a logical loop from model prediction to abnormal behavior identification, improving the system's sensitivity and responsiveness to abnormal cell-level power evolution.

[0018] Based on the normalized power perturbation state diagram, power distribution paths are mined in the continuous deviation accumulation area, and operation state simulation self-learning optimization is performed to build a fault prediction optimization model.

[0019] In this embodiment, an in-depth analysis of the identified continuous deviation accumulation area is performed based on the normalized power perturbation state diagram constructed in the early stage. The goal is to identify the power distribution conduction path and influence chain between the battery cells during the deviation accumulation process. In terms of method, the battery cells in the deviation accumulation area (such as battery cells #07 and #09) are first selected, and their power perturbation evolution trajectory before and after the deviation occurs is traced back. By comparing with the perturbation trajectory corresponding to it in the normal map, it is identified whether it deviates from the normal perturbation pattern, such as an increase in perturbation frequency, a sudden increase in amplitude, or a shortening of the period. To analyze the power path structure, a "power distribution influence network" can be constructed using a weighted directed graph. Each node in the graph represents a battery cell, and the edge weight represents the power distribution correlation between the two battery cells (such as the Pearson correlation coefficient, covariance, or synchronous perturbation index). In an experiment conducted in the SOC range of 40% to 60% with a stable system current of approximately 85A, it was found that before the deviation accumulation in cell #09 occurred, cells #07 and #08 exhibited significant synchronous perturbations under high-rate operating conditions. The edge weight increased from 0.12 under normal conditions to 0.48, suggesting a scheduling chain coupling relationship between the three in this operating range. This structure was identified as an "abnormal conduction path" in power distribution, a potential source of power anomalies. Based on the identification of the abnormal path, the system simulates power scheduling behavior and deviation evolution trends under different operating conditions to verify the sensitivity of fault evolution and train a fault prediction model. The core of the operational simulation is to establish a proxy model of power allocation behavior based on existing data. Scenario reconstruction is used to simulate the power scheduling results between cells under different parameter perturbations. Reinforcement learning or data-based system modeling (such as system identification and gray-box modeling) can be used to simulate and reproduce the BMS scheduling mechanism. In one experiment, by increasing the thermal response parameters of cell #07 by 10% and lowering the SOC threshold by 5%, and simulating its dispatch response trajectory under a similar ambient temperature (approximately 32°C), the system found that the power allocation deviation increased from an average of 1.2W to 3.8W within 15 minutes, inducing a synchronous power drift in cell #09, thus forming a new round of deviation coupling chain. This result was recorded by the system as a high-risk behavior chain for model training. Combined with the dataset generated by the operational simulation, the system used supervised learning methods (such as XGBoost, LightGBM, and multi-dimensional decision trees) to build a model to predict whether continuous deviation areas will evolve into systemic failures. The model inputs include: ① the current and historical perturbation feature vectors of each cell; ② the synchronous fluctuation index of the power conduction path to which it belongs; ③ the response sensitivity indicator from the personalized state profile; and ④ external environmental conditions such as temperature, SOC, and rate of change. The training objective is to predict, for any deviation point, whether it will cause a conduction deviation or systemic dispatch imbalance within a future period (e.g., 5 or 10 minutes).In the experimental training set, the model achieved 87.6% accuracy in predicting fault evolution on test data and successfully predicted the power collapse risk of a group of cells (#06, #08, and #10) below 35% SOC approximately 180 seconds in advance. Ultimately, this optimization model was deployed as part of the BMS scheduling assistance engine, enabling real-time prediction of potential deviation risks and dynamic scheduling path adjustments during operation. The system not only enables proactive intervention in fault risks but also redistributes cell power load based on risk weights, significantly improving the overall robustness and lifecycle management capabilities of the system.

[0020] In this embodiment, refer to Figure 2 The specific steps of calling BMS to collect power supply history monitoring logs, perform periodic power disturbance change analysis and normalized power perturbation evolution, and construct normalized power perturbation state diagrams of multiple battery cells are as follows: Call BMS to collect power supply history monitoring logs; Identify power cells based on historical power monitoring logs and analyze cell power under different load, temperature, and SOC conditions to obtain the power of each cell under different load conditions. Calculating the slight power fluctuation of each battery cell power and extracting the slight power fluctuation parameters of the battery cell; Performing a multi-scale wavelet transform on the tiny power fluctuation parameter to generate a power fluctuation spectrum diagram; Perform periodic power disturbance change analysis on the power fluctuation spectrum to extract the periodic power noise component of each battery cell; Normalized power perturbation evolution is performed based on the periodic power noise component to construct normalized power perturbation state diagrams of multiple battery cells.

[0021] In this implementation, a secure session is established with the BMS communication module via the vehicle-mounted or test-mounted CAN bus. The "Data-Logger" interface is then used to batch download the past three months of historical power monitoring logs. The logs contain over 20 fields, including date and time stamp, cell serial number, voltage, current, instantaneous power, temperature, SOC (state of charge), and cycle count, totaling approximately 30 GB of data. To ensure data continuity and integrity, the system first performs a hash check and incrementally fills in missing segments before importing the original CSV logs into distributed storage (such as HDFS). In a Spark environment, engineers resample the logs using trip ID as the partition key and a 10-minute window. Kalman filtering is also used to remove outlier current and voltage samples caused by communication interference. The pre-cleaned dataset exhibits monotonically increasing time series characteristics, making it suitable for direct input into the subsequent cell identification and power analysis processes. The collected monitoring logs are structured to identify and differentiate different cells and analyze their power performance under multiple operating conditions. First, the instantaneous power of each cell is calculated using the voltage and current data in the BMS log using the formula P_cell = V_cell × I_cell. Due to the large number of cells, each cell must be clearly identified by its physical number or internal BMS grouping information. Under each load condition (e.g., light load <30A, medium load 30-80A, and heavy load >80A), the average power and fluctuation range of each cell are calculated for different SOC ranges (in 10% intervals) and temperature ranges. For example, at 25°C, SOC between 50% and 60%, and a load current of 40A, the average power of cell #07 is 18.5W with a standard deviation of 1.2W. This process utilizes dimensionality reduction methods such as clustering algorithms (such as K-means) or principal component analysis (PCA) to classify the cell power characteristics and identify cells with similar power characteristics or those with abnormal drift. To ensure statistical significance, each operating condition must contain no fewer than 500 data points to support the accuracy of subsequent perturbation modeling. This study explores subtle fluctuations in cell power, specifically the fine-grained variations in cell power after removing the load trend. These fluctuations typically manifest as high-frequency jitter at the milliwatt level, a key indicator of the cell's internal dynamic response (such as chemical reaction rate and electrochemical polarization) and aging characteristics.

[0022] Each cell power time series needs to be smoothed to remove trends. Using a sliding average (window length, such as 30 seconds) or a local polynomial regression method (such as LOESS), the "residual" portion of the original power series is extracted as a small fluctuation signal. Next, the residual fluctuation is normalized to zero mean and unity variance to ensure comparability in subsequent analyses. To quantify small fluctuations, statistical features such as root mean square (RMS), kurtosis, kurtosis, and autocorrelation coefficient can be extracted. For example, the small power fluctuation of cell #15 at a load of 40A and a SOC of 70% has an RMS of 0.85W and a kurtosis of 3.1, indicating the presence of certain spike-like high-frequency perturbations. These fluctuation parameters can be used to preliminarily determine whether the cell has abnormal perturbation characteristics, providing a data foundation for subsequent spectral analysis. The cell's small power fluctuation signal is expanded in the frequency domain using the Multi-Scale Wavelet Transform (MSWT) to extract the energy distribution at different frequency levels. Compared to the Fourier transform, the wavelet transform is more suitable for processing non-stationary signals. It can capture local characteristics of power signals in both the time and frequency dimensions and is suitable for identifying fine-grained disturbance patterns. The Daubechies wavelet (e.g., db4) from the discrete wavelet transform (DWT) was used as the mother wavelet, with a decomposition level of 5 to 7 to extract the energy spectrum distribution from 0.1 Hz to several Hz. In the experiment, a 1-minute sliding window was used for the transform, with each window containing 60 sample points. The transform results are displayed as a scalogram, with time plotted on the horizontal axis and frequency levels plotted on the vertical axis, with color representing energy intensity. For example, within a certain window, the high-frequency energy (above 2 Hz) of cell #03 suddenly increases, potentially indicating rapid dynamic changes within the cell, such as short-term resistance variations or localized aging effects. The scalogram can be used to subsequently extract periodic disturbances and assist in locating cells experiencing power instability. Periodic disturbance components in cell power fluctuations are identified from wavelet spectrograms. These components may reflect periodic processes such as internal structural aging, load reflow, and thermal cycling. The analysis method uses spectral kurtosis, periodogram analysis, and time-frequency clustering methods (such as Short-Time Fourier Transform clustering) to extract periodic features. Specifically, the dominant frequency points in the spectrogram of each cell are extracted within a selected frequency band (e.g., 0.5Hz-2Hz), and their energy stability and periodic reproducibility across different windows are analyzed. For example, in the sample window, cell #12 exhibits a significant power disturbance with an 8-second period and a frequency of approximately 0.125Hz. This disturbance persists across five adjacent windows, indicating that the periodic disturbance is a stable noise component. This periodic power noise component can be quantified using metrics such as disturbance period, peak energy, and modulation depth, providing a foundation for subsequent normalized perturbation evolution models.

[0023] The periodic power disturbance components of each cell are standardized and time-series modeled to construct a "normalized power perturbation state diagram" that characterizes the health of the entire power system. First, periodic noise indicators (such as main frequency, energy density, and disturbance amplitude) are normalized to make the disturbance states of each cell comparable at different times and under different operating conditions. Using a sliding window method (e.g., a window every 5 minutes), the trend of the disturbance indicators within each window is analyzed to construct a cell perturbation evolution trajectory diagram. A multidimensional dynamic time warping (DTW) method is used to compare the similarity of the evolution paths between cells, identifying cells whose perturbation trends deviate from the normal state. The perturbation evolution trajectories of all cells are superimposed to construct a "normalized power perturbation state diagram," with time on the horizontal axis and cell number on the vertical axis. Color intensity indicates perturbation intensity. For example, after four hours of continuous operation, the disturbance energy of cell #18 gradually increases and becomes significantly higher than that of its neighboring cells, potentially indicating accelerated aging or localized failure. This state diagram can serve as a visual front-end for the BMS fault warning system, providing support for cell degradation prediction and power distribution optimization.

[0024] In this embodiment, refer to Figure 3 The specific steps of mining the evolution of each battery cell's state based on the power supply history monitoring log and performing multi-dimensional personalized state perception to build a personalized state portrait of each battery cell are as follows: Based on the power supply history monitoring log, the voltage, current and temperature generated during the battery cell's historical charge and discharge cycles are extracted to obtain the battery cell's high-frequency operation monitoring parameters. Perform time series behavior fitting on the high-frequency operation monitoring parameters of the battery cells to construct a time series behavior trajectory set for each battery cell; Mining the state evolution of each cell one by one on the set of time series behavior trajectories to extract the cell state evolution path; Perform long-term trend analysis and self-learning on the cell state evolution path to obtain the dynamic state transfer vector; Perform multi-dimensional personalized state perception on the dynamic state transfer vector to build a personalized state portrait of each battery cell.

[0025] In this example, voltage, current, and temperature change data for each battery cell over multiple complete charge and discharge cycles is extracted from the BMS's historical monitoring logs, focusing on dynamic performance parameters reflected during high-frequency operation. "High-frequency operation" refers to the battery cell's response to rapid voltage changes, frequent switching between operating modes, or high current loads in a short period of time. This includes situations such as rapid acceleration or sudden stopping, or frequent uphill and downhill driving in electric vehicles. A data sampling frequency of at least 1Hz is recommended, ideally 5-10Hz, to ensure capture of high-frequency dynamic details. In the experimental data, the monitoring cycle for each battery cell should cover at least 30 complete charge and discharge cycles, including peak current fluctuations and thermal environment changes under daily driving conditions, to ensure comprehensive and representative data. Extracted data fields include timestamp, cell number, voltage, current, surface temperature, and SOC. The data is then segmented along the timeline, organizing it into several "high-frequency response segments," such as segments with current change rates exceeding 10A / s or voltage drops exceeding 50mV / s, which serve as key event windows. High-frequency operating monitoring parameters extracted from these windows, such as voltage response time, current sudden change coefficient, and temperature rise rate, are key fundamental data reflecting the cell's responsiveness and degradation trends. The high-frequency monitoring parameters extracted in the first step are modeled using time series to form a structured "cell temporal behavior trajectory." Dynamic analysis of the cell's voltage, current, and temperature change sequences captures its behavioral patterns at different operating stages, and mathematical fitting is performed to extract its temporal characteristics. For each "high-frequency response segment," curve fitting techniques (such as nonlinear least squares, B-splines, or piecewise polynomials) are used to model the voltage-time, current-time, and temperature-time curves. The fitting process not only compresses the data dimension but also extracts representative temporal characteristic parameters, such as voltage dip curvature, current response slope, and temperature response lag. Subsequently, the fitted curve parameters for each cell at different time points and operating conditions are organized into a time series vector set to construct a "cell temporal behavior trajectory set." This trajectory set essentially represents the "time evolution space" of the cell over multiple operating cycles, with each trajectory point representing the dynamic characteristics of a micro-high-frequency response event. From the constructed battery cell time series behavior trajectory, we explore the evolution of the battery cell state over time and extract the state transition path, namely the "battery cell state evolution path." This path describes the process of battery cells gradually aging from a new state or transitioning from a healthy state to an abnormal state, and is an important basis for power allocation prediction.

[0026] Using a Markov Chain model, the cell behavior trajectory vector is discretized into states. For example, each response segment is mapped into several "discrete states" based on the threshold range of trajectory parameters (such as voltage response time, current peak, and temperature rise rate). Next, the transition probabilities between states are statistically analyzed to form a state transition matrix. Furthermore, a hidden Markov model (HMM) can be introduced to address the issue of unobservable states, allowing the underlying state evolution process to be inferred from the observation sequence. For example, cell #21 exhibits a high-frequency response characteristic that gradually evolves from an initial "fast response and slow temperature rise" state to a "sluggish response and rapid temperature rise" pattern, suggesting a risk of thermal runaway. For experimental data, it is recommended to construct a trajectory evolution sample library on a monthly basis, selecting data spanning three to six months to observe how cell states evolve. By comparing parameters such as state path length and state transition frequency across different cells, cells with abnormal evolution rates can be identified, facilitating subsequent trend analysis. Linear trend analysis and sliding regression techniques (such as Ridge regression) are used to predict the evolution of state transition probabilities over time. If the probability of a cell transitioning to a "high-risk state" shows a sustained upward trend, it may indicate accelerated degradation. These trend data are then reinforced using self-learning methods to model the state transition. For example, a time-difference-based reinforcement learning algorithm (such as TD(λ)) is employed to continuously adjust the state transition prediction strategy, thereby improving the ability to predict future state changes. Dynamic state transition vectors consist of several time-varying features, such as the direction, intensity, rate, and jump threshold of each state transition. These vectors not only record the current evolution trend of the cell but also its historical state dependencies, thus possessing a "memory" capability. In experimental environments, the number of daily cell operation samples can be set to exceed 5,000 state transition records to ensure the stability of the trend analysis model. Based on this, the dynamic state transition vectors can accurately reflect the abnormal paths that may be triggered by the cell during long-term operation, providing the necessary conditions for subsequent personalized modeling.

[0027] Dynamic state transition vectors are used to construct a personalized state profile for each battery cell, characterizing its unique state behavior characteristics, evolution patterns, and potential risk factors from multiple dimensions. By introducing a multidimensional modeling framework, the state transition vectors are converted into a structured state recognition model, enabling refined fault warning and differentiated power management. The dimensions of the personalized state profile include, but are not limited to, state change frequency (F_var), state jump amplitude (A_jump), response hysteresis (L_delay), load sensitivity (S_load), and thermal behavior characteristics (T_heat). These dimensions are fused using dimensionality reduction methods such as principal component analysis (PCA) to ultimately form a feature vector representation of each cell in a high-dimensional state space. Furthermore, the state profiles of multiple cells are visually clustered using self-organizing maps (SOM) or t-SNE techniques to identify similar state groups and potentially abnormal individuals. For example, the state profile of cell #07 deviates significantly from the normal cluster center, indicating possible internal aging inconsistencies, structural defects, or manufacturing deviations. This state profile serves as the basis for individual identification within the power distribution system, enabling scheduling algorithms to respond differently based on actual conditions (such as limiting the discharge power of high-risk cells or initiating maintenance warnings). By continuously updating the state profile, a dynamic, self-learning, and personalized cell state perception system can be built, significantly enhancing the BMS's ability to predict and control power distribution anomalies.

[0028] In this embodiment, the specific steps of calculating the requested power deviation of the battery cells, performing power allocation coordination analysis between the battery cells, and mining the power allocation game logic based on the personalized state profile to identify the power allocation game rules of the battery cells are as follows: Identify cell-by-cell power requests and actual allocated power parameters; Calculating a requested power deviation between the power request and the actual allocated power parameter; Calculating the total power of the power supply and calculating the cell-by-cell distribution ratio of the actual distributed power parameters to generate a time-series power distribution ratio curve; Performing inter-cell power allocation coordination analysis on the sequential power allocation ratio curve according to the requested power deviation to generate inter-cell power allocation coordination parameters; Based on the personalized state portrait, the power allocation game logic is mined for the power allocation coordination parameters between the battery cells to identify the power allocation game rules of the battery cells.

[0029] In this embodiment, the system identifies the "theoretical power request" and "actual allocated power" for each battery cell per unit time based on historical BMS monitoring data and control command logs to construct a complete power demand-allocation relationship model. The cell's power request, P_req^i(t), is calculated based on the cell's current state of charge (SOC), temperature, and load requirements. This is typically set by the energy management system (EMS) in power scheduling instructions (e.g., automatically assigning a high-priority charging power request to cells with an SOC < 40%). The actual allocated power, P_real^i(t), is calculated by sampling the cell's actual voltage and current values ​​during the current cycle: P_real^i = U_i(t) × I_i(t). The system samples within a time window with a 100ms resolution, collecting 600 power data points per cell in one minute. By analyzing this data cell by cell, the system can establish a "time series matrix" of power demand and actual supply across the entire battery system, providing a precise basis for subsequent deviation analysis and power coordination. In a real-world scenario, facing a 96-cell power battery system, the identification process completed the precise alignment of all power requests and responses within 500ms, with an accuracy rate of 99.2%. By comparing the requested power P_req^i(t) of each battery cell in the same sampling period with the actual allocated power P_real^i(t), the power deviation ΔP^i(t) = P_req^i(t) -P_real^i(t) is calculated to form a "request power deviation" sequence. This deviation value is used to measure the power scheduling accuracy of the BMS system for the battery cell under different control logic or load conditions. For greater analytical significance, the system further normalizes the deviation value to obtain the "relative power shortage rate" r^i(t) = ΔP^i(t) / P_req^i(t) for each battery cell, which is used to analyze possible resource allocation bias problems in the scheduling system. The system uses a sliding window (e.g., a 10s window with a 2s step) to smooth the deviation sequence and extract statistical features (mean, standard deviation, maximum under-matching rate, negative center of gravity, etc.). Experimental results show that in high-rate discharge scenarios, the mean cell deviation can reach ±12%, with some experiencing severe power deviations exceeding 15% for sustained periods. This information provides fundamental support for subsequent power coordination analysis and potential overheating / imbalance risk assessment.

[0030] The P_real^i(t) values ​​of all cells within each sampling period are summed to obtain the current total power of the battery pack, P_total(t) = Σ_i P_real^i(t). The power proportion of each cell at that moment is then calculated, q^i(t) = P_real^i(t) / P_total(t). These proportions form a "power allocation ratio curve" Q^i(t) for each cell on the timeline. To clearly visualize the dynamic allocation between cells, the system uses normalized stacked plots and heat maps to dynamically display Q^i(t). Its first-order derivative is also calculated to characterize the allocation fluctuation σ^i(t). Furthermore, a moving average curve and low-pass filtering are used to smooth noise in the ratio curve, improving the accuracy of the curve stability analysis. During multiple operating cycle tests, the fluctuation of the allocation ratio curve for cells with high fatigue levels was significantly higher than that for healthy cells (an average increase of over 23%), indicating that the curve not only reflects the scheduling strategy but also provides a certain indicator of health status. The system needs to assess whether there is any coordination imbalance in power allocation behavior. First, the requested power deviation ΔP^i(t) from step S35 is cross-analyzed with the allocation ratio curve Q^i(t) from step S36. The Pearson correlation coefficient and dynamic offset are calculated to identify whether allocation behavior is dynamically adjusted according to power demand. Second, the inter-cell covariance matrix and correlation coefficient network are used to analyze the linkage between the power allocation curves. If most cells Q^i(t) show synchronized fluctuations, the system can conclude that the current power allocation is well coordinated. If a split trend (e.g., a continuous decrease in the allocation ratio of high-quality cells) is observed, the allocation is considered unbalanced. Finally, the inter-cell "power allocation coordination parameter matrix" C(i,j,t) is generated to describe the degree of scheduling coupling and coordination risk between different cells. In high-power load experiments, this analysis successfully identified multiple "coordination breakdown" regions (power request imbalance > 25% and allocation difference covariance > 0.12), providing precise guidance for predictive maintenance and scheduling corrections. By combining cell behavior with state attributes, game theory is introduced to analyze the implicit "strategy confrontation" phenomenon in the power scheduling process. The system uses a personalized profile of each battery cell (including fatigue, temperature sensitivity, response delay, and more) to map this profile with the power coordination parameters C(i,j,t) into an "allocation preference function" U_i(t), thereby constructing a multi-agent non-cooperative game model. Using evolutionary game analysis tools (such as replicator dynamics), the system simulates the strategies of different battery cells as they adjust requests or adapt to allocations during multiple rounds of scheduling, identifying groups of cells with dominant strategies and those vulnerable to exclusion. The system further incorporates the Nash equilibrium search algorithm to determine whether the current scheduling has reached a stable equilibrium.Measured data revealed that some "high-response, high-health" cells could achieve a "distribution monopoly" by continuously requesting high power, while "low-health, delayed-response" cells were chronically disadvantaged in resources, accelerating their decline. The system ultimately constructed a map of inter-cell game patterns and mapped high-risk strategy trends within the prediction module, providing a support mechanism for proactive intervention and strategic balancing. This model improved pre-fault identification rates by approximately 18%, enabling more intelligent power management and evolutionary control.

[0031] In this embodiment, the specific steps of performing power allocation prediction and power allocation trajectory deviation comparison according to the power allocation game rule and extracting the continuous deviation accumulation area are as follows: Calculate the real-time power distribution data of the battery cells based on the BMS; Performing adaptive filtering and denoising and timestamp alignment processing on the real-time power allocation data to obtain time-aligned power allocation data; Based on the time-aligned power allocation data, the power allocation game rule is used to predict the power allocation for multiple time steps in the future, and the power allocation prediction trajectory at different time points is extracted; Comparing the power allocation prediction trajectory with the preset power allocation trajectory based on trajectory deviations, and marking power allocation trajectories with abnormal deviations; Deviation points are identified on the power distribution trajectory of the abnormal deviation, and deviation accumulation detection is performed to extract continuous deviation accumulation areas.

[0032] In this embodiment, during the online phase of vehicle operation or bench testing, the BMS pushes cell voltage U_i(t) and current I_i(t) at a high frequency of 100 ms to 500 ms. The system first calculates the real-time power of each cell based on the instantaneous power formula P_i(t) = U_i(t) × I_i(t). Simultaneously, it obtains the total bus power P_total(t) from the EMS (Energy Management System) and synchronizes the timestamp. Real-time power data undergoes rapid peak truncation and 16-bit fixed-point normalization at the FPGA preprocessing layer to prevent overflow in back-end floating-point processing. To prevent data window misalignment due to communication delays, the platform establishes an NTP-based time synchronization network, with clock drift correction accuracy controlled to ±1 ms. Statistics show that on a 96-cell battery pack, this process can stably generate 960 to 1920 single-cell power records per second. Combined with MQTT streaming buffering, bandwidth usage remains stable within 150 kB / s, laying the data foundation for real-time dynamic power analysis. High-frequency power data can be affected by Hall current sensor drift, CAN jitter, and switching noise. The system first applies an adaptive filter based on recursive least squares (RLS) to P_i(t), dynamically updating the weights w(t) to minimize e(t) = P_meas(t) – P_pred(t). A filter window length of 20 points removes approximately 70% of high-frequency noise without sacrificing response speed. The timestamp alignment process then begins. Because power data is aggregated from multiple CAN branches, with subframe delay variations, the system employs a multi-channel buffer alignment algorithm (Time-Aligned FIFO) based on event timestamps, interpolating and truncating data with a 5 ms threshold to ensure that each cell's power point corresponds strictly to the same system time. After alignment, distributed Z-score depolarization (µ±4σ) and Savitzky–Golay second-order smoothing are performed to form a noise-suppressed and time-consistent power allocation matrix P_sync(i,t). The overall noise power spectral density is reduced by 18 dB, and the timing error is less than 2 ms. Based on the obtained cell power allocation game law (implied Nash migration probability), a Seq2Seq-LSTM network with attention is introduced for multi-step rolling prediction. The input tensor is composed of the concatenation of P_sync(i,t) for the past 60 seconds, the coordination parameter matrix C(i,j,t), and the individual cell profile vector Profile_i. The decoder predicts the power allocation ratio q̂_i(t+τ) for every 1 second within the next T_pred=30 seconds, with τ=1…30.The network uses a two-layer LSTM (128 units) with temporal attention and an Adam optimizer with a learning rate of 1e-3. After 50 epochs of training on a three-month dataset, the MAPE achieved approximately 4.5%. The predicted outputs are aggregated by cell dimension to form a "power allocation prediction trajectory" \hat Q_i(t:t+30s). In offline cross-validation, this model detected power skew peaks under sudden high load conditions approximately 5–7 seconds in advance, meeting early warning requirements. A benchmark library Q_ref_i(τ|scenario) of "ideal power allocation trajectories" under typical operating conditions is established, and a tolerance deviation threshold δ(τ) is defined, which decays over time. The system selects the corresponding Q_ref based on the scenario label and calculates ΔQ_i(τ) = q̂_i(τ) – Q_ref_i(τ). If |ΔQ_i(τ)| > δ(τ) for k consecutive points (k=3), it is considered an "abnormal deviation trajectory." The system simultaneously records the deviation extreme value, the duration of the threshold violation, and the SOC / temperature context at the time of occurrence, generating a structured alarm: {cell_id, t_start, t_end, ΔQ_max, scenario}. In 1000 sliding window tests, this method achieved an F1 score of 0.92 for detecting potential power imbalances, with a false alarm rate of <5%. To avoid false alarms caused by short-term glitches, the system performs a more granular analysis of abnormal trajectories. First, using the ΔQ_i(τ) sequence within the abnormal segment as a sequence, density peak clustering (DPC) is used to identify local peak "deviation points." Then, using the CUSUM (Cumulative Sum Control Chart) detection method, the deviation point sequence is statistically analyzed for fluctuation direction, and the cumulative value S_k is calculated as max(0, S_{k-1}+ΔQ_i(τ_k)-κ). When S_k exceeds the threshold H, the segment is considered to have entered the "deviation accumulation state." Finally, the region of accumulated continuous deviations is marked as a potential fault hotspot, and the output is {cell_id, τ_start, τ_end, S_peak}. In a six-month retrospective of actual fleet data, cumulative testing successfully located 27 cells experiencing subsequent rapid capacity decay or sudden increases in internal resistance 10–14 days in advance, validating the method's value in proactively warning power distribution failures.

[0033] In this embodiment, the specific steps of mining the power distribution path of the continuous deviation accumulation area based on the normalized power perturbation state diagram and performing operation state simulation self-learning optimization to construct the fault prediction optimization model are as follows: Locate the deviation cell based on the continuous deviation accumulation area; Performing a long-term power prediction analysis on the deviation battery cell to extract a power change curve of the deviation battery cell; Performing abnormal fault cell diagnosis on the power variation curve based on the normalized power perturbation state diagram, and marking the abnormal fault cell; Mining the power distribution path of abnormal fault cells to predict the power distribution path of the fault cells; The power distribution topology correlation monitoring is carried out on abnormal fault cells, and the operation state simulation self-learning optimization is performed to build a fault prediction optimization model.

[0034] In this embodiment, the "continuous deviation accumulation area" in the identified power distribution trajectory is mapped to specific cells one by one to determine which cells have continuously experienced abnormal power distribution deviations over multiple time periods, thereby locating "deviation cells" that may have performance degradation, abnormal response, or scheduling incoordination. During the implementation process, the deviation accumulation area extracted in the previous stage is first timestamped and backtracked to accurately match it with the cell ID and power time series. For example, if the predicted power trajectory of cell #07 is continuously higher than the reference trajectory within the time interval [t1, t2], the cumulative total deviation exceeds 120W·s, and the deviation lasts for more than 30 seconds, cell #07 can be preliminarily determined to be a "deviation cell." Based on the historical monitoring data and real-time operation logs of the BMS, long-term power trend modeling and prediction are performed on the marked deviation cells, and then the power change curve is extracted to identify its energy response pattern, fluctuation pattern, and possible scheduling adaptability defects.

[0035] In the experiment, we recommend setting the following judgment thresholds as positioning criteria: continuous deviation duration ≥ 20s, deviation energy per unit time > 4W·s / s, and total cumulative deviation > 100W·s. Applying these criteria to large-scale operational samples (e.g., 8 hours of full-vehicle operating data) accurately identifies high-risk battery cells. Further cross-validation combined with state profiling can improve the accuracy and robustness of deviant cell identification. The complete power distribution data for the deviant cell over the past 30 minutes or 2 hours is extracted as a time series to construct its power evolution dataset {Pi(t0),Pi(t1),...,Pi(tn)}\{P_i(t_0), P_i(t_1), ..., P_i(t_n)\}{Pi(t0),Pi(t1),...,Pi(tn)}. Subsequently, sliding window modeling techniques (such as the LSTM long short-term memory model or the GRU gated recurrent network) are used to model the temporal correlation of its historical power behavior, predicting its power trend over the next 10 to 60 seconds and constructing a "power change prediction curve." In the experimental setup, a sliding window length of 60 seconds and a prediction step size of 30 seconds are recommended. Auxiliary variables such as SOC, temperature, and cell health index (SOH) are introduced as input features to enhance model robustness. For example, power trend modeling results for cell #11 show significant hysteresis and amplification effects in its power changes under rapid uphill acceleration, with a maximum prediction error exceeding 18%, indicating unstable dispatch response. Based on a previously constructed "normalized power perturbation state diagram" (a statistical representation of the power perturbation behavior of a cell under normal conditions), the power change curves of deviant cells are compared and matched with the spectrum to identify whether they deviate from the normal perturbation range, thereby accurately diagnosing and labeling abnormal cells. The normalized power perturbation state diagram typically consists of power perturbation patterns of multiple healthy cells under standard operating conditions. The spectrum is extracted using a wavelet transform and then subjected to multi-scale feature normalization to reflect the power perturbation stability of healthy cells under different load, temperature, and SOC conditions. The diagnostic method primarily uses cosine similarity or dynamic time warping (DTW) algorithms to match the actual power change curve of a battery cell with the disturbance curve of the corresponding state in the normal state diagram. For example, within the SOC range of 40% to 60% and the temperature range of 25°C to 35°C, the frequency principal component of the disturbance curve for cell #08 deviates from the normal mean by 3.1Hz, and the disturbance amplitude exceeds the normal fluctuation band by more than twice, preliminarily identifying it as an abnormally faulty cell. Based on the power scheduling behavior of the battery cell over a period of time, the system explores its "power distribution path"—the relationship between the scheduling link and the power load assumed by the battery cell during system operation. This path reveals the energy transmission role played by the battery cell in the entire system and is key to identifying sources of system imbalance and power scheduling blind spots.

[0036] Operationally, power allocation records are expanded along a timeline to form a power allocation flow map, recording the power share of each cell in each time segment. Combined with operating condition classifications (such as acceleration, constant speed, deceleration, and coasting), a scheduling map path is generated. For example, cell #14 consistently receives more than 15% of the system power during rapid acceleration. Its power allocation path behaves as the primary dispatch unit in all three high-load segments, suggesting it may continue to serve as the system's primary output. Path visualization (such as a Sankey diagram) can visually display cell power scheduling pathways and deviation hotspots, helping to identify imbalances or abnormal concentrations in scheduling strategies. In experiments, path analysis can also be combined with game strategy states to determine whether scheduling choices are rational, such as whether the potential of healthy cells is being neglected while abnormal cells are over-deployed. By constructing a "topological structure diagram" of inter-cell power allocation relationships, scheduling dependencies and abnormal transmission links between cells can be monitored. Simulation evolution and self-learning mechanisms are also introduced to gradually improve the system's ability to perceive and respond to fault risks over multiple iterations, ultimately building a fault prediction model for power allocation optimization. For topological monitoring, a cell-to-cell graph (Graph G = {V,E}) is constructed based on the power scheduling node relationships of each cell. V represents the set of cells, and E represents the power scheduling edge weight (representing scheduling intensity). By dynamically updating the edge weight trends in the graph, core nodes with concentrated scheduling pressure and potential bottleneck cells are identified. For example, during a system load surge, the edge weight associated with cell #09 increased threefold, indicating a temporary power pressure concentration point requiring focused monitoring. For operational simulation, an agent-based scheduling behavior simulation platform is constructed. By setting different power request scenarios and cell state combinations, the system simulates possible future scheduling trends. After each simulation round, the system adjusts scheduling parameters based on deviation feedback, implementing reinforcement learning-based optimization. In experiments, a neural network predictor (such as the Transformer architecture) is trained using actual electric vehicle operating data. The power scheduling strategy is continuously optimized through multiple rounds of "prediction-verification-feedback." Ultimately, a predictive system model is developed that enables fault pre-signal detection, scheduling path reconstruction, and game-based optimization. This significantly enhances the vehicle's BMS's ability to proactively respond to power distribution failures.

[0037] In this embodiment, the specific steps of performing power distribution topology correlation monitoring on abnormal fault cells, performing operation state simulation self-learning optimization, and constructing a fault prediction optimization model are as follows: Perform power distribution topology correlation monitoring on abnormal fault cells and extract topology-related cells; Perform fault tracing analysis based on the power distribution path and topologically associated cells to identify the power distribution fault tracing point; Perform fault trend evolution analysis on the power distribution fault source point to obtain the fault trend evolution characteristics; Based on the fault trend evolution characteristics, operation state simulation self-learning optimization is carried out to build a fault prediction optimization model.

[0038] In this example, a "power allocation topology graph" is constructed between the cells in the system. Structured modeling is used to reveal scheduling dependencies, interactions, and load response coordination between cells at the power scheduling level. For identified abnormally faulty cells, the topology structure allows for further analysis to determine whether they exhibit coupled power allocation behaviors with other cells, thereby identifying potential "topologically correlated cells." A directed graph G = (V, E) is constructed, with each cell as a node and the coordinated behaviors or substitution relationships between cells as edges. Edge weights are determined by calculating the Pearson correlation coefficient, mutual information coefficient, or Bayesian dependency metrics of the power responses between cells. For example, if the power fluctuation directions of cells #05 and #11 are highly consistent across multiple operating cycles, with a mutual information value greater than 0.85, a strongly coupled edge is considered. In experiments, real BMS operating data (e.g., power scheduling records for 10 cells over a 90-minute period) is used to construct an inter-cell correlation coefficient matrix, and a threshold (e.g., 0.7) is set to screen for valid topological edges. Ultimately, a "topologically associated set of cells" forming a power allocation coordination chain with the abnormally faulty cells can be extracted. For example, cells #03, #06, and #08 exhibit significant scheduling coupling relationships with the primary abnormal cell #04 in different time windows. Dynamic path tracing and causal analysis techniques enable traceability and diagnosis of power anomalies. This involves determining which cell's scheduling anomaly or behavioral deviation initially triggered the power imbalance, thereby locating the "source cell" of the power allocation failure. This traceability analysis primarily integrates two key pieces of information: the temporal sequence of abnormal power behavior and the causal relationships between different cells. Methods such as Granger causality tests, transfer entropy analysis, and time series regression can be used to determine whether power changes in one cell predict the power changes in another. For example, under stable operating conditions with a SOC of 65% to 85% and an ambient temperature of 28°C, cell #02 experiences a power jump first, followed by the same power changes in cells #05 and #06. The Granger test indicates that the power sequence #02 has a significant causal impact on #05 (p<0.01). Based on this, the system marks #02 as the "source" of this fault chain. The experiment recommends setting the analysis window length to 60-90 seconds, the power change threshold to ±10W, and multi-directional verification based on the edge weight transfer trend within the topological path. This analysis not only identifies the first abnormal cell but, more importantly, analyzes how the fault's impact propagates, expands, and amplifies within the power topology network. For cells identified as "sources," the power deviation behavior is examined to determine whether it exhibits a regular evolutionary pattern, moving from an initial minor anomaly to a systemic risk. Trend modeling and behavioral spectrum analysis can be used to extract the "fault trend evolution characteristics." Key analytical techniques include trend slope modeling, sliding standard deviation analysis, and wavelet energy change analysis.Taking trend slope analysis as an example, linear regression or polynomial fitting techniques are used to model the time-varying curve of the cell power deviation and extract the trend slope β value. If the slope remains positive for a long period of time, it indicates that the power deviation is continuously increasing. For example, during a 30-minute high-speed cruising experiment, the standard deviation of the power deviation of cell #07 increased from an initial 1.2W to 5.6W, and the dominant frequency of the wavelet spectrum energy gradually shifted from 1.4Hz to 4.8Hz, indicating a clear nonlinear trend change. This increase in high-frequency disturbances is a typical early sign of systemic scheduling imbalance.

[0039] Evolutionary characteristics should also be modeled using multivariate factors, including ambient temperature changes, SOC range drift, and load fluctuations. Ultimately, a set of "evolutionary indicator combinations" is formed to characterize potential system-level scheduling failure risks, such as the disturbance energy increase rate, scheduling interval reduction rate, and maximum instantaneous power mutation value. Based on the extracted fault trend characteristics, a "fault prediction and optimization model" with predictive and adaptive scheduling optimization capabilities is constructed through simulation of system operation and reinforcement learning strategies. This is a key step in the evolution from static diagnosis to dynamic intervention. Specifically, a simulation environment is first established, and the power response characteristics of normal and abnormal cells are input to simulate the system's evolutionary path under different scheduling strategies. Based on the existing trend characteristics, a multidimensional input vector (such as disturbance slope, power deviation, and topological edge coupling) is constructed to train a prediction model, such as a power predictor based on a Transformer or a time-series GNN (graph neural network). The optimization strategy utilizes reinforcement learning mechanisms (such as DDPG and PPO), and the scheduling strategy is self-learned by setting reward functions (such as delaying fault occurrence, improving overall energy efficiency, and mitigating anomaly accumulation). In each iteration, the model outputs a power scheduling strategy, and the simulation system evaluates fault probability and energy loss, forming a closed learning loop. In experimental deployments, it is recommended to use high-frequency data from real-vehicle BMSs as training samples for strategy replay training. This strategy is then compared to static scheduling models to analyze fault warning lead times (e.g., whether anomalies can be predicted 30 seconds in advance) and the degree of improvement in system power balancing capabilities. The resulting fault prediction optimization model not only enables early detection of abnormal cell-level behavior but also provides forward-looking recommendations for system scheduling adjustments, making it a key component in building intelligent BMS systems.

[0040] In this embodiment, a power distribution fault prediction system based on BMS data analysis and driving is provided, which is used to execute the power distribution fault prediction method based on BMS data analysis and driving as described above, including: The normalized power state module is used to call the BMS to collect power supply history monitoring logs, perform periodic power disturbance change analysis and normalized power perturbation evolution, and construct normalized power perturbation state diagrams for multiple battery cells; The personalized status module is used to mine the evolution of each battery cell's status based on the power supply historical monitoring log, and perform multi-dimensional personalized status perception to build a personalized status portrait for each battery cell; The power allocation rule module is used to calculate the requested power deviation of the battery cell, perform power allocation coordination analysis between the battery cells, and mine the power allocation game logic based on personalized status profiles to identify the power allocation game rules of the battery cells; An allocation trajectory deviation module is used to perform power allocation prediction and power allocation trajectory deviation comparison according to the power allocation game rule, and extract continuous deviation accumulation areas; The fault prediction optimization module is used to mine the power distribution path of the continuous deviation accumulation area based on the normalized power perturbation state diagram, and perform operation state simulation self-learning optimization to build a fault prediction optimization model.

[0041] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0042] The foregoing description is intended only to provide specific embodiments of the present invention, which are intended to enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest manner consistent with the principles and novel features disclosed herein.

Claims

1. A power distribution fault prediction method based on BMS data analysis and drive, characterized in that: The following steps are involved: Call the BMS to collect historical power monitoring logs, analyze periodic power disturbance changes and normalized power perturbation evolution, and construct normalized power perturbation state diagrams for multiple battery cells; Based on the historical power monitoring log, the battery cell status evolution is mined and multi-dimensional personalized status perception is performed to build a personalized status profile for each battery cell. Calculate the requested power deviation of the battery cells, perform power allocation coordination analysis between the battery cells, and mine the power allocation game logic based on the personalized status profile to identify the power allocation game rules of the battery cells; According to the power allocation game rule, power allocation prediction and power allocation trajectory deviation comparison are performed to extract the continuous deviation accumulation area; Based on the normalized power perturbation state diagram, power distribution paths are mined in the continuous deviation accumulation area, and operation state simulation self-learning optimization is performed to build a fault prediction optimization model.

2. The power distribution fault prediction method based on BMS data analysis and drive according to claim 1 is characterized in that: The specific steps of calling the BMS to collect power supply history monitoring logs, performing periodic power disturbance change analysis and normalized power perturbation evolution, and constructing normalized power perturbation state diagrams of multiple battery cells are as follows: Call BMS to collect power supply history monitoring logs; Identify power cells based on historical power monitoring logs and analyze cell power under different load, temperature, and SOC conditions to obtain the power of each cell under different load conditions. Calculating the slight power fluctuation of each battery cell power and extracting the slight power fluctuation parameters of the battery cell; Performing a multi-scale wavelet transform on the tiny power fluctuation parameter to generate a power fluctuation spectrum diagram; Perform periodic power disturbance change analysis on the power fluctuation spectrum to extract the periodic power noise component of each battery cell; Normalized power perturbation evolution is performed based on the periodic power noise component to construct normalized power perturbation state diagrams of multiple battery cells.

3. The power distribution fault prediction method based on BMS data analysis and drive according to claim 1 is characterized in that: The specific steps of mining the state evolution of each battery cell based on the power supply historical monitoring log and performing multi-dimensional personalized state perception to build a personalized state portrait of each battery cell are as follows: Based on the power supply history monitoring log, the voltage, current and temperature generated during the battery cell's historical charge and discharge cycles are extracted to obtain the battery cell's high-frequency operation monitoring parameters. Perform time series behavior fitting on the high-frequency operation monitoring parameters of the battery cells to construct a time series behavior trajectory set for each battery cell; Mining the state evolution of each cell one by one on the set of time series behavior trajectories to extract the cell state evolution path; Perform long-term trend analysis and self-learning on the cell state evolution path to obtain the dynamic state transfer vector; Perform multi-dimensional personalized state perception on the dynamic state transfer vector to build a personalized state portrait of each battery cell.

4. The power distribution fault prediction method based on BMS data analysis and drive according to claim 1 is characterized in that: The specific steps of calculating the requested power deviation of the battery cells, performing power allocation coordination analysis between the battery cells, and mining the power allocation game logic based on the personalized state portrait to identify the power allocation game rules of the battery cells are as follows: Identify cell-by-cell power requests and actual allocated power parameters; Calculating a requested power deviation between the power request and the actual allocated power parameter; Calculating the total power of the power supply and calculating the cell-by-cell distribution ratio of the actual distributed power parameters to generate a time-series power distribution ratio curve; Performing inter-cell power allocation coordination analysis on the sequential power allocation ratio curve according to the requested power deviation to generate inter-cell power allocation coordination parameters; Based on the personalized state portrait, the power allocation game logic is mined for the power allocation coordination parameters between the battery cells to identify the power allocation game rules of the battery cells.

5. The power distribution fault prediction method based on BMS data analysis and drive according to claim 1 is characterized in that: The specific steps of performing power allocation prediction and power allocation trajectory deviation comparison according to the power allocation game rule and extracting the continuous deviation accumulation area are as follows: Calculate the real-time power distribution data of the battery cells based on the BMS; Performing adaptive filtering and denoising and timestamp alignment processing on the real-time power allocation data to obtain time-aligned power allocation data; Based on the time-aligned power allocation data, the power allocation game rule is used to predict the power allocation for multiple time steps in the future, and the power allocation prediction trajectory at different time points is extracted; Comparing the power allocation prediction trajectory with the preset power allocation trajectory based on trajectory deviations, and marking power allocation trajectories with abnormal deviations; Deviation points are identified on the power distribution trajectory of the abnormal deviation, and deviation accumulation detection is performed to extract continuous deviation accumulation areas.

6. The power distribution fault prediction method based on BMS data analysis and drive according to claim 1 is characterized in that: The specific steps of mining the power distribution path of the continuous deviation accumulation area based on the normalized power perturbation state diagram, performing operation state simulation self-learning optimization, and constructing the fault prediction optimization model are as follows: Locate the deviation cell based on the continuous deviation accumulation area; Performing a long-term power prediction analysis on the deviation battery cell to extract a power change curve of the deviation battery cell; Performing abnormal fault cell diagnosis on the power variation curve based on the normalized power perturbation state diagram, and marking the abnormal fault cell; Mining the power distribution path of abnormal fault cells to predict the power distribution path of the fault cells; The power distribution topology correlation monitoring is carried out on abnormal fault cells, and the operation state simulation self-learning optimization is performed to build a fault prediction optimization model.

7. The power distribution fault prediction method based on BMS data analysis and drive according to claim 1 is characterized in that: The specific steps of performing power distribution topology correlation monitoring on abnormal fault cells, performing operation state simulation self-learning optimization, and constructing a fault prediction optimization model are as follows: Perform power distribution topology correlation monitoring on abnormal fault cells and extract topology-related cells; Perform fault tracing analysis based on the power distribution path and topologically associated cells to identify the power distribution fault tracing point; Perform fault trend evolution analysis on the power distribution fault source point to obtain the fault trend evolution characteristics; Based on the fault trend evolution characteristics, operation state simulation self-learning optimization is carried out to build a fault prediction optimization model.

8. A power distribution fault prediction system based on BMS data analysis and drive, characterized in that: The method for predicting power distribution faults based on BMS data analysis as claimed in claim 1 comprises: The normalized power state module is used to call the BMS to collect power supply history monitoring logs, perform periodic power disturbance change analysis and normalized power perturbation evolution, and construct normalized power perturbation state diagrams for multiple battery cells; The personalized status module is used to mine the evolution of each battery cell's status based on the power supply historical monitoring log, and perform multi-dimensional personalized status perception to build a personalized status portrait for each battery cell; The power allocation rule module is used to calculate the requested power deviation of the battery cell, perform power allocation coordination analysis between the battery cells, and mine the power allocation game logic based on personalized status profiles to identify the power allocation game rules of the battery cells; An allocation trajectory deviation module is used to perform power allocation prediction and power allocation trajectory deviation comparison according to the power allocation game rule, and extract continuous deviation accumulation areas; The fault prediction optimization module is used to mine the power distribution path of the continuous deviation accumulation area based on the normalized power perturbation state diagram, and perform operation state simulation self-learning optimization to build a fault prediction optimization model.

Images