Lithium ion battery fault prediction method and system based on BMS

Abstract

The invention relates to the field of battery fault prediction, in particular to a lithium ion battery fault prediction method and system based on a BMS. The method comprises the following steps: extracting multi-dimensional operation monitoring parameters of a battery through a BMS (Battery Management System), carrying out multi-state evolution perception and label mapping processing, and constructing a global multi-state perception map of the battery; short-term abnormal sudden change detection is carried out according to the multi-dimensional operation monitoring parameters of the battery, and normal characteristic deviation trend analysis is carried out, so that an abnormal fluctuation deviation evolution trajectory is constructed; and performing deep topological correlation learning on the global multi-state sensing map of the battery based on the abnormal fluctuation deviation evolution trajectory, performing heterogeneous node global sensing, performing abnormal behavior causal relationship mining on heterogeneous deviation nodes in the battery, and performing multi-causal fission simulation to generate a battery behavior deterioration chain under an abnormal trend. According to the method, accurate and efficient fault prediction is realized, transfer learning is carried out, and the perspectiveness of subsequent BMS fault prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of battery fault prediction, and in particular to a lithium-ion battery fault prediction method and system based on a BMS. Background Art

[0002] With the rapid development of technologies such as new energy vehicles, portable electronic devices, smart grids, and energy storage systems, lithium-ion batteries, as electrochemical energy storage devices with high energy density, long cycle life, and no memory effect, have been widely used in various critical scenarios. In particular, in electric vehicles and large-scale energy storage systems, lithium-ion batteries not only play a core role in power supply but also directly affect the operational stability and safety of the entire system. Furthermore, the importance of battery management systems (BMS), as key components for monitoring, controlling, and protecting lithium batteries, has become increasingly prominent.

[0003] During actual operation, lithium-ion batteries are subject to complex chemical reaction mechanisms and fluctuating external operating conditions, and their performance and safety can change over time. This is especially true under conditions such as prolonged charge and discharge cycles, high or low temperature environments, and excessive charge and discharge. Batteries can experience capacity decay, increased internal resistance, abnormal temperatures, and voltage imbalances. In severe cases, they can even cause safety incidents such as thermal runaway, swelling, fire, or explosion. Furthermore, with the increasing accuracy of BMS data collection, battery operating data is becoming increasingly large-scale, high-frequency, and multi-dimensional. Extracting effective information from this complex monitoring data to identify and warn of potential battery failures has become a core challenge in current lithium battery safety management.

[0004] Existing battery fault monitoring and early warning methods primarily rely on empirically set thresholds and manual rule-based judgments, or employ linear models based on a small number of characteristic parameters for estimation. For example, these methods, such as single-parameter overtemperature protection and voltage anomaly detection, offer some practical advantages. However, when faced with multi-factor coupled, nonlinear battery performance degradation, these methods often exhibit delayed response, crude recognition, and weak predictive capabilities. This is particularly true when batteries are in a sub-healthy or early-stage degradation state. Traditional methods struggle to effectively identify hidden fault signs, leading to delayed fault warnings and increased system operational risks. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a lithium-ion battery fault prediction method and system based on BMS to solve at least one of the above technical problems.

[0006] To achieve the above object, the present invention provides a lithium-ion battery fault prediction method based on BMS, comprising the following steps: Step S1: Extract multi-dimensional battery operation monitoring parameters through the BMS, perform multi-state evolution perception and label mapping processing, and build a global multi-state perception map of the battery; Step S2: Perform short-term abnormal mutation detection based on the multi-dimensional operation monitoring parameters of the battery, and perform normal feature deviation trend analysis to construct the abnormal fluctuation deviation evolution trajectory; Step S3: Based on the abnormal fluctuation deviation evolution trajectory, deep topological association learning is performed on the global multi-state perception map of the battery, and global perception of heterogeneous nodes is performed to extract heterogeneous deviation nodes in the battery; Step S4: mining abnormal behavior causal relationships for heterogeneous deviation nodes in the battery, and performing multi-causal fission simulation to generate a battery behavior deterioration chain under abnormal trends; Step S5: performing abnormal behavior trend evolution on the battery behavior deterioration chain under the abnormal trend, and performing multi-point failure probability calculation to obtain an abnormal behavior failure probability curve; Step S6: Perform multi-dimensional fault behavior encapsulation based on the abnormal behavior fault probability curve, and then perform fault feature migration training to perform battery fault similarity identification and prediction.

[0007] In this specification, a lithium-ion battery fault prediction system based on a BMS is provided, which is used to execute the lithium-ion battery fault prediction method based on a BMS as described above, including: The multi-state perception module is used to extract multi-dimensional battery operation monitoring parameters through the BMS, perform multi-state evolution perception and label mapping, and build a global multi-state perception map of the battery; The mutation detection module is used to detect short-term abnormal mutations based on the multi-dimensional operation monitoring parameters of the battery, and analyze the deviation trend of normal characteristics, thereby constructing the evolution trajectory of abnormal fluctuation deviation; The abnormal similarity analysis module is used to perform deep topological association learning on the global multi-state perception map of the battery based on the abnormal fluctuation deviation evolution trajectory, and perform global perception of heterogeneous nodes to extract heterogeneous deviation nodes in the battery; The fission simulation module is used to mine the causal relationship of abnormal behavior of heterogeneous deviation nodes in the battery and perform multi-causal fission simulation to generate the battery behavior deterioration chain under abnormal trends; The fault probability calculation module is used to analyze the abnormal behavior trend evolution of the battery behavior deterioration chain under abnormal trends, and perform multi-point failure probability calculation to obtain the abnormal behavior failure probability curve; The fault transfer learning module is used to encapsulate multi-dimensional fault behaviors based on abnormal behavior fault probability curves, and then conduct fault feature transfer training to perform battery fault similarity identification and prediction.

[0008] The beneficial effects of the present invention are specifically as follows: through multi-dimensional panoramic modeling of battery status, covering multiple key parameters such as temperature, voltage, current, internal resistance, SOC, etc., the system's ability to recognize complex operating states is enhanced; discrete parameter states are structured through a label mapping mechanism, so that the battery operating state has a semantic hierarchy, which is helpful for subsequent models to perform semantic analysis and trend identification; a state evolution map of the battery's entire life cycle is formed, providing a unified map basis for anomaly detection and behavior prediction; the BMS system's ability to analyze complex interactions in multiple states is improved, avoiding false alarms or omissions caused by single-dimensional judgments; and a structured, high-resolution state substrate is provided for subsequent behavior deviation and trend evolution analysis. Effectively identify tiny abnormal fluctuations in the latent stage, and improve the ability to detect early degradation by capturing abnormal mutation behaviors in a short period of time; realize trend modeling of small deviations in the battery operating state, avoiding the abnormality drowning caused by averaging in traditional methods; the constructed abnormal fluctuation deviation evolution trajectory provides a quantitative representation of the potential degradation path of the battery, which is convenient for predicting its subsequent development direction; significantly improve the system's sensitivity to nonlinear failure characteristics, breaking through the previous limitation of only being able to capture large anomalies; build an anomaly detection mechanism that emphasizes both data-driven and knowledge-driven, and improve the generalization and adaptability of the model. Through topological learning, abnormal behavior and normal behavior are clearly separated in the structural space, enhancing the system's ability to discriminate deviant behavior; using graph neural networks or similar mechanisms to conduct horizontal learning of multiple battery states, the system has swarm intelligence and can discover "abnormal" batteries that are hidden in individuals but significant in the global picture; the identification of heterogeneous deviation nodes can realize early clustering screening to avoid local anomalies being ignored due to non-aggregation significance; significantly improve the resolution and robustness of anomaly identification, which is particularly suitable for battery batch management scenarios; construct a multi-scale behavior deviation map for multi-battery systems, effectively supporting abnormal correlation judgment under complex working conditions. Causal modeling is introduced on the basis of traditional data prediction to reveal the potential driving mechanism behind abnormal behavior, enabling the prediction to leap from correlation to mechanism; multi-causal fission simulation constructs multiple possible development paths to improve the coverage and robustness of fault trend prediction; the battery behavior deterioration chain provides a path-level evolution reference for actual prediction, which is conducive to accurately predicting when and why failure will occur; structured modeling of the abnormal evolution process is realized to provide a causal basis for subsequent decision-making (such as battery replacement, load reduction, and alarm); avoids the disadvantages of failure under complex conditions based solely on time series prediction, and enhances the system's adaptability to dynamic evolution.Each stage of the behavioral evolution process is mapped to a probability node for fault occurrence, enabling a quantitative assessment of potential risks. The generated fault probability curve can be used as a dynamic safety indicator for real-time dynamic adjustment of system strategies. Multi-point probabilistic modeling breaks the binary judgment of "fault or not" and implements multi-level risk warning, facilitating early intervention. It effectively supports the system's dynamic risk tolerance threshold setting under different strategies, achieving flexible fault management. It enhances the predictive decision-making capabilities of the BMS, shifting the system from fault detection to fault risk control. Fault behavior encapsulation enables the compressed expression of complex abnormal patterns, allowing the system to accurately represent multiple potential faults with a small number of parameters. Through a transfer training mechanism, identified fault features are fitted and identified on batteries that have not experienced faults, improving the forward-looking nature of predictions. Fault similarity recognition enables the prediction model to not only detect known problems, but also identify new faults that are "close to known problems." It supports learning from individual cases and generalizing to the population, improving model training efficiency and application breadth. It provides "experience memory" capabilities for smart battery systems, enabling brain-like learning and adaptation, and possessing the potential for long-term autonomous evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 A schematic flow chart of the steps of a lithium-ion battery fault prediction method based on BMS of 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

[0010] 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.

[0011] This application provides a BMS-based lithium-ion battery fault prediction method and system. The execution entities of the BMS-based lithium-ion battery fault prediction method and system 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.

[0012] See also Figures 1 to 4 The present invention provides a lithium-ion battery fault prediction method based on BMS, comprising the following steps: Step S1: Extract multi-dimensional battery operation monitoring parameters through the BMS, perform multi-state evolution perception and label mapping processing, and build a global multi-state perception map of the battery; Step S2: Perform short-term abnormal mutation detection based on the multi-dimensional operation monitoring parameters of the battery, and perform normal feature deviation trend analysis to construct the abnormal fluctuation deviation evolution trajectory; Step S3: Based on the abnormal fluctuation deviation evolution trajectory, deep topological association learning is performed on the global multi-state perception map of the battery, and global perception of heterogeneous nodes is performed to extract heterogeneous deviation nodes in the battery; Step S4: mining abnormal behavior causal relationships for heterogeneous deviation nodes in the battery, and performing multi-causal fission simulation to generate a battery behavior deterioration chain under abnormal trends; Step S5: performing abnormal behavior trend evolution on the battery behavior deterioration chain under the abnormal trend, and performing multi-point failure probability calculation to obtain an abnormal behavior failure probability curve; Step S6: Perform multi-dimensional fault behavior encapsulation based on the abnormal behavior fault probability curve, and then perform fault feature migration training to perform battery fault similarity identification and prediction.

[0013] In the embodiment of the present invention, see Figure 1 , is a schematic flow chart of the steps of a BMS-based lithium-ion battery fault prediction method of the present invention. In this example, the steps of the BMS-based lithium-ion battery fault prediction method include: Step S1: Extract multi-dimensional battery operation monitoring parameters through the BMS, perform multi-state evolution perception and label mapping processing, and build a global multi-state perception map of the battery; In this embodiment, during the basic stage of battery fault prediction, the BMS (Battery Management System) first continuously extracts multi-dimensional monitoring parameter data during battery operation. These parameters primarily include, but are not limited to, voltage (V), current (I), temperature (T), internal resistance (R), state of charge (SOC), and remaining capacity (Capacity). The system sets a sampling interval of 10ms to 100ms, with a sampling frequency of at least 10Hz, to ensure accurate recording of dynamic changes in the battery's operation at different stages, under varying loads, and under varying temperature and humidity conditions. The acquired raw monitoring parameter data then enters a data preprocessing module for outlier removal, wavelet noise reduction, and linear interpolation to ensure data continuity and physical consistency. The system then employs a state evolution awareness algorithm to generate time series modeling for these processed parameters. Specifically, a sliding time window (e.g., a window length of 300s with a sliding step of 30s) is used to track and analyze battery state evolution trends. By normalizing key features such as voltage fluctuation, internal resistance growth rate, and SOC decline rate within each time window, the system can identify multiple typical battery states during operation, such as healthy charging state, mild aging state, and critical thermal runaway state. The system then assigns labels to each state category based on a multi-label classification mechanism. This label mapping process combines a rule engine with a historical sample learning mechanism to compare monitoring parameter combinations with known operating states. If, within a given window, the cell's internal resistance increases by more than 30%, the temperature rises continuously by more than 8°C / min, and the SOC decline rate deviates from the normal range, it is labeled "abnormal aging state." Finally, the system constructs a global multi-state perception map of the battery based on the state labels mapped across different time periods, modules, and cells. This map is presented as a multidimensional tensor, with the horizontal axis representing time nodes, the vertical axis representing cell number, and the depth dimension representing the activation level of each state label. This map can be used for subsequent abnormal trend identification, fault pattern mining, and evolution prediction, forming a crucial foundation for intelligent fault diagnosis.

[0014] Step S2: Perform short-term abnormal mutation detection based on the multi-dimensional operation monitoring parameters of the battery, and perform normal feature deviation trend analysis to construct the abnormal fluctuation deviation evolution trajectory; In this embodiment, during the mid-term perception phase of the battery fault prediction system, the collected multi-dimensional operating monitoring parameter data is subjected to short-term abnormal mutation detection and trend analysis of deviations from normal characteristics to promptly identify potential fault signs. First, the system uses a sliding time window to monitor key parameters such as voltage, current, temperature, internal resistance, and SOC in real time. In the experiment, the sliding window was set to 60 seconds with a step size of 10 seconds to ensure that the detection granularity captures both short-term mutations and critical minor fluctuations. Short-term mutation detection utilizes a combined identification mechanism using an improved CUSUM (Cumulative Sum Chart) method and the Z-score standard deviation method. For temperature, for example, if the temperature slope ΔT / Δt within a continuous window is greater than 0.5°C / min and the Z-score exceeds 3, an abnormal sudden increase or decrease trend is identified. This mechanism effectively monitors sudden thermal runaway, rapid discharge, or internal short circuit behavior. Similarly, when monitoring internal resistance, if a cell's resistance growth rate exceeds 0.2mΩ / min within adjacent time windows, it can be preliminarily determined to be experiencing capacity fade or contact failure. Secondly, the system incorporates a baseline comparison analysis method to analyze deviations from normal characteristics. This method uses data accumulated during healthy battery operation to construct a healthy baseline model (such as the average voltage curve and standard temperature fluctuation range), and then calculates the deviation from the currently collected data. If the SOC decline rate exceeds 1.5 times the baseline rate per unit time (e.g., every 10 minutes), it is marked as "deviant SOC behavior." If the internal resistance continuously fluctuates above the baseline and exhibits an increasing frequency of fluctuations, it is included in the "fluctuating impedance signature" sequence. Finally, the system combines all detected short-term mutation points and trend deviation points into a time series, and applies time series clustering and dynamic time warping (DTW) techniques to construct an evolutionary trajectory of abnormal fluctuation deviations. This evolutionary trajectory reflects the entire process of a battery's gradual deviation from a healthy state to a potentially abnormal state during a specific period of operation. This trajectory is presented in the graph as a progression from "steady-state smoothness" to "fluctuations" and finally to "trend deviation." This trajectory will serve as an important input for subsequent fault tendency assessment and abnormal pattern modeling.

[0015] Step S3: Based on the abnormal fluctuation deviation evolution trajectory, deep topological association learning is performed on the global multi-state perception map of the battery, and global perception of heterogeneous nodes is performed to extract heterogeneous deviation nodes in the battery; In this embodiment, after identifying short-term sudden changes and deviations from normal behavior, the system needs to further understand the dynamic behavioral differences of each cell within the battery system from a global perspective. Therefore, in step S3, we deeply integrate the "abnormal fluctuation deviation evolution trajectory" with the "global battery multi-state perception map" to achieve topological structured learning of the cell operating state and identify outlier nodes within it—cells or regions that exhibit significant deviations from the group state. First, the system treats the previously constructed battery multi-state perception map as an attribute graph, where each node represents the set of states of a specific cell within a certain time period, and edges represent correlations in operating behavior between different cells (such as temperature synchronization and voltage similarity). The feature vector of each node includes statistical characteristic values ​​(such as mean, variance, and slope) for dimensions such as voltage, temperature, SOC, internal resistance, and capacity. A sliding window strategy is used in the time dimension to encode dynamic behavior into a static graph input. Second, the system performs deep topological correlation learning on the perception map based on the Graph Neural Network (GNN) framework. In our experiments, we used the GraphSAGE architecture for training. It aggregates and propagates features within a graph, enabling the model to learn the state distribution pattern of each cell node within its topological neighborhood. Through multiple rounds of message propagation, the system can identify cells whose state change trajectories exhibit atypical behavior, either locally or globally. Even if a cell's temperature is within a safe range, if its state fluctuation pattern deviates significantly from that of its neighbors, it will be marked with a high weight. Subsequently, based on the trained GNN model, the system introduces an outlier node identification algorithm, primarily employing an isolation forest approach combined with graph outlier detection techniques (such as Deep Graph Infomax) to identify nodes with low-probability state distributions in the graph. For example, in one experiment, among nodes with a capacity below 40% and a temperature fluctuation exceeding ±5°C, cells with an anomaly score greater than 0.65 identified by the isolation forest model were included in the initial outlier screening pool. Ultimately, by analyzing the evolutionary trajectory of abnormal fluctuations in conjunction with the graph structure, the system can identify "abnormal deviation nodes" that are inconsistent with the global pattern and contradictory to the local topological behavior. These nodes represent cells that may be at high risk for future failures (such as thermal runaway, sudden capacity drop, and increased polarization), and will subsequently serve as core input for fault prediction and feature modeling.

[0016] Step S4: mining abnormal behavior causal relationships for heterogeneous deviation nodes in the battery, and performing multi-causal fission simulation to generate a battery behavior deterioration chain under abnormal trends; In this embodiment, after extracting heterogeneous deviation nodes from the battery system, it is necessary to further identify the causal relationships between these abnormal nodes during the fault formation process. To achieve a deeper understanding of the abnormal state evolution mechanism, step S4 uses causal inference methods combined with time series behavior fission simulation technology to construct a multi-path deterioration chain of abnormal battery operation behavior from its starting point to its end point, namely the "battery behavior deterioration chain." First, the system uses the multi-dimensional state changes of each heterogeneous deviation node as a time series input to mine the causal relationships of the abnormal behavior. The causal relationship modeling method employed here is a hybrid approach based on the Granger causality test and the structural causal model (SCM). Granger causality primarily identifies whether a certain indicator (such as internal resistance increase) can predict another indicator (such as temperature change) in the time dimension. The SCM model implements structural causal modeling between variables by constructing a directed acyclic graph (DAG). It was found in multiple test batteries that when the voltage fluctuation rate of a certain battery cell exceeds the threshold of ±0.15 V / min, its temperature rise rate within 10 minutes often exceeds 1.5°C. Such a relationship will be confirmed as a potential causal chain. Subsequently, the system performs "multi-causal fission simulation" based on the identified causal pairs. This process simulates the state evolution path of the battery system under different disturbance triggering conditions to identify node fission that may cause further state imbalance. The system will simulate whether a sudden drop in the SOC of a certain battery cell will cause the voltage of the adjacent battery cells to drop synchronously, whether it will cause an imbalance in the overall current scheduling, and ultimately cause heat diffusion and other phenomena. During the simulation process, a dynamic causal simulator based on a Bayesian network is used, combined with real monitoring data for training and deduction to ensure that the simulation results are physically reasonable. In a specific experimental environment, typical operating conditions, such as the temperature rise process under 0.5C constant current charging mode, were selected as the baseline background. By performing causal fission expansion on the operating trajectories of multiple battery cells with mild abnormal deviations (such as internal resistance increase >15% and capacity decrease >10%), the system generated multiple deterioration paths, including "voltage fluctuation → internal resistance increase → local temperature rise → SOC drift → system load transfer failure." Each path is represented as a behavioral chain with a probability weight, forming the final "battery behavior deterioration chain" database. Through the above-mentioned causal mining and simulated fission process, the BMS system can perceive the potential failure trends that are forming in the battery in advance, providing a detailed behavioral path basis for subsequent fault prediction and active safety management. This stage is a key bridge for achieving the transition from "abnormal phenomenon" to "failure mechanism."

[0017] Step S5: performing abnormal behavior trend evolution on the battery behavior deterioration chain under the abnormal trend, and performing multi-point failure probability calculation to obtain an abnormal behavior failure probability curve; In this embodiment, after constructing the abnormal behavior deterioration chain, the chain is dynamically analyzed to predict future abnormal behavior trends and accurately calculate multi-point failure probabilities, ultimately forming an "abnormal behavior failure probability curve" that evolves over time. This curve can be used to assess the likelihood of battery failure at different time points and provides critical data support for precise battery health management and proactive intervention. First, based on the constructed deterioration chain, the system uses a time-series-based evolutionary modeling approach to model the future evolution of each abnormal path. Specifically, a long short-term memory network (LSTM) or gated recurrent unit (GRU) model is used to predict the states of nodes in the deterioration chain. Training data is derived from high-frequency operating monitoring parameters recorded during actual test conditions, such as internal resistance, temperature, SOC, and voltage fluctuations. For example, under 0.8C discharge conditions, if a battery cell experiences a continuous voltage drop of 0.1V and a temperature rise of more than 3°C over 20 minutes, the system identifies this node as an evolution focus and simulates the probability of thermal runaway after 30 minutes. Next, the system calculates the probability of multiple failure points at each key node along each evolutionary path. The probabilistic modeling approach employed here is a Dynamic Bayesian Network (DBN), which dynamically updates transition probabilities between different states over time. This approach allows the system to update the probability weights for each state in real time and, by integrating dependencies between upstream and downstream nodes, achieve moment-by-moment probability estimation. If a dependency relationship exists between an abnormal increase in internal resistance and a sudden SOC change, a directed edge is established within the DBN structure, and the failure risk of the voltage decay node is corrected during the inference process. Through multi-path and multi-node evolution and probability calculation, an "abnormal behavior failure probability curve" is ultimately generated. This curve, with time as the horizontal axis and failure probability as the vertical axis, illustrates how the battery failure risk evolves from the current moment to several minutes or hours into the future. In experimental validation, the system generated corresponding failure probability curves for a set of battery cells operating under different temperature and humidity conditions (e.g., 25°C at room temperature and 45°C at high temperature). The results show that in high-temperature environments, the peak failure probability of certain nodes with critical internal resistance increases by 15 minutes earlier, with the maximum probability increasing by over 20%. Through this process, the BMS system can not only perceive the current abnormal state but also provide quantitative predictions of future failure risks. The abnormal behavior failure probability curve serves as the basis for subsequent fault warning, control strategy optimization, and energy scheduling adjustments. It builds a barrier for early prediction and precise control capabilities for intelligent battery systems and is a core step in future intelligent battery safety management.

[0018] Step S6: Perform multi-dimensional fault behavior encapsulation based on the abnormal behavior fault probability curve, and then perform fault feature migration training to perform battery fault similarity identification and prediction.

[0019] In this embodiment, based on the construction of the abnormal behavior failure probability curve, the fault characteristics of the battery under multi-dimensional parameters are further structured and vectorized to construct a standardized fault behavior representation system. The model is trained and generalized through the transfer learning mechanism, so that the BMS has the ability to identify similar patterns of potential faults in unknown scenarios and realize the prediction and identification of fault similarity across cells and batches. First, the system extracts the time points of key fault nodes from the abnormal behavior failure probability curve and retrieves the full-dimensional operation monitoring data at these moments, including multiple feature items such as voltage decay rate, internal resistance rise slope, instantaneous temperature fluctuation amplitude, SOC imbalance value, charge and discharge rate change trend, etc. Taking typical experimental data as an example, a certain battery cell has an internal resistance step change of 0.15Ω in the 55th minute under the constant temperature condition of 20℃ in the laboratory, accompanied by a continuous drop in voltage to below 3.1V and a temperature rise to 45℃. The system encapsulates this data as a "high internal resistance thermal evolution type" fault behavior and constructs a multi-dimensional feature vector: Next, the system normalizes the feature vectors of multiple similar fault behaviors and performs visual cluster analysis using PCA (Principal Component Analysis) or t-SNE dimensionality reduction methods to identify the distribution relationships of various fault modes in the feature space. By constructing a feature mapping function, the fault behaviors of different battery cells are uniformly mapped into a low-dimensional space, enhancing the model's ability to identify common features across samples. Next, the system enters the transfer learning training phase. This uses a deep neural network (DNN)-based transfer framework, specifically a domain-adaptive neural network (DANN) or a time series transfer model based on a modified BERT architecture. In the pre-training phase, the system trains the model using feature vectors of known faults from an annotated dataset, enabling the network to learn universal fault discrimination features. In the transfer phase, the system uses data from a new environment or a new batch of battery cells as target domain input. Through adversarial training or minimizing a domain-differential loss function, the system seamlessly transfers feature knowledge from the source domain to the target domain. Finally, the system uses the trained transfer model to identify similar fault features in live battery cells. The specific method is to calculate the cosine similarity or Mahalanobis distance between the current cell's real-time multidimensional feature vector and historical samples in the fault signature library to determine which historical faults the cell's current state is most similar to. Under 0.5C discharge conditions, if a cell exhibits behavior that is more than 87% similar to a "high-temperature rapid decay" fault, the system will trigger a fault warning and recommend appropriate control strategies.

[0020] In this embodiment, refer to Figure 2 , is a flowchart of the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Setting a high-frequency sampling time window; performing millisecond-level multi-dimensional data acquisition through the BMS based on the high-frequency sampling time window to extract multi-dimensional battery operation monitoring parameters, including voltage, current, temperature, internal resistance, state of charge (SOC), and capacity; Adaptively filtering the multi-dimensional operating monitoring parameters of the battery and correcting sensor acquisition deviations to obtain deviation-optimized monitoring parameters; Identify the current battery operating scenario based on the BMS; analyze and divide the time periods based on the battery operating scenario to obtain time periods of equal length; Performing parameter segmentation processing on the deviation optimization monitoring parameters according to the equal-length time periods to extract monitoring parameters for multiple time periods; The multi-state evolution perception of monitoring parameters in multiple time periods is performed, and label mapping processing is performed to construct a global multi-state perception map of the battery.

[0021] In this embodiment, to achieve high-precision monitoring of the lithium-ion battery's operating status, a high-frequency sampling time window must first be set. Traditional BMS systems typically use a sampling frequency of 1 Hz or lower, which is insufficient for fault detection under dynamic conditions and makes it difficult to capture short-term, sudden changes in behavior. A time window in milliseconds is set, with a sampling frequency of 10 ms (i.e., 100 Hz) or higher, depending on hardware capabilities and communication bandwidth. Within this time window, the BMS must collect multiple key operating parameters from the battery, including but not limited to: cell voltage (V), charge and discharge current (A), battery surface and core temperature (°C), equivalent internal resistance (mΩ), state of charge (SOC) (%), and current battery capacity (Ah). This acquisition process relies on the BMS's integrated high-precision sensor modules and ADC conversion circuits, which offer fast response and noise immunity. Voltage acquisition accuracy must be within ±5 mV, current error within ±0.1 A, and temperature accuracy better than ±0.5°C. To support high-frequency sampling, the data acquisition architecture needs to be optimized. It is recommended to use DMA (Direct Memory Access) for cached reads to reduce the load on the main MCU. High-frequency data can also be uploaded to a host computer or edge processing unit via the CAN-FD bus or Ethernet. Furthermore, to ensure acquisition synchronization, all channels must be calibrated and time-aligned to avoid multidimensional parameter mismatches caused by time drift. This process provides a solid data foundation for subsequent data filtering, state identification, and fault analysis. Due to limitations in sensor accuracy, communication noise, and environmental interference, the raw acquired data may contain significant random errors and systematic biases. Therefore, in this step, adaptive filtering is required to correct for biases introduced by various sensors to improve the accuracy of subsequent model analysis. First, a weighted moving average (WMA) or adaptive Kalman filter (AKF) algorithm is applied to denoise the multidimensional data. Kalman filtering, due to its state estimation capabilities, is particularly suitable for time series data. It can dynamically adjust the noise covariance matrix under different operating conditions to improve filtering performance. To address spike disturbances in the current signal, the observation error covariance R can be adaptively adjusted within the range of 0.1–0.3, while the state prediction error covariance Q is set to 0.01–0.05. For voltage sampling, to address bias caused by ADC nonlinearity, an error correction function is established through offline calibration and interpolation mapping. Furthermore, a dynamic threshold algorithm based on extreme value identification is introduced to eliminate obvious erroneous values, such as negative temperature values ​​and sudden voltage changes.

[0022] Furthermore, to further enhance data reliability, cross-validation of data from different sensors is necessary. Temperature data can be collected simultaneously using a thermistor and an infrared temperature sensor. By comparing the trends of these two data sources, it is possible to determine which data source exhibits deviation. After the deviations are corrected, all parameters are normalized to facilitate subsequent model analysis and state identification. The resulting deviation-optimized monitoring parameters exhibit greater temporal continuity and physical consistency, serving as the core input for building state-aware models. The operating state of lithium-ion batteries is highly dependent on operating conditions. Different operating scenarios (such as static, charging, constant current discharge, and pulsed load) trigger distinct electrochemical responses and fault manifestations. Therefore, this step analyzes the deviation-optimized monitoring parameters to identify the current operating scenario of the battery and divides the analysis time periods accordingly. Operating scenario identification typically utilizes state machine logic or decision tree-based rule classification methods. For example, a current threshold within ±0.5 A is defined as static, a current continuously exceeding 1 A is defined as discharge, and a voltage rise accompanied by a positive current is defined as charging. To improve accuracy, machine learning classification models such as support vector machines (SVMs) or random forests can be introduced to perform label classification training on time windows in the sampling sequence. Model inputs include first-order differential features such as the voltage, current, and temperature change rate during the current period. Based on the identified operating scenarios, the data stream is divided into multiple fixed-length time segments (e.g., 5 or 10 seconds each, depending on BMS performance and battery response speed). This segmentation avoids the introduction of inconsistent temporal features across different scenarios and facilitates stable modeling of the state vector. Using a 10-second segment as a segment, 360 high-dimensional data segments are generated per hour, each containing multiple instantaneous or averaged statistical values, laying the foundation for state modeling and evolution analysis. Each segment is also accompanied by a scenario label to facilitate subsequent differentiated modeling based on operating conditions. Based on the time segment division, this step performs structured slicing of the multi-dimensional parameters after deviation optimization into time segments of equal length. Each time segment is treated as an independent "state segment," from which the numerical features and changing trends of key operating parameters are extracted through statistical analysis and feature extraction methods. Taking a 10-second segment with a 100 Hz sampling frequency as an example, each segment will contain 1,000 data points. For each parameter (voltage, current, temperature, etc.), statistical features (mean, maximum, minimum, standard deviation), variation characteristics (slope, first-order difference), and frequency domain features (FFT energy, frequency band amplitude) can be extracted. For current signals, Fast Fourier Transform (FFT) can be used to identify periodic fluctuations to determine whether there are high-frequency oscillations or load fluctuations. For temperature signals, the slope and temperature rise rate are focused on to determine thermal anomaly trends. All features can be combined into a high-dimensional vector. For example, if five statistical features are extracted for each of the six original parameters, the feature dimension of a state segment will be 30.To ensure feature comparability, all paragraph feature vectors are standardized (such as Z-score standardization).

[0023] This feature slicing process creates a time-series dataset, where each segment contains both operating context and physical characteristics, as well as fault-sensitive evolutionary clues. This serves as a crucial input for building dynamic prediction models. Parameters from all time periods are aggregated into a multi-segment time series matrix to support subsequent state-aware evolution modeling. A global state evolution map of the lithium-ion battery is constructed from the segmented monitoring data, enabling comprehensive temporal and spatial understanding of its operating trends, health degradation, and abnormal behavior. First, time series clustering algorithms (such as DTW-KMeans and HDBSCAN) are used to cluster parameters from multiple time periods to identify different state categories. Each category represents an operating state, such as healthy discharge, slow charge, fast charge, or high temperature anomaly. Neural network models such as autoencoders or transformers can then be introduced to reduce the dimensionality of the high-dimensional state vector and perform time series modeling to achieve more refined state identification. To enhance the interpretability and visualization of state evolution, all identified states are organized into a directed graph structure. The graph's nodes represent battery states at different time periods, while edges represent state transitions. Edge weights can represent transition probabilities or the intensity of state changes. By constructing a Markov transition matrix or state transition diagram, the battery's evolution path from normal to abnormal states can be analyzed, identifying typical fault warning patterns. If a state sequence of "high current + rising temperature + rising internal resistance" often evolves to a "rapid capacity decline" state after three consecutive segments, this can be defined as a fault precursor pattern and labeled. Labels are included for both the current state type (e.g., "abnormal charging") and its risk level (e.g., "fault warning level"). Ultimately, all data segments, state types, label mapping results, and state evolution paths together form a "global battery multi-state perception graph." This graph can be updated in real time and compared with a historical fault database for early warning, lifespan prediction, and fault classification decisions. It is the core output of the BMS-based intelligent fault prediction system.

[0024] In this embodiment, the specific steps of performing multi-state evolution perception on monitoring parameters in multiple time periods and performing label mapping processing to construct a global multi-state perception map of the battery are as follows: Extract nonlinear features of monitoring parameters in multiple time periods one by one to identify sudden increases and decreases in battery performance, nonlinear jitter, and pre-critical oscillation characteristics, thereby obtaining the battery status characteristics in different time periods. Performing deep state analysis on the state features and generating state labels; Performing multi-time period fluctuation mining on the state characteristics to obtain a fluctuation trend of the battery state characteristics; Based on the state labels, the multi-state evolution perception of the battery state characteristic fluctuation trend is carried out, and label mapping processing is performed to construct a multi-state perception map.

[0025] In this embodiment, during the operation of the lithium-ion battery, its state variables (such as voltage, current, temperature, internal resistance, etc.) may exhibit certain nonlinear change characteristics, especially before a fault or abnormal state. These nonlinear characteristics are difficult to fully identify through linear statistical parameters such as mean and variance. Nonlinear feature extraction is performed for each time period to identify sudden increases and decreases (such as voltage drops), short-term nonlinear jitter (such as internal resistance fluctuations at the millisecond level), and potential pre-critical oscillation phenomena (such as SOC oscillation rebound). The processing flow is based on the analysis of the data of the cut time period. Each time period is about 10 seconds, the sampling frequency is 100 Hz, and contains 1000 sets of data.

[0026] Using differencing and sliding window functions, the first- and second-order derivatives of the time series are analyzed for trends, and volatility, nonlinearity (e.g., based on fractal dimension, Lyapunov exponents), and mutation indicators (e.g., CUSUM detection) are calculated. For voltage signals, the per-second slope, peak-to-peak value, kurtosis, and skewness are extracted as indicators of nonstationarity. Sudden spikes and dips are detected by setting thresholds. If the voltage drops by more than 50mV within 10ms or the internal resistance jumps by 20% within 500ms, a sudden state is initially identified. To further identify potential pre-critical states (pre-failure oscillations), empirical mode decomposition (EMD) or wavelet transforms can be applied to perform multi-scale signal analysis to extract high-frequency jitter components and identify areas of abnormal high-frequency energy concentration. This step generates a state feature vector for each time period. This vector contains multiple nonlinear features, characterizing the battery's health status and potential abnormalities during that period, and serves as an important input for fault identification models. After state feature extraction, a systematic and in-depth analysis is required to accurately determine and classify the battery's state of health. This step combines statistical and deep learning methods to identify state evolution patterns through multidimensional feature modeling and generate state labels based on these patterns. First, the nonlinear feature vectors for all time periods are normalized and dimensionality reduced (e.g., using PCA or t-SNE to reduce the dimensionality to two or three dimensions) to facilitate pattern recognition and visualization analysis. Next, clustering algorithms such as Gaussian mixture models (GMMs), DBSCAN, or spectral clustering are used to perform unsupervised classification of the states for all time periods, identifying natural state groupings such as "steady-state operation," "minor anomalies," "violent fluctuations," and "short-term mutations." These labels provide a preliminary representation of the battery's current state of health. Simultaneously, a supervised learning-based classification network (such as a deep neural network (DNN) or a one-dimensional convolutional neural network (1D-CNN)) can be introduced to train historically labeled samples and automatically label unknown state segments, forming an "intelligent state labeling system." In the experimental setup, "violent fluctuations" in historical data are typically accompanied by high internal resistance increases (greater than 10 mΩ) and temperature rise rates exceeding 0.1°C / s. "Minor abnormalities" are characterized by voltage perturbations less than 50 mV, with jitter cycles concentrated in the 1–5 Hz range. A multidimensional labeling system creates four levels of status labels: "normal-abnormal-warning-fault," with each segment of data mapped to a label category. These labels not only represent the current health status but also serve as foundational information for subsequent fluctuation trend mining and state evolution identification. Through cross-timeframe analysis, long-term evolutionary trends in battery status characteristics can be identified, particularly those dynamic processes that progress from minor fluctuations to severe abnormalities.First, a time series is constructed based on the state feature vectors and corresponding state labels generated for each time period in the previous step. A sliding window mechanism (for example, covering 10 segments with a total duration of 100 seconds) is used to continuously extract local sequences and calculate the changing trends of the state features along the time axis. State trend mining focuses on the following indicators: the rate of change of feature mean (Δμ), variance volatility (Δσ²), the frequency of state label changes (state jump frequency), and multi-dimensional state co-volatility (e.g., current fluctuations accompanied by synchronous changes in temperature rise and internal resistance). Time series modeling methods such as moving autoregressive (AR) and long short-term memory (LSTM) networks can be used to model the state series and predict its trend. Experiments have found that batteries often experience an increase in the frequency of fluctuations in multiple state indicators before failure. For example, 120 minutes before failure, the internal resistance increase rate increases from 0.2 mΩ per hour to 0.5 mΩ, and the SOC estimation residual jumps from 0.5% to 1.5%. Once these fluctuation trends form a specific pattern, they can serve as an important basis for fault prediction.

[0027] Ultimately, this step outputs a set of time series trend features, including increasing amplitude and frequency trends, and label transition frequencies. These features depict the battery's evolutionary path from "stable" to "abnormal" or "warning," providing dynamic input for the state evolution graph. After obtaining the state labels and feature fluctuation trends, the evolutionary process between states over multiple time periods is structured and modeled to construct a multi-state battery perception graph. This graph not only records static state characteristics but also describes the transition paths, transition intensities, and risk levels between states, serving as a core tool for fault prediction and health management. First, a state transition matrix is ​​constructed using the state label sequence, defining the transition probability between each state. The transition frequency from "normal" to "minor abnormality" is p1, and from "abnormal" to "warning" is p2. An evolution perception mechanism is then introduced: using a Bayesian network or Markov chain to analyze the stability and criticality of state transition paths, identifying those paths as "high-risk evolutionary paths." The graph structure is a directed graph, with nodes representing different states and edges representing transition paths with causal evolutionary relationships in time or features. A slight voltage drop → increased temperature fluctuations → a sudden increase in internal resistance → entering the "warning state." Edge weights are calculated based on the transition probability, average evolution time interval, and risk factor calculated from historical samples. Label mapping processes state classification and risk tagging based on the graph. If the current path appears in 80% of historical fault evolutions, the node is mapped to the "pre-fault state." If the state exhibits stable fluctuations across multiple evolution paths, it is marked as the "operational alert state." The resulting "multi-state perception graph" can be dynamically updated and visualized graphically. Combined with real-time BMS data, it enables fault trend tracking and risk warning. This graph not only reflects the current battery status but also provides a scientific basis for operation and maintenance decisions, maintenance scheduling, and lifespan prediction.

[0028] In this embodiment, refer to Figure 3 , is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Perform short-term abnormal mutation detection based on multi-dimensional battery operation monitoring parameters and extract abnormal mutation characteristics; Based on the preset standard battery behavior database, the normal fluctuation range of abnormal mutation characteristics is judged. If it is determined to be an abnormal fluctuation beyond the range, it is marked as an abnormal fluctuation event; Calculating the frequency of occurrence, deviation amplitude and abnormal duration of the abnormal fluctuation event; Multidimensional cluster analysis is performed on the occurrence frequency, deviation amplitude and abnormal duration to obtain the statistical characteristics of short-term mutation behavior; The normal characteristic deviation trend of the statistical characteristics of short-term mutation behavior is analyzed and dynamically evolved to construct the abnormal fluctuation deviation evolution trajectory.

[0029] In this embodiment, lithium-ion batteries often exhibit short-term, abnormal, and sudden changes in behavior during operation, particularly during changes in charge and discharge loads, ambient temperature disturbances, or the early stages of aging. These changes can include voltage drops, sudden increases in internal resistance, or unusual temperature jumps. To identify these abnormal behaviors, this step establishes a detection mechanism based on high-frequency, multi-dimensional parameters (such as voltage, current, temperature, internal resistance, and SOC) collected by the BMS. The detection process utilizes a combination of time-domain difference (TDD) and statistical mutation detection. First, a sliding window is constructed with a 10 ms time resolution to measure the first- and second-order differences of each parameter. Dynamic thresholds are then calculated based on historical baseline fluctuations. A voltage drop exceeding 30 mV or a temperature change rate exceeding 0.2°C / s over three consecutive sampling points is initially defined as an "abnormal mutation point." Subsequently, an outlier aggregation strategy, such as the density-based temporal clustering algorithm based on adjacent windows, is used to aggregate multiple single-point anomalies into a single short-term mutation event for subsequent analysis. In actual experiments, the collected current signal fluctuates violently under high-power pulse loads. Short-term surges can reach up to 20 A, making it difficult to identify anomalies using conventional averaging methods. Therefore, the Z-score method is used to normalize each point, setting a threshold of ±3σ as the mutation identification limit to effectively detect abnormal fluctuation points. The final output of this step is a number of events labeled "short-term mutations," along with parameters such as the mutation occurrence time, fluctuation type, initial value, and change amount, which form the basic features for subsequent analysis. To distinguish natural operating fluctuations from true abnormal mutations, a "standard battery behavior database" is constructed as a comparison benchmark. This database consists of multiple sets of experimental data from healthy lithium batteries operating under typical operating conditions (constant current charging, constant current-constant voltage discharge, and static conditions). It covers the typical parameter fluctuation ranges across different temperature, rate, and SOC ranges. At 25°C and 1C charging, the normal cell voltage fluctuation range is ±20 mV, the temperature rise rate is 0.05–0.1°C / s, and the internal resistance fluctuation range is ±5 mΩ. After identifying short-term mutations, their corresponding characteristics, such as amplitude, rate, and duration, are compared item by item with the upper limits of fluctuations in a standard database. If any parameter exceeds the maximum value of normal fluctuations, the mutation is labeled an "abnormal fluctuation event." This comparison is not limited to absolute values; relative changes (such as ΔV / V%, ΔT / Δt) and the rate of change must also be considered. To improve the accuracy of judgments, a fuzzy judgment mechanism based on dynamic thresholds can be introduced, such as determining whether the mutation amplitude exceeds 1.5 times the upper limit of healthy samples and persists for more than a preset time threshold (such as 500 ms).

[0030] In a set of experimental data, the temperature fluctuation range under normal operating conditions is ±1°C. However, a sudden temperature increase of 3°C within 3 seconds significantly exceeds the boundaries defined in the normal behavior database and is clearly labeled as an "abnormal temperature rise event." This identification mechanism effectively filters out false positives, such as short-term jumps caused by load fluctuations, which are not fault indicators. Ultimately, the "abnormal fluctuation event" output from this step serves as an important label for abnormal battery behavior and is used for subsequent frequency and evolution analysis. After identifying abnormal fluctuation events, these events need to be further quantitatively analyzed to quantify the occurrence patterns and severity of short-term abnormal battery behavior. Specifically, for each type of abnormal fluctuation (such as voltage sag, temperature surge, and internal resistance surge), the frequency (times / hour), average deviation (e.g., ΔV of 75 mV), and duration of each event (e.g., exceeding 0.8 seconds) within a set time window are counted. The calculation process uses a fixed time period (e.g., 10 minutes, 30 minutes, or 1 hour) as a benchmark, and categorizes and counts abnormal fluctuation events recorded within each period. Within one hour, six temperature surge events occurred, with an average temperature rise rate of 0.25°C / s and a maximum rate of 0.4°C / s. The average amplitude of the voltage dip events was 60 mV, and the maximum duration was 2.2 seconds. These statistical parameters serve as key indicators of battery operational robustness, indicating whether the battery system exhibits significant instability. Furthermore, to enhance the accuracy of abnormal behavior assessment, it is recommended to incorporate an Anomaly Severity Index (ASI). This index combines the three aforementioned factors and uses a weighted scoring model. Frequency is weighted 0.4, amplitude is weighted 0.3, and duration is weighted 0.3, ultimately yielding a comprehensive severity score for abnormal fluctuations at a given stage. This quantitative score facilitates horizontal comparisons between different batteries and longitudinal analysis of historical trends. A multidimensional clustering method is used to classify and model the statistical characteristics of abnormal fluctuation behavior, identifying different types of abnormal mutation patterns. Using a three-dimensional vector of frequency, deviation amplitude, and duration as input features, each time period (e.g., every 30 minutes) is considered a sample for clustering. Common clustering algorithms include K-means, DBSCAN, or Gaussian mixture models (GMM). K-means is suitable for processing large datasets, while GMM is suitable for mining state subgroups with overlapping distributions. In experiments, the abnormal statistical features of the battery every 30 minutes over 6 hours were input into the K-means model (with the number of clusters K=3). Three types of abnormal behavior patterns were identified: low-frequency mild fluctuation (Type A), medium-frequency moderate fluctuation (Type B), and high-frequency strong mutation (Type C). Type C anomalies are more likely to occur during high-rate discharge or when the SOC is below 20%, and are more likely to indicate failure.After clustering, the statistical centroid of each cluster can be analyzed to construct a typical abnormal behavior profile. Type A features a frequency of <2 events / 30 minutes, a deviation amplitude of <40 mV, and a duration of <1 second; while Type C features a frequency of >5 events, a deviation amplitude of >80 mV, and a duration of >2 seconds. This clustering process compresses complex, sudden changes in behavior into a finite set of states, facilitating subsequent state evolution modeling. The final output is a set of cluster labels and corresponding behavioral pattern descriptions, providing a classification basis for abnormal behavior trajectory modeling and input state features for risk warning models. The statistical features of the clustered abnormal behavior are dynamically tracked and analyzed for deviations over time to identify any continuous evolution from "normal characteristics" to "abnormal trajectories." This analysis is similar to modeling health degradation curves, but focuses on the evolutionary trajectory of "abnormal fluctuating behavior."

[0031] First, a baseline for "normal characteristics" is defined: the mean and standard deviation range of each statistical characteristic during a fault-free or non-abnormal phase. Under normal conditions, the frequency of sudden events is less than 2 per 30 minutes, the deviation amplitude is less than 30 mV, and the duration is less than 1 second. Z-score normalization is performed on the characteristics during actual operation, and the degree of deviation of the current characteristic value from the baseline (Z deviation) is calculated. If the Z deviation exceeds a threshold (e.g., Z > 2.5) for three consecutive time periods, the system is marked as entering an "increasing deviation phase." Next, dynamic time warping (DTW) and sliding window analysis are used to track the trajectory of the characteristic trajectory, generating an "abnormal fluctuation deviation evolution curve." This curve depicts the temporal evolution from healthy operation to initial deviation to obvious abnormality. Evolution stage identification models can be further developed, such as "initial deviation phase," "accelerated evolution phase," and "out-of-control phase." In a real-world case study, sudden events in a battery increased from one to six per hour within five hours, with the deviation amplitude increasing from 20 mV to 90 mV. The evolution curve exhibits an exponential growth trend. Based on this curve, the system can issue early warning signals, indicating possible internal short circuits, contact anomalies, or overheating. Ultimately, the abnormal fluctuations generated by this step deviate from the evolution trajectory and serve as a key reference for battery status assessment and fault prediction. This information, integrated into the battery's multi-state perception map, becomes a crucial basis for full lifecycle monitoring and decision-making.

[0032] In this embodiment, reference Figure 4 The above is a flowchart of the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Position and identify abnormal battery nodes based on abnormal mutation characteristics; Based on the abnormal fluctuation deviation evolution trajectory, the battery global multi-state perception map is analyzed for the similarity of the behavior of all battery nodes, and the battery nodes with similar abnormal fluctuations are identified; Identify the differences and synchronization of abnormal battery nodes and battery nodes with similar abnormal fluctuations; Performing deep topological association learning on the differences and synchronizations to extract the group topological mapping relationship of the battery nodes; Based on the group topology mapping relationship and the abnormal fluctuation deviation evolution trajectory, global perception of heterogeneous nodes is performed to extract heterogeneous deviation nodes in the battery.

[0033] In this embodiment, in a battery system, particularly a power battery pack or energy storage system, each battery node in a battery array composed of multiple cells may exhibit independent mutational behavior due to manufacturing variations, thermal imbalance, or varying degrees of aging. Based on the "abnormal mutation signature" identified in the previous step, the core task of this step is to conduct a horizontal comparison of all cells to identify those exhibiting abnormal mutational behavior within a specific time window, namely "abnormal battery nodes." Specifically, within each cycle (e.g., 10 minutes), three metrics—frequency, magnitude, and duration—are counted for all cells to construct a mutational signature vector for each node. Nodes with higher mutation severity are flagged by setting thresholds (e.g., frequency exceeding 3 events / 10 minutes, magnitude exceeding 80 mV, duration exceeding 1.5 seconds) or by using statistical deviation (Z-score > 2.5). In the experiment, in a system consisting of 100 battery strings, 95% of the nodes experienced an average mutation frequency of 1.2 times per 10 minutes. However, node #27 experienced six mutations, with a maximum internal resistance variation of 18 mΩ, nearly six times the average of 3.2 mΩ for all other nodes. This node was subsequently labeled an "abnormal battery node." The BMS system performed real-time online calculations and generated a fault risk report based on the location number and sensor layout. Ultimately, individual batteries whose mutation signals significantly deviated from the statistical patterns of the group were located, providing a direct basis for subsequent fault risk control and replacement decisions. After locating a single abnormal node, this step aims to globally analyze whether other battery nodes exhibit similar behavioral patterns and similar evolutionary trajectories. The goal is to identify "potential abnormal nodes" with a common tendency toward operational instability, enabling early warning and group health analysis. This analysis relies on the previously constructed "abnormal fluctuation deviation evolution trajectory" and "battery multi-state perception map." First, a state evolution vector sequence was constructed for each battery node. Core variables included the mutation frequency change rate, fluctuation intensity slope, state jump frequency, and label transition path, forming a trajectory vector. Nodes #15 and #27 saw their frequency increase from 1 to 6 times within 30 minutes, with deviation amplitudes both exceeding 100 mV. Dynamic time warping (DTW), cosine similarity, or Euclidean distance were then used to measure the pairwise similarity between the trajectory sequences of all nodes. In the experimental setup, a similarity threshold of 0.85 was set. Four nodes (#15, #26, #38, and #59) were found to have highly similar behavioral trajectories to the abnormal node #27. Although not all of these nodes exceeded the mutation threshold, their evolutionary paths converged and they were labeled as "nodes with similar abnormal fluctuations." This type of analysis can effectively identify battery cells on the verge of abnormality or those that have not yet experienced failure, providing technical support for introducing the "trend identification" dimension into BMS systems.After identifying a group of battery nodes exhibiting abnormal fluctuation trends, further analysis is needed to determine whether these nodes exhibit consistent state mutation rhythms and whether they respond simultaneously to external operating disturbances or system load changes. This analysis not only confirms whether the problem stems from a local physical defect but also determines whether the failure mode carries the risk of mass transmission. In this step, feature vector sequences are first extracted for all abnormal and similar nodes, including the mutation time point, amplitude, duration, and trend. Subsequently, methods such as mutual information analysis, covariance matrix analysis, and cross-correlation function (CCF) are used to measure the consistency of fluctuation responses across nodes. In one experimental batch, it was found that the starting points of voltage mutations for nodes #27 and #15 differed by no more than 100 ms across multiple charge and discharge cycles, and their temperature rise curves showed a synchronization of up to 0.92 (based on the Pearson correlation coefficient). This indicates that the two nodes respond in a highly consistent manner to changes in system load or heat conduction, demonstrating "behavioral synchronization." Conversely, while node #38 exhibits similar behavioral trajectories, its mutation phase lags by more than 2 seconds, and its response amplitude is only 60% of that of the primary abnormal node, resulting in "asynchronous but similar behavior." This analysis results are ultimately output as a "group behavior difference-synchronicity matrix" to guide the next steps in topological mapping and global perception map construction. After identifying the differences and synchronicities between abnormal behavior groups, a topological mapping relationship between battery nodes is constructed to reveal possible physical coupling, thermal conduction paths, load response chains, or controller synergy between nodes. This topological structure is not limited to geometric connections but also emphasizes the implicit connection between electrochemical behavior and dynamic response. This is achieved by constructing a topological learning model based on a graph neural network (GNN). The battery nodes are treated as nodes in the graph structure, and each node is characterized by its fluctuating behavior parameter vector. Edge weights are set based on the aforementioned synchronization and difference calculations. For example, node pairs with a mutual information value greater than 0.8 are assigned an edge weight of 1, and the remaining weights are normalized according to a similarity function. The training process uses an unsupervised or semi-supervised approach, taking as input the graph structure and the states of some anomaly-labeled nodes. The output is a group topology embedding vector for each node, describing its behavioral independence and connectivity within the entire system. In experiments using a 100-node GNN model, the post-training output embedding space exhibits a clear clustering trend: core anomaly nodes, synchronous anomaly nodes, and normal nodes form separable partitions in the embedding space.

[0034] The topological learning results can be viewed as a "behavioral correlation map" of the battery system. They not only reveal possible collective coordinated response mechanisms among nodes but also provide a structured cognitive basis for ultimately identifying "deviant nodes" in the system. After understanding anomalous nodes, similar behavioral groups, and topological connectivity within the battery system, it is necessary to identify "deviant nodes" that do not belong to any cluster center and whose trajectory evolution paths significantly deviate from those of the majority of nodes. These nodes may represent atypical faults, controller imbalances, cell inconsistencies, and other issues, and are therefore valuable for fault monitoring. This is achieved by fusing the topological embedding vector generated in the previous step with the characteristics of abnormal fluctuation trajectories to construct a high-dimensional behavior space. Node behavior is scored for anomaly using isolation forests, local outlier factors (LOFs), or density-based anomaly detection methods (such as one-class support vector machines). Nodes with scores above a set threshold (e.g., the 0.95th percentile) are considered "deviant nodes." In a certain energy storage system, although node #88 did not show significant mutations, its topological embedding characteristics were clearly isolated, and its behavioral evolution path was significantly different from all similar groups. After disassembly and inspection, it was found that the node had a slight controller communication delay problem, which affected its real-time response capability and, in turn, the energy management balance of the entire package. The core value of this step lies in the "heterogeneity identification" of abnormal behavior. Different from group consistency risk detection, it provides a data-driven intelligent perception mechanism for the identification of atypical hidden dangers in complex systems. All heterogeneous nodes will be added to the battery multi-state perception map for dynamic updates, intelligent warnings, and maintenance strategy adjustments.

[0035] In this embodiment, step S4 includes the following steps: Based on the abnormal fluctuation deviation evolution trajectory, the abnormal behavior causal relationship of heterogeneous deviation nodes in the battery is mined to obtain the abnormal behavior causal graph; Model the timeline of the abnormal behavior causal graph and conduct abnormal behavior triggering factors and diffusion analysis to obtain abnormal triggering factors and abnormal diffusion paths; Based on the statistical characteristics of short-term mutation behavior, abnormal trend characteristics are learned and control parameters are defined to generate abnormal evolution parameters under the current trend; Based on the abnormal evolution parameters, a multi-causal fission simulation is performed on the abnormal triggering factors and the abnormal diffusion path to generate a battery behavior deterioration chain under the abnormal trend.

[0036] In this example, "abnormal deviation nodes" have been identified in the battery system. The core task of this step is to deeply explore the causal relationships between the abnormal behavior of these nodes and other nodes, constructing an "abnormal behavior causal graph" to reveal potential fault propagation paths and system-level risks. The specific approach is as follows: First, based on the previously constructed "abnormal fluctuation deviation evolution trajectory," feature extraction is performed on the abnormal behavior of each node to form a behavioral feature vector. Then, methods such as Bayesian networks, Granger causality tests, or structural equation models (SEM) are used to analyze the causal relationships between nodes. By calculating the causal influence between nodes, an "abnormal behavior causal graph" is constructed. The nodes in the graph represent battery nodes, the edges represent the causal relationships between nodes, and the edge weights represent the causal influence. In this experiment, assume that in a battery system, the voltage at node #15 experiences abnormal fluctuations, which in turn causes a corresponding change in the current at node #27. Granger causality tests reveal that the voltage change at node #15 has a significant causal influence on the current change at node #27. Based on this, the constructed "abnormal behavior causal graph" shows a directed edge from node #15 to node #27, with the weight representing the causal influence of voltage changes on current changes. First, using time series analysis methods such as long short-term memory (LSTM) networks and gated recurrent units (GRUs), the battery node behavior is modeled to capture its temporal dependencies. Then, combined with the "abnormal behavior causal graph," the triggering factors of abnormal behavior are analyzed, identifying which nodes' abnormal behavior occurs before others. Next, the diffusion path of abnormal behavior is analyzed, identifying how abnormal behavior propagates between nodes, forming a fault chain. In the experiment, assuming that in a certain battery system, abnormal voltage fluctuations at node #15 occur before abnormal current fluctuations at node #27, the LSTM model is used to model the behavior of nodes #15 and #27. It is found that the abnormal voltage fluctuations at node #15 have a significant temporal dependency on the abnormal current fluctuations at node #27. Combined with the "Causal Graph of Abnormal Behavior," the abnormal voltage fluctuation at node #15 was identified as the trigger of the abnormal behavior, and the abnormal current fluctuation at node #27 as the response. A time-dependent path from #15 to #27 was identified between the two. The trigger of the abnormal behavior and its diffusion path have been identified. The core task of this step is to learn the abnormal trend characteristics based on the statistical characteristics of short-term mutation behavior, define control parameters, and generate abnormal evolution parameters under the current trend. This provides a basis for subsequent anomaly prediction and control strategy optimization.

[0037] The specific approach is as follows: First, based on the statistical characteristics of short-term mutation behavior, such as mutation frequency, mutation amplitude, and duration, an abnormal trend feature vector is constructed. Then, machine learning methods such as support vector machines (SVM) and random forests (RF) are used to learn the abnormal trend features and identify the evolution patterns of abnormal behavior. Next, control parameters such as thresholds and weights are defined to construct an abnormal evolution model and generate abnormal evolution parameters for the current trend. In the experiment, assume that in a battery system, abnormal voltage fluctuations at node #15 occur at a frequency of 5 times / hour, an amplitude of 100 mV, and a duration of 2 seconds. Learning these features using the SVM model reveals a positive correlation between the frequency and amplitude of abnormal voltage fluctuations, and a negative correlation between the duration and amplitude. Based on this, control parameters are defined, such as a frequency threshold of 5 times / hour, an amplitude threshold of 100 mV, and a duration threshold of 2 seconds, to construct an abnormal evolution model and generate abnormal evolution parameters for the current trend. Based on the abnormal evolution parameters, a multi-causal fission model is constructed to simulate the diffusion process of abnormal behavior. Then, the Monte Carlo simulation method was used to simulate the model multiple times to generate battery behavior deterioration chains under different scenarios. Next, the key nodes and key paths in the deterioration chain were analyzed to identify potential failure risk points. In the experiment, it was assumed that in a certain battery system, the frequency of abnormal voltage fluctuations at node #15 was 5 times / hour, the amplitude was 100 mV, and the duration was 2 seconds. The diffusion process of abnormal behavior was simulated through a multi-causal fission model, and it was found that abnormal voltage fluctuations at node #15 may trigger abnormal current fluctuations at node #27, thereby affecting the performance of the entire battery system. Through Monte Carlo simulation, battery behavior deterioration chains under different scenarios were generated, and nodes #15 and node #27 were identified as key risk nodes In this embodiment, step S5 includes the following steps: The abnormal behavior trend evolution of the battery behavior deterioration chain under abnormal trend is carried out to characterize the abnormal behavior evolution; Conduct long-term progressive predictions on the evolution characteristics of abnormal behaviors to obtain long-term abnormal evolution prediction data; Conduct in-depth fault tendency mining on long-term abnormal evolution prediction data to obtain long-term evolution fault tendency information; Multi-point failure probability calculation is performed on the failure tendency information to obtain an abnormal behavior failure probability curve.

[0038] In this example, based on the previously constructed "abnormal behavior causal graph" and "abnormal diffusion path," time series modeling is performed for each deterioration chain to capture its evolution patterns. Deep learning methods, such as long short-term memory (LSTM) networks and gated recurrent units (GRUs), are then used to model the evolution of the deterioration chain and extract "abnormal behavior evolution characteristics." These characteristics include, but are not limited to, evolution speed, evolution amplitude, and evolution period. In this experiment, assume that in a battery system, abnormal voltage fluctuations at node #15 trigger abnormal current fluctuations at node #27, which in turn affects the temperature at node #35. Using the LSTM model to model this deterioration chain, it was found that the evolution speed of the abnormal voltage fluctuations was 0.5 mV / s, the evolution amplitude of the abnormal current fluctuations was 20 mA, and the evolution period of the abnormal temperature fluctuations was 5 hours. Based on these "abnormal behavior evolution characteristics," potential future failure modes, such as thermal runaway caused by excessive temperatures, can be predicted. The "abnormal behavior evolution characteristics" have been extracted in the aforementioned steps. The core task of this step is to perform "long-term progressive prediction" of these features to obtain "long-term abnormal evolution prediction data" for the future, supporting fault warning and maintenance decisions. The specific approach is as follows: First, based on the "abnormal behavior evolution features" extracted earlier, a multi-step prediction model, such as a sequence-to-sequence (seq2seq) model or a Transformer model, is constructed. These models are then used to predict the "abnormal behavior evolution features" for the future, generating "long-term abnormal evolution prediction data." This data includes, but is not limited to, future trends in parameters such as voltage, current, and temperature. In this experiment, assume that in a battery system, abnormal voltage fluctuations at node #15 trigger abnormal current fluctuations at node #27, which in turn affects the temperature at node #35. Using the seq2seq model, a multi-step prediction of this degradation chain is performed, predicting that over the next 24 hours, the voltage will continue to rise by 0.2 mV / s, the current will increase by 10 mA, and the temperature will rise by 2°C. Based on this predicted data, proactive measures can be taken, such as adjusting the charging strategy and increasing heat dissipation, to prevent failures. In the aforementioned steps, the "long-term abnormal evolution prediction data" has been obtained. The core task of this step is to conduct "deep mining of fault tendencies" on these data, identify potential fault risk points, and provide a basis for fault warning and maintenance decisions.

[0039] The specific approach is as follows: First, based on previously acquired long-term abnormal evolution prediction data, failure propensity assessment models, such as support vector machines (SVMs) and random forests (RFs), are constructed. These models are then used to analyze the long-term abnormal evolution prediction data over a period of time to identify potential failure risk points. The failure risk of each node is assessed by calculating a failure propensity index. In this experiment, assume that in a battery system, abnormal voltage fluctuations at node #15 trigger abnormal current fluctuations at node #27, which in turn affects the temperature of node #35. Using the SVM model to analyze this degradation chain, it is found that the temperature of node #35 will rise by 2°C over the next 24 hours. The failure propensity index is 0.8, indicating a high failure risk. Based on this information, proactive maintenance measures, such as replacing cooling equipment and adjusting charging strategies, can be implemented to reduce the failure risk. This information is then used to calculate the probability of multi-point failure and construct an abnormal behavior failure probability curve to quantify the failure risk of each node, providing a basis for fault warning and maintenance decisions. The specific approach is as follows: First, based on the previously identified "failure risk points" and "failure propensity index," a multi-point failure probability model, such as a multivariate Gaussian process (MGP) or Copula model, is constructed. These models are then used to calculate the failure probability of each node, resulting in an "abnormal behavior failure probability curve." This curve displays the failure probability of each node over a period of time, reflecting the overall health of the system. In an experiment, assume that in a battery system, abnormal voltage fluctuations at node #15 trigger abnormal current fluctuations at node #27, which in turn affects the temperature at node #35. Using the Copula model, the multi-point failure probability calculation for this deterioration chain reveals a failure probability of 0.9 for node #35, indicating an extremely high failure risk. Based on this information, measures can be taken, such as premature battery replacement and charging strategy adjustments, to prevent failures.

[0040] In this embodiment, step S6 includes the following steps: Identify the peak point of failure probability based on the abnormal behavior failure probability curve; Identify the battery behavior data, voltage response curve, state evolution trajectory and battery topology evolution data of the fault probability point; Perform multi-dimensional fault behavior encapsulation on the battery behavior data, voltage response curve, state evolution trajectory, and battery topology structure evolution data to obtain a fault behavior feature vector set; Performing enhanced transfer learning on the fault behavior feature vector set to construct a battery fault prototype clone; Performing fault feature migration training on the BMS based on the battery fault prototype clone, thereby building a fault feature prediction and recognition model; Based on the fault feature prediction and recognition model, battery fault similarity recognition and prediction are performed.

[0041] In this embodiment, after obtaining the abnormal behavior failure probability curve generated during the battery's operating cycle, the system first identifies key points along the entire curve, focusing on time periods where the probability suddenly increases. These high-probability regions typically correspond to moments when the potential failure trend is nearing its breaking point, and are therefore referred to as failure probability peaks. Identifying these peaks requires analyzing the slope and acceleration of the curve's first- and second-order derivatives to determine whether the probability mutation point represents a random fluctuation or a trend-based jump. If the failure probability jumps from 12% in the previous 30 seconds to 41% at the 3840th second of battery operation, the system automatically identifies this point as a potential failure trigger window. To enhance stability, the system sets a minimum threshold change amplitude (ΔP ≥ 0.15) and a minimum duration window (T ≥ 45 seconds) as criteria for peak point validity. Furthermore, the system uses historical sample statistical data to fit the curve distribution and employs a Gaussian mixture model to categorize the curve into multiple risk-level distribution regions. The multiple identified failure probability peaks are numbered and archived for subsequent behavior tracking and feature learning. After successfully identifying one or more peak points of fault probability, the system traces key behavioral data from several time windows around these points to construct a complete fault context. First, the system extracts battery behavior data from this period, including current fluctuations, temperature change rates, and internal resistance dynamic trends. It also simultaneously extracts voltage response curves—the voltage recovery trajectory of the battery cell after the application or removal of an external load. Under abnormal conditions, such curves typically exhibit delayed voltage recovery, frequent fluctuations, and shortened peak-to-valley intervals. Second, the system dynamically models the changes in state of charge (SOC) and state of hydration (SOH) during this period, generating a battery "state evolution trajectory." Before a particular fault peak, the SOC may exhibit a continuous error increase, while the SOH may exhibit a transient downward trend. Furthermore, combined with data on the evolution of the physical connection topology between the battery cells, the system tracks changes in conduction state within the topological connection paths, as well as the reconfiguration of offset paths. This information, combined with the evolution of the physical connection topology between the battery cells, forms a comprehensive dataset of the events surrounding the fault node, providing a multi-dimensional perspective across time, space, logic, and physics, laying the foundation for subsequent fault mode encapsulation and transfer learning. After comprehensively collecting behavioral data associated with peak fault probability points, the system formats, encapsulates, and uniformly encodes this data to construct a multi-dimensional fault behavior feature vector set. Specifically, the system first extracts and normalizes each type of data: Voltage response curves are transformed using Fourier transforms to extract the primary frequency distribution features, state evolution trajectories are measured using the DTW (Dynamic Time Warping) method to measure trajectory offsets, and topological evolution is converted into a vector representation of the graph structure change rate (such as changes in node centrality and the number of connected paths).Afterward, all feature dimensions are concatenated into a high-dimensional feature vector. The system uses PCA (Principal Component Analysis) to reduce the feature dimension to between 128 and 256, ensuring learnability and computational efficiency while preserving information. Feature vectors may include metrics such as voltage hysteresis mean, internal resistance surge rate, thermal gradient differential, and topology reconstruction frequency. The packaged fault behavior feature vector set is used to construct a battery "fault behavior prototype," a high-dimensional representation of a specific fault state, mapped to an actual failure mode (such as gassing, SEI film disintegration, or electrode corrosion) in the form of data labels. To generalize battery fault signatures and enable cross-module and cross-device transfer and adaptation, the system utilizes a reinforcement transfer learning mechanism to model the fault feature vector set and construct multiple fault prototype clones. The core approach is to transfer high-value behavioral patterns from the source task (known fault feature set) to the target task (unknown sample identification). At this stage, the system introduces a dual-tower neural network: one tower processes historical fault samples (the source domain), while the other processes live samples from the BMS system to be trained (the target domain). Joint comparative learning is used to optimize transfer performance. A policy gradient boosting mechanism is incorporated into the training process, enabling the model to dynamically adjust attention weights when recognizing ambiguous sample features, enhancing its ability to represent key metrics. When internal resistance changes are subtle but thermal gradients are severe, the system tends to prioritize thermal features in prototype construction. The resulting battery fault prototype clones are fault feature data models with transfer and generalization capabilities, serving as labels to guide subsequent model training tasks. After the fault prototype clones are constructed, they are input as training samples into the learning engine embedded in the BMS for fault feature transfer training. The model utilizes an incremental training approach, gradually embedding the fault prototypes into the existing recognition framework to avoid overfitting and concept drift. The training architecture is based on a fusion of the Transformer and multi-task learning architecture, ensuring time series modeling capabilities while supporting the coordinated optimization of subtasks such as voltage response, thermal evolution, and SOC fluctuation. The training goal is to enable the model to recognize similar failure modes in different battery modules, different operating conditions, and different lifecycle stages. The high-temperature jump behavior exhibited by cell number C17 under operating condition A should be recognized by the model as highly similar to the electrode aging failure of C25 under operating condition B. The fault feature prediction and recognition model ultimately trained has prototype-like recognition capabilities and global feature generalization capabilities, and can be widely deployed in the BMS platform to achieve identification and early warning of future unknown faults. After completing the fault feature model training, the BMS system converts the real-time collected battery operation data into feature vectors during actual operation, compares and analyzes them with the fault prototypes in the model, and thus performs the battery fault similarity identification and prediction task. The system calculates the matching degree of the current feature vector with all known fault prototypes using indicators such as Euclidean distance, cosine similarity, and KL divergence to determine whether it is close to a specific fault evolution pattern.At the same time, the system uses linear extrapolation of historical trends and nonlinear prediction models, combined with the current similarity growth rate, to form a probability prediction for possible faults within a certain future time window. Within the prediction window T = 24 hours, if the similarity between the current feature vector of the C33 battery cell and the "high resistance-thermal expansion-electrode delamination" prototype is greater than 0.91, and the similarity growth rate is +0.06 / h, the system will determine that its probability of fault occurrence within T exceeds 70%. Ultimately, this prediction result not only provides early fault warning but also collaborates with the vehicle control system to implement active protection, current limiting strategies, dynamic balancing adjustments, and other operations, comprehensively improving the reliability and safety of the battery system.

[0042] In this embodiment, a lithium-ion battery fault prediction system based on a BMS is provided, which is used to execute the lithium-ion battery fault prediction method based on a BMS as described above, including: The multi-state perception module is used to extract multi-dimensional battery operation monitoring parameters through the BMS, perform multi-state evolution perception and label mapping, and build a global multi-state perception map of the battery; The mutation detection module is used to detect short-term abnormal mutations based on the multi-dimensional operation monitoring parameters of the battery, and analyze the deviation trend of normal characteristics, thereby constructing the evolution trajectory of abnormal fluctuation deviation; The abnormal similarity analysis module is used to perform deep topological association learning on the global multi-state perception map of the battery based on the abnormal fluctuation deviation evolution trajectory, and perform global perception of heterogeneous nodes to extract heterogeneous deviation nodes in the battery; The fission simulation module is used to mine the causal relationship of abnormal behavior of heterogeneous deviation nodes in the battery and perform multi-causal fission simulation to generate the battery behavior deterioration chain under abnormal trends; The fault probability calculation module is used to analyze the abnormal behavior trend evolution of the battery behavior deterioration chain under abnormal trends, and perform multi-point failure probability calculation to obtain the abnormal behavior failure probability curve; The fault transfer learning module is used to encapsulate multi-dimensional fault behaviors based on abnormal behavior fault probability curves, and then conduct fault feature transfer training to perform battery fault similarity identification and prediction.

[0043] 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.

[0044] 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 lithium-ion battery fault prediction method based on BMS, characterized in that: The following steps are involved: Step S1: Extract multi-dimensional battery operation monitoring parameters through the BMS, perform multi-state evolution perception and label mapping processing, and build a global multi-state perception map of the battery; Step S2: Perform short-term abnormal mutation detection based on the multi-dimensional operation monitoring parameters of the battery, and perform normal feature deviation trend analysis to construct the abnormal fluctuation deviation evolution trajectory; Step S3: Based on the abnormal fluctuation deviation evolution trajectory, deep topological association learning is performed on the global multi-state perception map of the battery, and global perception of heterogeneous nodes is performed to extract heterogeneous deviation nodes in the battery; Step S4: mining abnormal behavior causal relationships for heterogeneous deviation nodes in the battery, and performing multi-causal fission simulation to generate a battery behavior deterioration chain under abnormal trends; Step S5: performing abnormal behavior trend evolution on the battery behavior deterioration chain under the abnormal trend, and performing multi-point failure probability calculation to obtain an abnormal behavior failure probability curve; Step S6: Perform multi-dimensional fault behavior encapsulation based on the abnormal behavior fault probability curve, and then perform fault feature migration training to perform battery fault similarity identification and prediction.

2. The lithium-ion battery fault prediction method based on BMS according to claim 1, characterized in that: The specific steps of step S1 are: Setting a high-frequency sampling time window; performing millisecond-level multi-dimensional data acquisition through the BMS based on the high-frequency sampling time window to extract multi-dimensional battery operation monitoring parameters, including voltage, current, temperature, internal resistance, state of charge (SOC), and capacity; Adaptively filtering the multi-dimensional operating monitoring parameters of the battery and correcting sensor acquisition deviations to obtain deviation-optimized monitoring parameters; Identify the current battery operating scenario based on the BMS; analyze and divide the time periods based on the battery operating scenario to obtain time periods of equal length; Performing parameter segmentation processing on the deviation optimization monitoring parameters according to the equal-length time periods to extract monitoring parameters for multiple time periods; The multi-state evolution perception of monitoring parameters in multiple time periods is performed, and label mapping processing is performed to construct a global multi-state perception map of the battery.

3. The lithium-ion battery fault prediction method based on BMS according to claim 2, characterized in that: The specific steps of performing multi-state evolution perception on monitoring parameters in multiple time periods and label mapping processing to construct a global multi-state perception map of the battery are as follows: Extract nonlinear features of monitoring parameters in multiple time periods one by one to identify sudden increases and decreases in battery performance, nonlinear jitter, and pre-critical oscillation characteristics, thereby obtaining the battery status characteristics in different time periods. Performing deep state analysis on the state features and generating state labels; Performing multi-time period fluctuation mining on the state characteristics to obtain a fluctuation trend of the battery state characteristics; Based on the state labels, the multi-state evolution perception of the battery state characteristic fluctuation trend is carried out, and label mapping processing is performed to construct a multi-state perception map.

4. The lithium-ion battery fault prediction method based on BMS according to claim 1, characterized in that: The specific steps of step S2 are: Perform short-term abnormal mutation detection based on multi-dimensional battery operation monitoring parameters and extract abnormal mutation characteristics; Based on the preset standard battery behavior database, the normal fluctuation range of abnormal mutation characteristics is judged. If it is determined to be an abnormal fluctuation beyond the range, it is marked as an abnormal fluctuation event; Calculating the frequency of occurrence, deviation amplitude and abnormal duration of the abnormal fluctuation event; Multidimensional cluster analysis is performed on the occurrence frequency, deviation amplitude and abnormal duration to obtain the statistical characteristics of short-term mutation behavior; The normal characteristic deviation trend of the statistical characteristics of short-term mutation behavior is analyzed and dynamically evolved to construct the abnormal fluctuation deviation evolution trajectory.

5. The lithium-ion battery fault prediction method based on BMS according to claim 1, characterized in that: The specific steps of step S3 are: Position and identify abnormal battery nodes based on abnormal mutation characteristics; Based on the abnormal fluctuation deviation evolution trajectory, the battery global multi-state perception map is analyzed for the similarity of the behavior of all battery nodes, and the battery nodes with similar abnormal fluctuations are identified; Identify the differences and synchronization of abnormal battery nodes and battery nodes with similar abnormal fluctuations; Performing deep topological association learning on the differences and synchronizations to extract the group topological mapping relationship of the battery nodes; Based on the group topology mapping relationship and the abnormal fluctuation deviation evolution trajectory, global perception of heterogeneous nodes is performed to extract heterogeneous deviation nodes in the battery.

6. The lithium-ion battery fault prediction method based on BMS according to claim 1, characterized in that: The specific steps of step S4 are: Based on the abnormal fluctuation deviation evolution trajectory, the abnormal behavior causal relationship of heterogeneous deviation nodes in the battery is mined to obtain the abnormal behavior causal graph; Model the timeline of the abnormal behavior causal graph and conduct abnormal behavior triggering factors and diffusion analysis to obtain abnormal triggering factors and abnormal diffusion paths; Based on the statistical characteristics of short-term mutation behavior, abnormal trend characteristics are learned and control parameters are defined to generate abnormal evolution parameters under the current trend; Based on the abnormal evolution parameters, a multi-causal fission simulation is performed on the abnormal triggering factors and the abnormal diffusion path to generate a battery behavior deterioration chain under the abnormal trend.

7. The lithium-ion battery fault prediction method based on BMS according to claim 1, characterized in that: The specific steps of step S5 are: The abnormal behavior trend evolution of the battery behavior deterioration chain under abnormal trend is carried out to characterize the abnormal behavior evolution; Conduct long-term progressive predictions on the evolution characteristics of abnormal behaviors to obtain long-term abnormal evolution prediction data; Conduct in-depth fault tendency mining on long-term abnormal evolution prediction data to obtain long-term evolving fault tendency information; Multi-point failure probability calculation is performed on the failure tendency information to obtain an abnormal behavior failure probability curve.

8. The lithium-ion battery fault prediction method based on BMS according to claim 1, characterized in that: The specific steps of step S6 are: Identify the peak point of failure probability based on the abnormal behavior failure probability curve; Identify the battery behavior data, voltage response curve, state evolution trajectory and battery topology evolution data of the fault probability point; Performing multi-dimensional fault behavior encapsulation on the battery behavior data, voltage response curve, state evolution trajectory, and battery topology structure evolution data to obtain a fault behavior feature vector set; Performing enhanced transfer learning on the fault behavior feature vector set to construct a battery fault prototype clone; Performing fault feature migration training on the BMS based on the battery fault prototype clone, thereby building a fault feature prediction and recognition model; Based on the fault feature prediction and recognition model, battery fault similarity recognition and prediction are performed.

9. A lithium-ion battery fault prediction system based on BMS, characterized in that: The method for predicting lithium-ion battery failure based on a BMS according to claim 1 comprises: The multi-state perception module is used to extract multi-dimensional battery operation monitoring parameters through the BMS, perform multi-state evolution perception and label mapping, and build a global multi-state perception map of the battery; The mutation detection module is used to detect short-term abnormal mutations based on the multi-dimensional operation monitoring parameters of the battery, and analyze the deviation trend of normal characteristics, thereby constructing the evolution trajectory of abnormal fluctuation deviation; The abnormal similarity analysis module is used to perform deep topological association learning on the global multi-state perception map of the battery based on the abnormal fluctuation deviation evolution trajectory, and perform global perception of heterogeneous nodes to extract heterogeneous deviation nodes in the battery; The fission simulation module is used to mine the causal relationship of abnormal behavior of heterogeneous deviation nodes in the battery and perform multi-causal fission simulation to generate the battery behavior deterioration chain under abnormal trends; The fault probability calculation module is used to analyze the abnormal behavior trend evolution of the battery behavior deterioration chain under abnormal trends, and perform multi-point failure probability calculation to obtain the abnormal behavior failure probability curve; The fault transfer learning module is used to encapsulate multi-dimensional fault behaviors based on abnormal behavior fault probability curves, and then conduct fault feature transfer training to perform battery fault similarity identification and prediction.

Images