A battery cell abnormality determination method and system based on BMS data driving

Abstract

The application discloses a kind of based on BMS data driving's cell abnormality determination method and system, utilize multiple sensors and BMS collection multidimensional cell operating data, and the data of collection is preprocessed to form basic data set;The basic data set obtained is reduced in dimension and feature extraction, the characteristic value extracted is compared with the dynamic self-adapting threshold value obtained based on dynamic threshold adjusting mechanism to obtain cell health state evaluation result, generates abnormal early warning signal and locates potential fault point according to cell health state evaluation result;Abnormal grading is carried out to abnormal early warning signal and maintenance suggestion is generated, generates exception record and is stored to cloud database.The application realizes the accurate identification and fast response of cell abnormality.

Description

Technical Field

[0001] This invention relates to the field of battery cell anomaly detection technology, specifically to a battery cell anomaly detection method and system based on BMS data. Background Technology

[0002] With the rapid development of Battery Management System (BMS) technology, data-driven cell anomaly detection methods have played a crucial role in improving battery pack safety, reliability, and lifespan, particularly in the field of quality and safety prediction for communication backup batteries. By building algorithms and utilizing multi-dimensional data collected by sensors and BMS to conduct fault diagnosis analysis and output conclusions, maintenance personnel can quickly grasp the health status of cells, promptly address potential faults, and prevent safety accidents. However, existing cell anomaly detection technologies still have shortcomings in data processing efficiency, dynamic operating condition adaptability, and fault identification accuracy, making it difficult to fully meet the needs of real-time monitoring of cell health status under complex operating conditions. Therefore, there is an urgent need for a cell anomaly detection method and system that can optimize data processing flows, introduce adaptive algorithms, and enhance the dynamic update capability of models to improve the efficiency and accuracy of cell anomaly detection, thereby meeting the needs of battery management systems for efficient and intelligent monitoring. Summary of the Invention

[0003] To address this, embodiments of the present invention provide a cell anomaly determination method and system based on BMS data-driven approach, in order to solve the problems of low data processing efficiency, poor adaptability to dynamic operating conditions, and difficulty in meeting the demand for efficient monitoring of cell health status under complex operating conditions in existing cell anomaly detection technologies.

[0004] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0005] According to a first aspect of the present invention, the present invention provides a cell anomaly determination method based on BMS data driving, the method comprising:

[0006] Multiple sensors and BMS are used to collect multi-dimensional cell operation data, and the collected data is preprocessed to form a basic dataset;

[0007] The obtained basic dataset is subjected to dimensionality reduction and feature extraction. The extracted feature values ​​are compared with the dynamic adaptive threshold obtained based on the dynamic threshold adjustment mechanism to obtain the cell health status assessment results. Based on the cell health status assessment results, abnormal early warning signals are generated and potential fault points are located.

[0008] The system classifies abnormal early warning signals into different levels and generates maintenance recommendations, while also generating abnormal records and storing them in a cloud database.

[0009] Furthermore, multiple sensors and a BMS are used to collect multi-dimensional cell operation data, specifically including:

[0010] The sensors used include temperature sensors, voltage sensors, current sensors, and ambient humidity sensors; the collected data includes cell temperature, individual cell voltage, charging and discharging current, and ambient humidity.

[0011] Furthermore, the collected data undergoes preprocessing to form a basic dataset, specifically including:

[0012] The collected raw data is cleaned and preliminarily classified to remove noise points and outliers. Data cleaning includes using a sliding window filtering algorithm to remove noise. Preliminary classification includes dividing the collected data into static basic data and dynamic running data according to data attributes.

[0013] Furthermore, the obtained basic dataset undergoes dimensionality reduction and feature extraction, specifically including:

[0014] Based on principal component analysis, and by introducing dynamic feature weighting and physical constraint mechanisms, dimensionality reduction is performed on multi-dimensional data.

[0015] Furthermore, dynamic feature weighting specifically includes:

[0016] Different weights are assigned based on the differences in importance of different time-series features, and a dynamic weight vector is generated. The original feature matrix is ​​weighted, and the formula is expressed as:

[0017]

[0018] in, This is the original time-series window feature matrix. This is a weighted diagonal matrix. This is the weighted feature matrix.

[0019] Furthermore, the physical constraint mechanism specifically includes:

[0020] Embedding physical constraint terms into the covariance matrix is ​​expressed by the following formula:

[0021]

[0022] In the formula, This is the covariance matrix that is finally obtained after dynamic feature weighting and embedding physical constraint terms; This is a weighted mean matrix; N is the sample size. Let be the physical constraint matrix, and satisfy: if the characteristic With features If there is a physical relationship, then the element in the i-th row and j-th column of matrix C... =1, otherwise =0; These are constraint coefficients;

[0023] The covariance matrix Eigenvalue decomposition yields multiple eigenvalues ​​and eigenvectors; the first eigenvalues ​​are selected. Construct a projection matrix from eigenvectors Finally, the dimensionality-reduced matrix is ​​obtained. for:

[0024]

[0025] This allows us to extract representative key operational features.

[0026] Furthermore, the extracted feature values ​​are compared with a dynamically adaptive threshold obtained based on a dynamic threshold adjustment mechanism to obtain the cell health status assessment result, specifically including:

[0027] A dynamic adaptive threshold is obtained by employing a dynamic threshold adjustment mechanism that incorporates reinforcement learning, including:

[0028] Introducing a deep reinforcement learning training adjustment coefficient , The data varies with time t; the reward function is to minimize the false alarm rate of the cell health status assessment result; the state space consists of the current operating parameters and the prediction error. ,in Here are the predicted values ​​of the cell performance parameters at time t. The measured values ​​of the cell performance parameters at time t; the action space is... That is, it covers the 90%-99.7% confidence interval in statistics, which is obtained based on the statistical characteristics of the normal distribution;

[0029] Then, the policy network is iteratively optimized using a proximal policy optimization algorithm to output real-time adjustment coefficients:

[0030]

[0031] in, This is the current state vector; This represents the near-end strategy optimization algorithm;

[0032] The dynamic adaptive threshold is obtained based on the reinforcement learning algorithm, and the calculation formula is as follows:

[0033]

[0034] in, For dynamic adaptive threshold, Here are the predicted values ​​of the cell performance parameters at time t. Standard deviation, This is the adjustment coefficient.

[0035] Furthermore, the extracted feature values ​​are compared with a dynamically adaptive threshold obtained based on a dynamic threshold adjustment mechanism to obtain the cell health status assessment result. Based on the cell health status assessment result, an abnormality warning signal is generated and potential fault points are located, specifically including:

[0036] When the extracted feature values ​​exceed the set dynamic adaptive threshold, an abnormal warning signal is generated.

[0037] Potential fault points are identified using a localization algorithm, including: based on the differences in operating characteristics between battery cells, cluster analysis is used to distinguish abnormal battery cells from normal battery cells, thereby determining the location of abnormal battery cells.

[0038] Furthermore, the abnormal early warning signals are classified into anomaly levels and maintenance suggestions are generated, specifically including:

[0039] Based on the severity of the anomaly, the anomaly warning signal is divided into three levels: minor anomaly, moderate anomaly, and severe anomaly;

[0040] The grading standard is established based on the degree to which abnormal feature values ​​deviate from the dynamic adaptive threshold. The specific formula is as follows:

[0041]

[0042] in, These are abnormal characteristic values. For dynamic adaptive threshold, For the percentage deviation; when It was initially determined to be a minor abnormality. It was determined to be moderately abnormal at that time. It was determined to be a serious abnormality at that time.

[0043] According to a second aspect of the present invention, an embodiment of the present invention provides a cell anomaly determination system based on BMS data, the system comprising:

[0044] The data acquisition module is used to collect multi-dimensional cell operation data using various sensors and BMS, and to preprocess the collected data to form a basic dataset.

[0045] The intelligent analysis module is used to perform data dimensionality reduction and feature extraction on the obtained basic dataset. It compares the extracted feature values ​​with the dynamic adaptive threshold obtained based on the dynamic threshold adjustment mechanism to obtain the cell health status assessment results. Based on the cell health status assessment results, it generates abnormal early warning signals and locates potential fault points.

[0046] The maintenance decision module is used to classify abnormal warning signals into different levels and generate maintenance suggestions, generate abnormal records and store them in the cloud database.

[0047] Compared with existing technologies, the present invention provides a cell anomaly determination method and system based on BMS data-driven analysis. By preprocessing multi-dimensional data, it ensures data accuracy and consistency while eliminating redundant information and improving data utilization. Combining cell operating characteristics and dynamic health assessment results, it achieves accurate identification and rapid response to potential faults, ensuring the stable operation of communication backup power systems. Furthermore, by comprehensively considering the cell's operating environment and aging trends, it generates targeted maintenance strategies, reducing the risk of safety accidents and improving the overall reliability and lifespan of the battery pack, providing strong support for the efficient operation and maintenance of communication backup power systems. Attached Figure Description

[0048] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0049] Figure 1 A schematic diagram of the overall workflow of a cell anomaly determination method based on BMS data provided in an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of a cell anomaly determination system based on BMS data drive, provided in an embodiment of the present invention. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0052] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0053] The first embodiment of this invention provides a cell anomaly determination method based on BMS data drive. The following is in conjunction with... Figure 1 Please provide a detailed explanation.

[0054] like Figure 1 As shown, in step S100, multiple sensors and BMS are used to collect multi-dimensional cell operation data, and the collected data is preprocessed to form a basic dataset.

[0055] The above steps specifically include:

[0056] S110, Data Acquisition.

[0057] This system utilizes multiple sensors and a battery management system (BMS) to acquire multi-dimensional operational data. These sensors include temperature sensors, voltage sensors, current sensors, and ambient humidity sensors, whose layout is optimized to minimize the impact of external environmental interference on data acquisition. For example, in practical applications of communication backup power systems, sensor locations are strategically distributed in key areas of the battery pack, such as the positive and negative terminals of the cells, the surface of the battery pack casing, and near internal heat dissipation channels, ensuring that the collected data comprehensively reflects the cell's operating status. In this embodiment, the collected data includes cell temperature, individual cell voltage, charging and discharging current, and ambient humidity.

[0058] S120, Data Preprocessing.

[0059] In this embodiment, the collected raw data is first cleaned and preliminarily classified by a preprocessing algorithm to remove noise points and outliers, forming a basic dataset.

[0060] Specifically, the data cleaning uses a sliding window filtering algorithm, with the following formula: , where x Represents the original data sequence. To adjust the sliding window size, This is the cleaned data. The algorithm effectively removes high-frequency noise by averaging the data within a time window, while retaining the main trend characteristics of the data.

[0061] The initial classification is based on data attributes, dividing it into static basic data (such as nominal parameters of the battery cell) and dynamic operating data (such as real-time operating status) so that the subsequent analysis modules can call them efficiently.

[0062] like Figure 1 As shown, in step S200, the obtained basic dataset is subjected to data dimensionality reduction and feature extraction. The extracted feature values ​​are compared with the dynamic adaptive threshold obtained based on the dynamic threshold adjustment mechanism to obtain the cell health status assessment result. Based on the cell health status assessment result, an abnormal warning signal is generated and potential fault points are located.

[0063] The above steps specifically include:

[0064] S210, Feature Extraction.

[0065] Based on traditional PCA (principal component analysis) techniques, dynamic feature weighting and physical constraint mechanisms are introduced to perform dimensionality reduction on multi-dimensional data. The specific techniques are as follows:

[0066] 1. Construction of dynamically weighted original matrix

[0067] The model introduces a dynamic weight vector to address the differences in the physical importance of different time-series features (e.g., voltage features have a higher weight in health assessment than temperature). ,in (Voltage-related feature weights) ≥0.05, weight of temperature-related features ≤0.02), where 30 is the number of features in this embodiment, determined by the data region, or it can be other constants; the original feature matrix is ​​weighted:

[0068]

[0069] in, The original time-series window feature matrix is ​​N, which is the number of samples, i.e., the total number of cell operation data (total number of time points) used for analysis. This is a weighted diagonal matrix; This is the weighted feature matrix. Weighting prioritizes physically significant features during dimensionality reduction.

[0070] 2. Calculation of Covariance Matrix with Physical Constraints

[0071] To avoid losing core physical relationships (such as the Ohm's law relationship between voltage and current) during dimensionality reduction, physical constraint terms are embedded in the covariance matrix:

[0072]

[0073] In the formula, It is a weighted mean matrix. , here For the sample size, For the i-th weighted feature, This is the physical constraint matrix, where 30 represents the existence of 30 features in the data, determined by the data region, but can also be other constants. It also satisfies: if the features... and If there is a physical relationship (such as voltage and current), then =1, otherwise =0; =0.01 is the constraint coefficient, which ensures that the covariance of the physical association features is not over-compressed.

[0074] 3. Feature Vector Selection and Dimensionality Reduction Mapping

[0075] right Perform eigenvalue decomposition to obtain eigenvalues. and eigenvectors Before selection Construct a projection matrix from eigenvectors (i.e., eigenvectors with large cumulative contribution rates). The final dimension-reduced matrix is:

[0076] This allows us to extract representative key operational features.

[0077] This improved PCA algorithm can map high-dimensional data to a low-dimensional space, significantly reducing computational complexity while preserving key information. Extracted features include the rate of change of cell internal resistance, voltage consistency deviation, and temperature gradient, which directly reflect the health status of the cell.

[0078] S220, Health Status Assessment.

[0079] In this embodiment, an adaptive threshold mechanism is introduced to dynamically adjust the threshold range based on historical data and current operating conditions.

[0080] Specifically, the formula for calculating the adaptive threshold is: ,in, For dynamic adaptive threshold, Here are the predicted values ​​of the cell performance parameters at time t. Standard deviation This is the adjustment coefficient.

[0081] During the health status assessment phase, the intelligent analysis module employs a dynamic threshold adjustment mechanism integrated with reinforcement learning, overcoming the limitations of traditional fixed-coefficient thresholds and achieving precise adaptive adjustment of the threshold range. The specific technical solution is as follows:

[0082] This stage introduces Deep Reinforcement Learning (DRL) to train adjustment coefficients. The reward function is "minimizing the false alarm rate of the health assessment results". The state space consists of the current operating parameters (temperature, current, SOH (State of Health, battery health status)) and the prediction error. ,in The predicted values ​​of cell performance parameters (including SOC, i.e., State of Charge, remaining capacity) at time t. Let be the measured values ​​of the cell performance parameters at time t, and let be the action space. That is, it covers the 90%-99.7% confidence interval in statistics.

[0083] Next, in this stage, the policy network is iteratively optimized using the PPO (proximal policy optimization) algorithm to output real-time adjustment coefficients:

[0084]

[0085] in, This represents the current state vector. For example, in low-temperature operating conditions (ambient temperature...),... < At 10℃, Automatically increases to 2.5 or higher to tolerate greater characteristic fluctuations; under stable operating conditions of healthy batteries (SOH≥90%), Reduced to below 1.5 to improve anomaly detection sensitivity.

[0086] Based on the above reinforcement learning algorithm, the formula for calculating the dynamic threshold is as follows:

[0087]

[0088] The above adjustment coefficients are based on reinforcement learning. It can dynamically balance the "missed detection rate" and "false alarm rate" according to the working conditions, solving the problem of insufficient adaptability of fixed coefficients in complex scenarios.

[0089] By combining improved physical features extracted by PCA, threshold adjustment is deeply correlated with core indicators of battery health status, thereby enhancing the physical interpretability of the assessment.

[0090] S230, Abnormal Warning.

[0091] In this embodiment, By introducing dynamic statistics, the evaluation results become more environmentally adaptable. When a certain feature value exceeds a set adaptive threshold, the intelligent analysis module generates an anomaly warning signal and identifies potential fault points through a location algorithm.

[0092] The localization algorithm, based on the differences in operational characteristics between battery cells, uses cluster analysis to distinguish abnormal cells from other normal cells. The mathematical model for cluster analysis is as follows: ,in Indicates the first A cluster, For the sample to be classified, As cluster center, Let be the distance between the sample and the cluster center, and k represent the index of any other category different from the current target category i. This algorithm can quickly locate abnormal battery cells, providing precise guidance for subsequent maintenance.

[0093] like Figure 1 As shown, in step S300, the abnormal warning signal is classified into abnormal levels and maintenance suggestions are generated. An abnormal record is generated and stored in the cloud database.

[0094] The above steps specifically include:

[0095] S310, Abnormal Classification.

[0096] In the anomaly classification stage, this embodiment divides the warning signal into three levels according to the severity of the anomaly: minor anomaly, moderate anomaly, and severe anomaly.

[0097] The grading standard is based on the degree to which abnormal feature values ​​deviate from an adaptive threshold, and the specific formula is as follows: ,in These are abnormal characteristic values. For dynamic adaptive threshold, Percentage deviation. When It was initially determined to be a minor abnormality. It was determined to be moderately abnormal at that time. It was determined to be a serious abnormality at that time.

[0098] S320, Strategy Generation.

[0099] During the strategy generation phase, this embodiment outputs corresponding maintenance recommendations based on different anomaly levels. For example, for minor anomalies, it is recommended to increase the monitoring frequency and record relevant data; for moderate anomalies, it is recommended to conduct local inspections and replace some aging components; for severe anomalies, it is recommended to immediately disconnect the power for repair and replace the faulty battery cell.

[0100] S330, data storage.

[0101] During the data storage phase, this embodiment generates anomaly records, including historical data and processing results, and uploads them to a cloud database. The cloud database employs a distributed storage architecture, supporting efficient access and analysis of large-scale data. The stored data includes basic datasets, anomaly characteristic values, maintenance suggestion records, and user feedback information, providing crucial data for subsequent system optimization.

[0102] Corresponding to the BMS data-driven cell anomaly determination method disclosed in the above embodiments, this invention also discloses a BMS data-driven cell anomaly determination system, specifically including:

[0103] The data acquisition module is used to collect multi-dimensional cell operation data using various sensors and BMS, and to preprocess the collected data to form a basic dataset.

[0104] The intelligent analysis module is used to perform data dimensionality reduction and feature extraction on the obtained basic dataset. It compares the extracted feature values ​​with the dynamic adaptive threshold obtained based on the dynamic threshold adjustment mechanism to obtain the cell health status assessment results. Based on the cell health status assessment results, it generates abnormal early warning signals and locates potential fault points.

[0105] The maintenance decision module is used to classify abnormal warning signals into different levels and generate maintenance suggestions, generate abnormal records and store them in the cloud database.

[0106] It should be noted that for a detailed description of the cell anomaly determination system based on BMS data provided in the embodiments of the present invention, please refer to the relevant description of the cell anomaly determination method based on BMS data provided in the embodiments of the present invention, which will not be repeated here.

[0107] Application example:

[0108] In a backup power system of a communication base station, the cell anomaly detection method of this invention is used to manage the health of a group of lithium-ion batteries. The battery pack consists of 16 individual cells connected in series, with a rated voltage of 51.2V and a capacity of 100Ah. During system operation, the data acquisition module collects cell voltage, temperature, and current data every second and transmits the data to the intelligent analysis module via a wireless communication interface. The intelligent analysis module extracts cell operating characteristics in real time and assesses health status. It detects that cell #12 has a voltage consistency deviation exceeding the adaptive threshold in three consecutive samplings, triggering an anomaly warning signal. The maintenance decision module determines the anomaly as moderate based on the degree of deviation of the anomaly characteristic values ​​and outputs maintenance recommendations: It is recommended to perform a partial inspection of cell #12 and replace aging components. Operators conduct on-site inspection according to the recommendations and find slight corrosion on the cell's connection terminals, which is promptly repaired. Through the application of this invention, the stability and safety of the communication backup power system are significantly improved, while the overall service life of the battery pack is extended by approximately 15%.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for determining an abnormality of a battery cell based on BMS data, characterized in that, The method comprises: Collecting multi-dimensional cell operation data by using various sensors and the BMS, and preprocessing the collected data to form a basic data set; Performing data dimension reduction and feature extraction on the obtained basic data set, comparing the extracted feature values with dynamic adaptive thresholds obtained based on a dynamic threshold adjustment mechanism to obtain a cell health state evaluation result, generating an abnormal early warning signal and locating a potential fault point according to the cell health state evaluation result; The data dimension reduction and feature extraction on the obtained basic data set specifically include: Based on principal component analysis, and introducing a dynamic feature weighting and physical constraint mechanism, the multi-dimensional data is reduced in dimension; The dynamic feature weighting specifically includes: Different weights are given according to the importance difference of different time sequence characteristics, and a dynamic weight vector is generated The original feature matrix is weighted, and the formula is represented as wherein, is the original time window feature matrix, is a weight diagonal matrix, is the weighted feature matrix; The physical constraint mechanism specifically includes: A physical constraint term is embedded in the covariance matrix, and the formula is: In the formula, is the covariance matrix finally obtained after dynamic feature weighting processing, which embeds a physical constraint term; is a weighted mean matrix; N is the number of samples; is a physical constraint matrix, and satisfies: if the feature is physically associated with the feature , the element in the i-th row and the j-th column of the matrix C is =1, otherwise =0; is a constraint coefficient; The covariance matrix is calculated as follows Eigenvalue decomposition is performed to obtain a plurality of eigenvalues and eigenvectors, and the first eigenvectors are selected to construct a projection matrix , and finally a reduced matrix is obtained : Thus, the representative key operation features are extracted; The extracted feature values are compared with dynamic adaptive thresholds obtained based on a dynamic threshold adjustment mechanism to obtain a cell health state evaluation result, which specifically includes: The dynamic adaptive thresholds are obtained by using a dynamic threshold adjustment mechanism based on reinforcement learning, which includes: Introducing deep reinforcement learning to train regulation coefficients , as a function of time t; reward function is to minimize false positive rate of state of health estimation; state space is current operating parameters and prediction error wherein is the predicted value of the battery performance parameter at time t, is the measured value of the battery performance parameter at time t; action space is i.e. covering 90%-99.7% confidence interval in statistics, obtained according to statistical properties of normal distribution; Then, the policy network is iteratively optimized by using a proximal policy optimization algorithm, and a real-time adjustment coefficient is output: wherein, is the current state vector; denotes a proximal policy optimization algorithm; The dynamic adaptive thresholds are obtained based on a reinforcement learning algorithm, and the calculation formula is: wherein, is a dynamic adaptive threshold value, is a prediction of the cell performance parameter at time t, is a standard deviation, is an adjustment factor; The extracted feature values are compared with dynamic adaptive thresholds obtained based on a dynamic threshold adjustment mechanism to obtain a cell health state evaluation result, and an abnormal early warning signal is generated and a potential fault point is located according to the cell health state evaluation result, which specifically includes: When the extracted feature values exceed the set dynamic adaptive thresholds, an abnormal early warning signal is generated; The positioning algorithm is used to determine the potential fault point, which includes: based on the differences in operation features between cells, the clustering analysis method is used to distinguish abnormal cells from normal cells, thereby determining the location of the abnormal cells; The abnormal early warning signal is classified and maintenance suggestions are generated, and an abnormal record is generated and stored in a cloud database.

2. The method of claim 1, wherein the method is based on BMS data driving. Multi-dimensional cell operation data is collected by using various sensors and the BMS, specifically including: The sensors used include temperature sensors, voltage sensors, current sensors, and environmental humidity sensors; the collected data includes cell temperature, single cell voltage, charge and discharge current, and environmental humidity.

3. The method of claim 1, wherein the method is based on BMS data driving. The collected data is preprocessed to form a basic data set, specifically including: The collected raw data is cleaned and preliminarily classified to eliminate noise points and outliers; wherein, the data cleaning includes removing noise by using a sliding window filtering algorithm; the preliminary classification includes dividing the collected data into static basic data and dynamic operation data according to the data attributes.

4. The method of claim 1, wherein the method is based on BMS data driving. The abnormal early warning signal is classified and maintenance suggestions are generated, specifically including: The abnormal early warning signal is divided into three levels according to the severity of the abnormality: slight abnormality, moderate abnormality and severe abnormality; Wherein, the classification standard is established based on the degree of deviation of the abnormal feature value from the dynamic adaptive threshold, and the specific formula is: wherein, is an abnormality feature value, is a dynamic adaptive threshold, is a deviation percentage; when is determined to be a slight abnormality, is determined to be a moderate abnormality, is determined to be a severe abnormality. 5.A system for determining abnormality of a battery cell based on BMS data, characterized in that, The system comprises: A data collection module for collecting multi-dimensional cell operation data by using various sensors and the BMS, and preprocessing the collected data to form a basic data set; The intelligent analysis module is configured to perform data dimension reduction and feature extraction on the obtained basic data set, compare the extracted feature values with dynamic adaptive thresholds obtained based on a dynamic threshold adjustment mechanism to obtain a battery health state evaluation result, generate an abnormal early warning signal and locate a potential fault point according to the battery health state evaluation result. The data dimension reduction and feature extraction on the obtained basic data set specifically include: Based on the principal component analysis method, and by introducing a dynamic feature weighting and a physical constraint mechanism, the multi-dimensional data is processed by dimension reduction; The dynamic feature weighting specifically includes: Different weights are given according to the importance difference of different time sequence characteristics, and a dynamic weight vector is generated The original feature matrix is weighted, and the formula is represented as wherein, is the original timing window feature matrix, is a weight diagonal matrix, is the weighted feature matrix; The physical constraint mechanism specifically includes: The physical constraint term is embedded in the covariance matrix, and the formula is as follows: In the formula, is the covariance matrix finally obtained after dynamic characteristic weighting processing, which embeds a physical constraint term; is a weighted mean matrix; N is the number of samples; is a physical constraint matrix, and satisfies: if the characteristics and the characteristics are physically associated, the element in the i-th row and the j-th column of the matrix C is =1, otherwise =0; is a constraint coefficient; The covariance matrix is calculated as follows Eigenvalue decomposition is performed to obtain a plurality of eigenvalues and eigenvectors, and the first eigenvectors are selected to construct a projection matrix , and finally a reduced matrix is obtained : Thus, the representative key operating features are extracted; The extracted feature values are compared with dynamic adaptive thresholds obtained based on a dynamic threshold adjustment mechanism to obtain a battery health state evaluation result, and the battery health state evaluation result is used to generate an abnormal early warning signal and locate a potential fault point, specifically including: The dynamic threshold adjustment mechanism based on reinforcement learning is used to obtain the dynamic adaptive thresholds, including: Introducing deep reinforcement learning to train regulation coefficients , as a function of time t; reward function is to minimize false positive rate of state of health estimation; state space is current operating parameters and prediction error wherein is the predicted value of the battery performance parameter at time t, is the measured value of the battery performance parameter at time t; action space is i.e. covers 90%-99.7% confidence interval in statistics, obtained according to statistical properties of normal distribution; Then, the policy network is iteratively optimized by using a proximal policy optimization algorithm, and a real-time adjustment coefficient is outputted: wherein, is the current state vector; denotes a proximal policy optimization algorithm; The dynamic adaptive thresholds are obtained based on the reinforcement learning algorithm, and the calculation formula is as follows: wherein, is a dynamic adaptive threshold value, is a prediction of the cell performance parameter at time t, is a standard deviation, is an adjustment factor; The extracted feature values are compared with dynamic adaptive thresholds obtained based on a dynamic threshold adjustment mechanism to obtain a battery health state evaluation result, and the battery health state evaluation result is used to generate an abnormal early warning signal and locate a potential fault point, specifically including: When the extracted feature values exceed the set dynamic adaptive thresholds, an abnormal early warning signal is generated; The positioning algorithm is used to determine the potential fault point, including: based on the operating feature difference between the batteries, the clustering analysis method is used to distinguish the abnormal battery from the normal battery, and thus the position of the abnormal battery is determined; The maintenance decision module is configured to perform abnormal classification on the abnormal early warning signal and generate a maintenance suggestion, generate an abnormal record and store the abnormal record to a cloud database.

Images