A battery pack fault early warning method, device and equipment

Abstract

The application discloses a battery pack fault early warning method, device and equipment, comprising: obtaining historical operation data, dividing the historical operation data to obtain a plurality of data sets, constructing data features according to the data sets, and calculating a feature mean vector and a covariance matrix according to the data features; constructing a multivariate Gaussian distribution model according to the feature mean vector and the covariance matrix, and calculating a first probability density corresponding to each data set; setting an abnormality judgment threshold according to the first probability density, and performing abnormality judgment and early warning on current operation data. The application constructs data features through historical operation data of any time period, analyzes the probability density of the historical operation data and the judgment threshold of system failure through the multivariate Gaussian distribution model, and avoids the problems that the failure early warning is not timely and the early warning efficiency is low due to the analysis, investigation and early warning through single data when performing abnormality judgment early warning.

Description

Technical Field

[0001] This invention relates to the field of battery pack technology, and in particular to a battery pack fault early warning method, device, and equipment. Background Technology

[0002] Currently, a considerable number of secondary battery fault early warning methods have been proposed and implemented. Among them, online battery anomaly early warning based on big data can provide more comprehensive and timely diagnostic functions compared to traditional offline inspections. Of the many data-based anomaly analysis methods, statistical methods are often widely used in fault diagnosis and battery pack safety management due to their simplicity, convenience, and interpretability. However, these methods are mainly based on the voltage data of individual cells in the battery pack and cannot accurately diagnose faults and anomalies caused by changes in other parameters such as current, temperature, or even power demand. Summary of the Invention

[0003] This invention provides a battery pack fault early warning method, apparatus, and device to solve the technical problem that existing battery pack early warning schemes rely on single data for early warning, resulting in untimely fault warnings and low warning efficiency.

[0004] To address the aforementioned technical problems, embodiments of the present invention provide a battery pack fault early warning method, comprising:

[0005] Acquire historical operating data, divide the historical operating data into several datasets, construct data features based on the datasets, and calculate the feature mean vector and covariance matrix based on the data features. The historical operating data includes multiple types of data.

[0006] A multivariate Gaussian distribution model is constructed based on the feature mean vector and covariance matrix, and the first probability density corresponding to the dataset is calculated.

[0007] An anomaly detection threshold is set based on the first probability density, and anomaly detection and warning are issued for the current running data.

[0008] This invention constructs data features using historical operating data from any time period and analyzes the probability density of the historical operating data and the system fault determination threshold using a multivariate Gaussian distribution model. By analyzing the overall operating data, it avoids the problems of untimely fault warnings and low warning efficiency caused by analyzing and troubleshooting based on single data points. Furthermore, by analyzing historical operating data of the battery pack at any time period, it overcomes the limitation of relying on data from specific charging and discharging periods for anomaly judgment, thus improving the versatility of anomaly judgment.

[0009] Furthermore, the historical operating data includes voltage data, which includes at least one of the following: voltage data of each individual battery cell at each moment, total voltage within the battery pack, average voltage, voltage difference within the battery pack, and highest and lowest individual cell voltages within the battery pack.

[0010] The historical operating data also includes at least one of the following: battery pack temperature data, system current, and system displayed SOC.

[0011] Furthermore, the step of dividing each first data set into several datasets based on its time-scale distribution characteristics specifically involves:

[0012] The distribution critical value of each first dataset is obtained based on the time scale distribution characteristics.

[0013] Furthermore, the historical operational data includes multiple first data sets, each first data set including one type of data; the step of dividing the historical operational data into several datasets specifically involves:

[0014] Based on the time-scale distribution characteristics of the reference data, a distribution threshold for the reference data is determined, and the first dataset of the reference data is divided into multiple second datasets using the distribution threshold, wherein the reference data is one type of the historical running data;

[0015] Anomaly detection algorithms are used to clean the data of each second dataset to obtain the normal and abnormal datasets of each second dataset.

[0016] Determine the time scale distribution of each group of normal and abnormal datasets, and divide the other first datasets into normal and abnormal datasets according to the time scale distribution.

[0017] This invention analyzes different types of data in historical operational data by dividing each first data set in the historical operational data according to the distribution characteristics of the time scale, so as to avoid the interaction of different features affecting the accuracy of anomaly judgment.

[0018] Furthermore, determining the distribution critical value of the reference data based on its time-scale distribution characteristics includes:

[0019] Multiple data points of the reference data are used as the critical values ​​of the hypothetical distribution, and the reference data is divided into various indicator datasets using the critical values ​​of the hypothetical distribution.

[0020] Perform a KS test on each indicator dataset assuming a Gaussian distribution, and sum the P-values ​​of the KS tests for each indicator dataset to obtain P1. Determine the critical value of the hypothetical distribution corresponding to the maximum P1 as the critical value of the distribution of the reference data.

[0021] Further, the step of constructing data features based on the dataset and calculating the feature mean vector and covariance matrix based on the data features specifically involves:

[0022] For each time scale of the second dataset, calculate the feature mean vector and covariance matrix of the normal dataset and the abnormal dataset;

[0023] The step of constructing a multivariate Gaussian distribution model based on the feature mean vector and covariance matrix, and calculating the first probability density corresponding to the dataset, specifically involves:

[0024] A multivariate Gaussian distribution model is constructed based on the first density formula, combined with the characteristic mean vector and covariance matrix.

[0025] Using the aforementioned multivariate Gaussian distribution model, the first probability density of normal datasets and the first probability density of abnormal datasets of various data types are calculated for each time scale.

[0026] This invention constructs a multivariate Gaussian distribution model based on the feature mean vector and covariance matrix, thereby calculating the joint probability density and introducing anomaly detection from historical operating data. It utilizes different features to comprehensively assess the system's operational safety status, identify system faults, and achieve rapid early warning.

[0027] Furthermore, the step of setting an anomaly determination threshold based on the first probability density and performing anomaly determination on the current running data specifically involves:

[0028] A binary classifier is trained based on a classification algorithm. For each time scale of the second data set, the first probability density of the normal dataset and the first probability density of the abnormal dataset are compared based on the binary classifier to obtain the anomaly judgment threshold corresponding to the time scale of each second data set.

[0029] Anomaly detection and early warning are performed on the current running data based on the anomaly detection threshold corresponding to the time scale of each second data set.

[0030] Furthermore, the step of performing anomaly detection and early warning on the current running data based on the anomaly detection threshold corresponding to the time scale of each second data set specifically involves:

[0031] Obtain the current running data, and determine the second probability density of the current running data and the corresponding anomaly detection threshold;

[0032] The second probability density is compared with the corresponding anomaly determination threshold. If the second probability density is less than the corresponding anomaly determination threshold, the system is operating abnormally and an early warning is issued.

[0033] This invention trains a binary classifier using a classification algorithm, comparing normal and abnormal data in any historical operational dataset to obtain an anomaly detection threshold. This eliminates the need to rely on data from specific charging / discharging periods for analysis, improving the versatility of anomaly detection. Furthermore, it can quickly identify system anomalies based on the second probability density of the current operational data and the anomaly detection threshold, improving the real-time performance of anomaly detection and avoiding detection delays.

[0034] Secondly, embodiments of the present invention also provide a battery pack fault early warning device, including: a feature extraction module, a density calculation module, and an anomaly determination module;

[0035] The feature extraction module is used to acquire historical running data, divide the historical running data into several datasets, construct data features based on the datasets, and calculate the feature mean vector and covariance matrix based on the data features.

[0036] The density calculation module is used to construct a multivariate Gaussian distribution model based on the feature mean vector and covariance matrix, and to calculate the first probability density corresponding to each dataset.

[0037] The density calculation module is used to set an anomaly detection threshold based on the first probability density, and to perform anomaly detection and early warning on the current running data.

[0038] Thirdly, embodiments of the present invention also provide a computer device, including: a processor, a communication interface, and a memory, wherein the processor, the communication interface, and the memory are interconnected, wherein the memory stores executable program code, and the processor is used to call the executable program code to execute the battery pack fault warning method. Attached Figure Description

[0039] Figure 1 This is a schematic flowchart of a battery pack fault early warning method provided in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram illustrating the division of individual cell voltage data in the battery pack fault early warning method provided in this embodiment of the invention;

[0041] Figure 3 This is a schematic diagram of a battery pack fault early warning device provided in an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1

[0044] Please refer to Figure 1 , Figure 1 A schematic flowchart of a battery pack fault early warning method provided in an embodiment of the present invention includes steps 101 to 103, as detailed below:

[0045] Step 101: Obtain historical running data, divide the historical running data into several datasets, construct data features based on the datasets, and calculate the feature mean vector and covariance matrix based on the data features;

[0046] In this embodiment, if the battery pack is currently running or can still operate normally, when detecting whether a fault exists, its operating data over a set time period, i.e., historical operating data, can be obtained when the battery pack is running or in sleep mode. Historical operating data includes multiple types of data, including two or more types. In one possible implementation, the historical operating data includes at least voltage data, which should include at least one of the following: individual cell voltage data at each moment, total voltage within the battery pack, average voltage, battery pack differential voltage, and the highest and lowest individual cell voltages in the battery pack. The battery pack includes multiple individual cells. In addition to voltage data, the historical operating data should include at least one of the following: battery pack temperature data, battery pack current data, and battery pack displayed SOC. The battery pack temperature data should include at least one of the following: temperature at each sampling point at each moment, average battery pack temperature, battery pack temperature difference, and the highest and lowest battery pack temperatures. The types of historical operating data described above are merely illustrative examples. In actual vehicle operation, other data may be included to further determine multiple fault types; this embodiment is not limited to these.

[0047] In this embodiment, the historical operational data includes multiple first data sets, each first data set including one type of data; the historical operational data is divided into several datasets, including:

[0048] Based on the time-scale distribution characteristics of the reference data, the distribution critical value of the reference data is determined, and the first dataset of the reference data is divided into multiple second datasets using the distribution critical value.

[0049] Anomaly detection algorithms are used to clean the data of each second dataset to obtain the normal and abnormal datasets of each second dataset.

[0050] Determine the time-scale distribution of each normal and abnormal dataset, and then divide the other first datasets into normal and abnormal datasets according to the time-scale distribution.

[0051] The reference data is any one of the historical operating data selected, that is, data from a selected first data set. In this embodiment, each first data set includes any one type of data from the historical operating data. For example, when the historical operating data includes both voltage data of individual cells and average battery pack temperature, the historical operating data includes two first data sets: a data set of voltage data of individual cells and a data set of average battery pack temperature. Either the voltage data of individual cells or the average battery pack temperature can be selected as the reference data. Similarly, when the historical operating data includes other types of data, this will not be elaborated further. As a specific example of this embodiment, the historical operating data includes several first data sets, and each first data set includes one of the following: average battery pack temperature, battery pack temperature difference, battery pack pressure difference, highest individual cell voltage in the battery pack, and total internal pressure of the battery pack. Then, based on the time scale distribution characteristics, each first data set is divided into several datasets.

[0052] In this embodiment, taking lithium-ion batteries as an example, the charge-discharge characteristics and randomness of experimental statistics of lithium-ion batteries indicate that, given a sufficient number of samples during operation, the sample characteristics within a specific range will approximately follow a Gaussian distribution. Therefore, the principle for dividing historical operating data is that the distribution of each indicator in each state data after division approximates a Gaussian distribution. The distribution critical value for dividing the data is obtained through big data analysis of historical operating data. The first data set in the historical operating data is divided into several second datasets using the distribution critical value.

[0053] In one possible implementation, determining the distribution critical value of the reference data based on its time-scale distribution characteristics includes:

[0054] Multiple data points of the reference data are used as the critical values ​​of the hypothetical distribution, and the reference data is divided into various indicator datasets using the critical values ​​of the hypothetical distribution.

[0055] Perform a KS test on each indicator dataset assuming a Gaussian distribution, and sum the P-values ​​of the KS tests for each indicator dataset to obtain P1. Determine the critical value of the hypothetical distribution corresponding to the maximum P1 as the critical value of the distribution of the reference data.

[0056] The reference data is individual cell voltage data, which includes charging data and discharging data. The distribution of charging data during the charging process and discharging data during the discharging process is extracted from the first dataset of individual cell voltage data. Each value D in the charging and discharging data is used as a hypothetical distribution critical value to divide the first dataset. The first dataset is then divided using these hypothetical distribution critical values ​​to obtain the divided datasets of each index. Next, a KS test assuming a Gaussian distribution is performed on each index dataset, and the p-values ​​of the KS tests for each index dataset are summed to obtain P1. Finally, the hypothetical distribution critical value at which P1 is maximized is selected as the distribution critical value for dividing the individual cell voltage data.

[0057] For example, please refer to Figure 2 , Figure 2 This diagram illustrates the partitioning of individual cell voltage data in the battery pack fault warning method provided in this embodiment of the invention. As a specific example of this embodiment, when partitioning data using individual cell voltage data (i.e., the first data set) from historical operating data, the critical value for the distribution of charging data is determined to be 3.3V, and the critical value for the distribution of discharging data is determined to be 3.35V. Therefore, the individual cell voltage data X is divided into four second datasets: X... chgl X chgs X dchl X dchs Among them, X chgl This refers to the dataset of moments during the charging process where the individual cell voltage is ≥3.3V; X chgs This refers to the dataset of moments during the charging process when the cell voltage is <3.3V; X dchl This refers to the dataset of moments during the discharge process where the cell voltage is <3.35V; X dchs This refers to the dataset of moments during discharge when the individual cell voltage is ≥3.35V, where the individual cell voltage can be calculated from the total internal voltage of the battery pack. It should be noted that the four second datasets mentioned above can originate from a single energy storage terminal or be a collection of multiple terminals partitioned according to this classification standard.

[0058] In this embodiment, the cell voltage can be either the average cell voltage or the cell voltage data of a single cell in the battery pack.

[0059] After splitting the dataset into two sub-datasets, anomaly detection algorithms are used to clean the data in each sub-dataset. Data cleaning methods include, but are not limited to, empirical knowledge, the 3σ rule, upper and lower k-quantiles, or other machine learning anomaly detection methods. Through data cleaning, the normal and anomalous datasets in each sub-dataset can be identified.

[0060] Taking the above example of dividing the individual voltage data into four second datasets, each second dataset is further divided into normal datasets and abnormal datasets. Therefore, the individual voltage data can be divided into 4 groups of normal datasets and abnormal datasets.

[0061] After determining the normal and abnormal datasets for each second dataset in the individual voltage data, the time scale distribution of each set of normal and abnormal datasets is determined, and the other first datasets are divided into normal and abnormal datasets according to the time scale distribution.

[0062] For example, the time scale for dividing the individual voltage data into each second dataset can be determined. Then, other first datasets in the historical data can be divided into multiple second datasets according to this time scale. Furthermore, each second dataset can be divided into normal datasets and abnormal datasets according to the time scale distribution of the normal and abnormal datasets of the individual voltage data within that dataset. In other possible implementations, the other data in the first dataset can also be directly divided into multiple normal datasets and multiple abnormal datasets according to the time scale distribution of the normal and abnormal datasets of each second dataset in the individual voltage data.

[0063] For example, historical operating data includes individual cell voltage data, average battery pack temperature T, battery pack temperature difference ΔT, battery pack voltage difference ΔU, and highest individual cell voltage U. max Taking the time span of the historical operating data as b as an example, the individual voltage data and other types of data in the historical operating data include data at times t0, t1, t2...tb. Specifically, taking b as 20 as an example, the individual voltage data includes t0, t1, t2, ..., t19, t20. The time scale distribution for dividing the data into the second dataset is determined by the individual voltage data as follows: First time scale - [t0, t1, t2, t11, t12, t13], Second time scale - [t3, t4, t5, t14, t15, t16], Third time scale - [t6, t7, t8, t17, t18], Fourth time scale - [t9, t10, t19, t20]. Each second dataset is further divided into normal and abnormal datasets. Taking the first time scale of the second dataset, i.e., [t0, t1, t2, t11, t12, t13], as an example, the time scale distribution of the normal dataset is the first sub-time scale [t0, t11, t12, t13], and the time scale distribution of the abnormal dataset is the second sub-time scale [t1, t2].

[0064] Then, the average temperature T of the battery pack, the temperature difference ΔT of the battery pack, the voltage difference ΔU of the battery pack, and the highest single-cell voltage U of the battery pack are respectively calculated. maxData within the first, second, third, and fourth time scales is divided into four second datasets for each data type. These are then further divided into normal and abnormal datasets based on sub-time scales. For example, the data within the second dataset corresponding to the first time scale is further classified as the normal dataset, and the data within the second dataset corresponding to the second sub-time scale is classified as the abnormal dataset. This process continues until the normal and abnormal datasets for each second dataset are obtained. This embodiment is only used as an example and is not limited to this specific implementation.

[0065] It should be noted that the embodiments of this application only use the selection of single-unit voltage data as a reference data for illustration. In other possible implementations, other data can also be used as reference data, and the first data set can be divided into several datasets according to the time scale distribution characteristics. Other methods can also be used to achieve this.

[0066] In this embodiment, after determining the normal and abnormal datasets in each second dataset, the step of constructing data features based on the datasets and calculating the feature mean vector and covariance matrix based on the data features specifically involves:

[0067] For each time scale of the second dataset, calculate the feature mean vector and covariance matrix of the normal dataset and the abnormal dataset.

[0068] Each second data set is a second data set partitioned from the reference data. Taking the four second data sets X of the aforementioned individual voltage data as an example... chgl X chgs X dchl and X dchs Data cleaning was performed on all datasets, and normal and abnormal datasets were obtained for each dataset. Examples are given using the second dataset X from the individual voltage data. chgl For example, the second dataset X chgl After cleaning, the normal dataset X in this state is obtained. chgl-norm and the abnormal dataset X chgl-abn The following explanation uses the calculation of the feature mean and covariance matrix of a normal dataset as an example. Normal dataset X chgl-norm The time scale of the operational data samples is d. For each time t in the time scale d of the historical operational data samples, the average battery pack temperature T, battery pack temperature difference ΔT, battery pack voltage difference ΔU, and highest single-cell voltage U of the battery pack are obtained or calculated for each time t. max From the data, a feature vector is obtained, and the feature vector is specifically:

[0069]

[0070] Then for the entire normal dataset X chgl-norm The time scale d, and the feature matrix are as follows:

[0071]

[0072] The eigenvalue mean vector and covariance matrix are as follows:

[0073]

[0074] Among them, the characteristic mean Feature Sample Variance Two-feature sample covariance Let i be the value of feature i at time t. Let M be the mean of feature i over time scale d. pt Let p be the value of the feature at time t. Let M be the mean of feature p over time scale d. qt Let q be the value of the feature at time t. Let q be the mean of feature q over time scale d, where feature i, feature q, and feature p all refer to the data types of historical operating data. Taking the above as an example, i represents the average temperature T of the battery pack, or the temperature difference ΔT of the battery pack, or the pressure difference ΔU of the battery pack, or the highest single-cell voltage U of the battery pack. max .

[0075] Following the above method, calculate the historical operating data respectively in X chgl X chgs X dchl and X dchs The feature mean vectors and covariance matrices over the time scales of the normal and abnormal datasets can be used to obtain 8 sets of feature mean vectors and covariance matrices.

[0076] In this embodiment of the application, after determining the feature mean vector and covariance matrix of the historical running data, step 102 is further included: constructing a multivariate Gaussian distribution model based on the feature mean vector and covariance matrix, and calculating the first probability density corresponding to the dataset.

[0077] In one possible implementation, the step of constructing a multivariate Gaussian distribution model based on the feature mean vector and covariance matrix, and calculating the first probability density corresponding to the dataset, specifically involves:

[0078] A multivariate Gaussian distribution model is constructed based on the first density formula, combined with the feature mean vector and covariance matrix. Using the multivariate Gaussian distribution model, the first probability density of normal datasets and the first probability density of abnormal datasets of various data types are calculated for each time scale.

[0079] In this embodiment, the second dataset X is used as an example. chgl Normal dataset X in chgl-norm μ is calculated on the corresponding time scale chgl , Σ chgl Taking μ as an example for illustration chgl Σ chgl Substituting the first density formula to construct a multivariate Gaussian distribution model, and then applying it to the normal dataset X. chgl-norm Substituting a normal dataset within the same time scale into a Gaussian distribution model yields the first probability density of the normal dataset at that time scale:

[0080]

[0081] Where n is the sample size. norm For historical running data and normal dataset X chgl-norm Data sets at the same time scale, p chgl (x norm μ chgl ,Σ chgl ) is a normal dataset X chgl-norm The first probability density of the corresponding time scale.

[0082] In this embodiment, the method for calculating the first probability density of the abnormal dataset can refer to the method for calculating the first probability density of the normal dataset, and will not be repeated here. Referring to the above method, the second dataset X also needs to be... chgs X dchl and X dchs The first probability density is calculated for the data within the time scale corresponding to the normal dataset and the abnormal dataset, and finally four sets of first probability densities are obtained for the four second datasets. Each set of first probability densities includes the first probability density of the normal dataset and the first probability density of the abnormal dataset.

[0083] In this embodiment of the application, after calculating the first probability density corresponding to the dataset, step 103 is further included: setting an anomaly judgment threshold according to the first probability density, and performing anomaly judgment and warning on the currently running data.

[0084] In this embodiment, the step of setting an anomaly detection threshold based on the first probability density and performing anomaly detection and early warning on the current running data specifically includes:

[0085] A binary classifier is trained based on a classification algorithm. For each time scale of the second data set, the first probability density of the normal dataset and the first probability density of the abnormal dataset are compared based on the binary classifier to obtain the anomaly judgment threshold corresponding to the time scale of each second data set.

[0086] Anomaly detection and early warning are performed on the current running data based on the anomaly detection threshold corresponding to the time scale of each second data set.

[0087] In this embodiment, taking the above example, the time scales of the second dataset are the first time scale, the second time scale, the third time scale, and the fourth time scale. In this embodiment, a simple binary classifier is used to determine the anomaly detection threshold. The classification algorithm for training the binary classifier includes, but is not limited to, logistic regression, SVM, and neural networks. Taking the second dataset X... chgl Taking the corresponding time scale as an example, the input data of the binary classifier are the first probability density of the normal dataset and the first probability density of the abnormal dataset within that time scale, with labels of normal (1) and abnormal (0), respectively. After training, the binary classifier obtains its weight parameters to get the anomaly judgment threshold. When the sample size difference between the abnormal dataset and the normal dataset is too large, upsampling and oversampling methods can be used to balance the sample size. Referring to the above method, the anomaly judgment threshold corresponding to the time scale of each second dataset is finally obtained. Taking the above as an example, the historical running data at X are obtained respectively. chgl X chgs X dchl and X dchs Anomaly detection thresholds at different time scales. Then, based on the anomaly detection thresholds corresponding to the time scale of each second data set, anomaly detection and early warning are performed on the current running data.

[0088] In this embodiment, the step of performing anomaly detection and early warning on the current running data based on the anomaly detection threshold corresponding to the time scale of each second data set specifically includes:

[0089] Obtain current operating data, determine the second probability density of the current operating data and the corresponding anomaly judgment threshold; compare the second probability density with the corresponding anomaly judgment threshold, and if the second probability density is less than the corresponding anomaly judgment threshold, the system is operating abnormally and an early warning is issued.

[0090] For example, if the current operating data shows a single-cell voltage ≥ 3.3V during charging, then the anomaly detection threshold for the current operating data is the historical operating data at X... chgl The anomaly detection threshold at the time scale, that is, the anomaly detection threshold corresponding to the first time scale.

[0091] It should be noted that the data types contained in the current operating data and the historical operating data are consistent. The calculation of the second probability density can refer to the method used to calculate the first probability density, which will not be repeated in the embodiments of this application. Taking the above as an example, when the second probability density is less than the anomaly judgment threshold corresponding to the first time scale, it is determined that the system is faulty at that moment; otherwise, it is normal. When the system is faulty, an early warning is issued according to a preset method.

[0092] In this embodiment, a multivariate Gaussian distribution model is constructed based on the feature mean vector and covariance matrix to calculate the joint probability density. This allows for a comprehensive assessment of the system's operational safety status using different features. Furthermore, based on the second probability density of the current operating data and the anomaly detection threshold, system anomalies can be quickly identified, improving the real-time performance of anomaly detection, avoiding delayed judgments of system faults, and enabling rapid early warning.

[0093] To further improve the accuracy of fault warnings, the binary classifier can be iteratively trained using a second probability density. Specifically, the second probability density and its labels are input into the binary classifier. To ensure reliability, when determining the labels for the second probability density, samples with second probability densities near the anomaly threshold are manually labeled. The labeling criteria are: all samples classified as anomalous are labeled as anomalous; for samples classified as normal, a preset number or proportion of samples are labeled as normal, and the rest are labeled as anomalous.

[0094] In this embodiment, by periodically updating and iterating the binary classifier, the anomaly detection threshold is updated, thereby avoiding anomaly detection errors caused by system instability and improving the accuracy of anomaly detection.

[0095] In this embodiment, after determining the anomaly of the current running data, the following steps are also included:

[0096] If a system malfunction is detected, the system acquires the operating data within a preset interval at the time of the malfunction and uses this data to determine the malfunction type.

[0097] This application embodiment does not impose specific restrictions on the preset interval, which can be selected according to the actual situation. When using the operating data within the preset interval at the time of the fault to determine the fault type, the abnormal data type can be determined by comparing each type of data in the operating data with the standard value of each type of data, and then the fault type can be determined.

[0098] In this embodiment, taking the abnormal temperature data as an example, if t1 is the fault time, then all operating data near time t1 are retrieved, and each operating data is compared with the standard value. When the temperature data is found to be abnormal and does not match the standard value, the system abnormality type can be determined to be temperature. Then, based on the cause of the abnormal temperature data and the time of the abnormal temperature data, the system fault type can be accurately located, thereby improving the accuracy of fault judgment.

[0099] In this embodiment, data features are constructed using historical operating data from any time period, and the probability density of the historical operating data is analyzed using a multivariate Gaussian distribution model. This addresses the problem of different features interacting and affecting anomaly detection, improving the accuracy of anomaly detection under multidimensional data and avoiding missed detections. Furthermore, analyzing historical operating data of the battery pack at any time period overcomes the limitation of anomaly detection relying on data from specific charging and discharging periods, improving the versatility of anomaly detection.

[0100] Please refer to Figure 3 , Figure 3 A schematic diagram of a battery pack fault early warning device provided in an embodiment of the present invention includes: a feature extraction module 401, a density calculation module 402, and an anomaly determination module 403;

[0101] The feature extraction module 401 is used to acquire historical running data, construct data features from the historical running data, and calculate the feature mean vector and covariance matrix based on the data features.

[0102] The density calculation module 402 is used to construct a multivariate Gaussian distribution model based on the feature mean vector and covariance matrix, and to calculate the first probability density corresponding to each dataset.

[0103] The anomaly determination module 403 is used to set an anomaly determination threshold based on the first probability density, and to determine and warn of anomalies in the current running data.

[0104] This invention also provides a computer device, including: a processor, a communication interface, and a memory, wherein the processor, the communication interface, and the memory are interconnected, wherein the memory stores executable program code, and the processor is used to call the executable program code to execute the battery pack fault warning method.

[0105] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A battery pack failure early warning method, characterized in that, include: The process involves: acquiring historical operational data, which includes multiple first datasets, each containing one type of data; determining a distribution threshold for the reference data based on its time-scale distribution characteristics, and using this threshold to divide the first datasets of the reference data into multiple second datasets, wherein the reference data is one type of data from the historical operational data; cleaning each second dataset using an anomaly detection algorithm to obtain normal and abnormal datasets for each second dataset; determining the time-scale distribution of each set of normal and abnormal datasets, and dividing the other first datasets into normal and abnormal datasets according to this time-scale distribution; For each time scale of the second dataset, calculate the feature mean vector and covariance matrix of the normal dataset and the abnormal dataset; A multivariate Gaussian distribution model is constructed based on the feature mean vector and covariance matrix, and the first probability density corresponding to each dataset is calculated; specifically: a multivariate Gaussian distribution model is constructed based on the first density formula combined with the feature mean vector and covariance matrix; using the multivariate Gaussian distribution model, the first probability density of normal datasets and the first probability density of abnormal datasets of various data types are calculated for each time scale. An anomaly detection threshold is set based on the first probability density, and anomaly detection and warning are issued for the current running data.

2. The battery pack failure early warning method of claim 1, wherein, The step of determining the distribution critical value of the reference data based on its time-scale distribution characteristics includes: Multiple data points of the reference data are used as the critical values ​​of the hypothetical distribution, and the reference data is divided into various indicator datasets using the critical values ​​of the hypothetical distribution. Perform a KS test on each indicator dataset assuming a Gaussian distribution, and sum the P-values ​​of the KS tests for each indicator dataset to obtain P1. Determine the critical value of the hypothetical distribution corresponding to the maximum P1 as the critical value of the distribution of the reference data.

3. The battery pack fault early warning method as described in claim 1, characterized in that, The step of setting an anomaly detection threshold based on the first probability density and performing anomaly detection and early warning on the current operating data specifically includes: A binary classifier is trained based on a classification algorithm. For each time scale of the second dataset, the first probability density of the normal dataset and the first probability density of the abnormal dataset are compared based on the binary classifier to obtain the anomaly judgment threshold corresponding to the time scale of each second dataset. Anomaly detection and early warning are performed on the current running data based on the anomaly detection threshold corresponding to the time scale of each second dataset.

4. The battery pack fault early warning method as described in claim 3, characterized in that, The process of determining and issuing warnings for anomalies in the current running data based on the anomaly determination threshold corresponding to the time scale of each second dataset specifically involves: Obtain the current running data, and determine the second probability density of the current running data and the corresponding anomaly detection threshold; The second probability density is compared with the corresponding anomaly determination threshold. If the second probability density is less than the corresponding anomaly determination threshold, the system is operating abnormally and an early warning is issued.

5. The battery pack fault early warning method as described in claim 1, characterized in that, After performing anomaly detection on the current running data, the following is also included: If a system malfunction is detected, the system acquires the operating data within a preset interval at the time of the malfunction and uses this data to determine the malfunction type.

6. The battery pack fault early warning method as described in claim 1, characterized in that, The historical operating data includes voltage data, which includes at least one of the following: voltage data of each individual battery cell at each moment, total voltage in the battery pack, average voltage, voltage difference in the battery pack, and the highest and lowest individual cell voltages in the battery pack. The historical operating data also includes at least one of the following: battery pack temperature data, system current, and system displayed SOC.

7. A battery pack fault early warning device, characterized in that, include: Feature extraction module, density calculation module, and anomaly detection module; The feature extraction module is used to acquire historical operational data, which includes multiple first data sets, each containing one type of data; determine the distribution threshold of the reference data based on its time-scale distribution characteristics, and use the distribution threshold to divide the first dataset of the reference data into multiple second datasets, wherein the reference data is one type of the historical operational data; perform data cleaning on each second dataset using an anomaly detection algorithm to obtain normal and abnormal datasets for each second dataset; determine the time-scale distribution of each set of normal and abnormal datasets, and divide the other first datasets into normal and abnormal datasets according to the time-scale distribution; For each time scale of the second dataset, calculate the feature mean vector and covariance matrix of the normal dataset and the abnormal dataset; The density calculation module is used to construct a multivariate Gaussian distribution model based on the feature mean vector and covariance matrix, and to calculate the first probability density corresponding to each dataset; specifically, it constructs a multivariate Gaussian distribution model based on the first density formula combined with the feature mean vector and covariance matrix; and uses the multivariate Gaussian distribution model to calculate the first probability density of normal datasets and abnormal datasets of various data types at each time scale. The density calculation module is used to set an anomaly detection threshold based on the first probability density, and to perform anomaly detection and early warning on the current running data.

8. A computer device, characterized in that, include: The system includes a processor, a communication interface, and a memory, which are interconnected. The memory stores executable program code, and the processor is used to call the executable program code to execute the battery pack fault warning method as described in any one of claims 1 to 7.

Images