A battery safety feature extraction method
By using a multi-dimensional data-driven approach to screen out battery safety features, this method solves the applicability problem of existing battery safety management methods in complex vehicle environments, achieves highly robust screening and online monitoring of battery safety features, and supports early safety warnings for electric vehicles.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-29
AI Technical Summary
Existing battery safety management methods are difficult to apply to real, complex, and ever-changing in-vehicle environments, and fail to effectively distinguish the differences in battery voltage characteristics between vehicle driving and charging states. This increases the difficulty of extracting battery safety features and makes it impossible to effectively support early safety warnings and online management of electric vehicles.
A multi-dimensional data-driven approach is adopted. The raw data of electric vehicles is acquired, preprocessed, and divided into driving and charging segments. A voltage matrix is constructed and basic statistical characteristics are calculated. The optimal battery safety characterization features, including mean and range, are screened using a multi-iteration Gaussian mixture model and hypothesis testing method. Multi-dimensional statistical analysis is then performed to verify its robustness.
It achieves highly robust screening of battery safety features in real-world environments, ensuring the effectiveness and applicability of selected features under different conditions, supporting the development of online battery safety monitoring and early warning systems, and providing a basis for differentiated management strategies.
Smart Images
Figure CN122109840A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electric vehicle battery safety management and fault early warning technology, specifically involving a data-driven battery safety feature extraction method, which is suitable for the identification and online monitoring of early battery safety hazards. Background Technology
[0002] With the increasing popularity of electric vehicles, battery safety issues are becoming increasingly prominent, especially serious faults such as battery thermal runaway that threaten the lives of occupants. Currently, various battery safety management methods and safety characterization features have been proposed, such as fault detection based on features like voltage residual, abnormal heating, sample entropy, and Z-value. However, most of these methods are based on laboratory conditions or single fault modes, and are not compared under unified standards, making them difficult to apply to the real, complex, and variable in-vehicle environment.
[0003] In the real world, the operation of electric vehicles is affected by a combination of factors, including season, mileage, state of charge (SOC), and driving behavior. Issues such as sensor errors and uneven data sampling intervals further increase the difficulty of extracting safety features. Furthermore, most existing studies fail to distinguish between vehicle driving and charging states, ignoring the fundamental differences in battery voltage characteristics between these two states.
[0004] The power battery is the energy source of electric vehicles, but it is also the part most prone to failure. With continuous development, safety issues such as battery thermal runaway are becoming more and more common. According to statistics from the China Automotive Technology Research Center, the latest statistics show that the fire accident rate of new energy vehicles has reached 0.09%, which is the most urgent problem to be solved in the process of battery development.
[0005] Therefore, there is an urgent need for a method that can extract and screen battery safety characterization features that have strong discriminative power, are applicable to multiple states, and conform to actual distribution characteristics, in order to support the development of early safety warning and online management systems. Summary of the Invention
[0006] The purpose of this invention is to provide a battery safety feature extraction method that can select the features that best represent the battery safety state from multiple dimensions and features, and verify its effectiveness and robustness in real complex environments.
[0007] The present invention adopts the following technical solution: Step 1: Obtain the original vehicle data of the electric vehicle, which includes information such as battery voltage, battery current, SOC, vehicle speed, and time.
[0008] Step 2: Data preprocessing. For the above raw vehicle data, perform data preprocessing.
[0009] Step 3: State Division: The preprocessed vehicle data is divided into "driving segment" and "charging segment" to account for the differences in battery voltage characteristics under different states.
[0010] Step 4: Statistical feature construction: For each driving segment or charging segment, construct the corresponding voltage matrix, including time step and number of batteries; Based on the voltage matrix, the following ten basic statistical characteristics are calculated: maximum value, minimum value, mean, variance, median, absolute energy, kurtosis, mean absolute change, sample entropy, and skewness. Statistical features, including range and variance, are calculated for each of the basic features, resulting in a total of twenty statistical features used to characterize the inconsistency of battery voltage.
[0011] Step 5: Safety Characteristic Feature Screening: Based on the aforementioned statistical features, the actual feature distribution is constructed using a multi-iteration Gaussian mixture model, and the optimal battery safety characterization features are screened from three dimensions using hypothesis testing methods. Dimension 1 involves constructing the actual distribution for each statistical feature using a multi-iteration Gaussian mixture model (MIGMM) and performing a Kolmogorov-Smirnov test against the theoretical Gaussian distribution. Dimension 2 involves performing a KS test on the feature distributions of normal and unsafe batteries to determine whether the features can significantly distinguish between them. Dimension 3 involves combining the screening results under both driving and charging conditions to select features that are effective in both states. Finally, the mean-range is selected as the optimal battery safety characterization feature.
[0012] Step 6: Multi-dimensional statistical analysis: Based on the optimal battery safety characterization features, establish a statistical dimension analysis model, and perform feature verification and statistical analysis from multiple dimensions; use data from other normal vehicles and vehicles that have experienced thermal runaway for verification, and perform feature verification and multi-dimensional statistical analysis based on the selected mean range, and establish a statistical dimension analysis model.
[0013] The multiple dimensions include: first, the seasonal dimension, analyzing the distinguishing ability of features under different temperature conditions; second, the mileage dimension, analyzing the impact of battery health status on features; third, the SOC dimension, analyzing the performance of features under different states of charge; and fourth, the vehicle state dimension, analyzing the applicability of features under driving and charging conditions. Through statistical distribution comparison and hypothesis testing, the robustness of the selected features under multidimensional conditions is verified, and common patterns are summarized to support the formulation of battery safety management strategies.
[0014] Compared with existing technologies, this invention combines big data mining and machine learning methods, and proposes a systematic battery safety feature selection framework using a data-driven approach. It is the first to combine Gaussian distribution fit, battery state differentiation, and multi-state applicability to ensure the selected features have high robustness and practicality. A multi-iterative Gaussian mixture model is used to address noise and outliers in the data, improving the accuracy of feature distribution modeling. Analysis of the processing results yields conclusions with more practical guiding significance, supporting the development of online battery safety monitoring and early warning systems. Furthermore, through multi-dimensional statistical analysis, it reveals the variation patterns of battery safety features under different seasons, SOC, mileage, and states, providing a basis for differentiated safety management strategies. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the overall process of the battery safety feature extraction method described in this invention; Figure 2 A schematic diagram illustrating the method for distinguishing vehicle status; Figure 3 For security characterization feature screening process; Detailed Implementation
[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0017] like Figure 1 As shown, a battery safety feature extraction method includes the following steps: Step 1: Obtain the original vehicle data of the electric vehicle, which includes information such as battery voltage, battery current, SOC, vehicle speed, and time.
[0018] Step 2: Data preprocessing.
[0019] For the above raw vehicle data, missing and outlier values are processed, and noise is smoothed; data integration is performed by merging data from multiple data sources.
[0020] Step 3: State division.
[0021] The preprocessed vehicle data is divided into "driving segment" and "charging segment" to account for the differences in battery voltage characteristics under different conditions.
[0022] like Figure 2 As shown, the time series data is divided into states based on vehicle speed and battery current: if the vehicle speed is greater than 0 or the current is greater than 0, it is determined to be in the driving state; if the vehicle speed is less than or equal to 0 and the current is stable and negative, it is determined to be in the charging state; the time steps of consecutive states are merged into "segments", and finally the driving segment and charging segment are obtained. The thousands of driving segments and charging segments of each vehicle constitute the driving dataset and charging dataset, respectively.
[0023] Step 4: Statistical feature construction.
[0024] For each driving segment or charging segment, a voltage matrix R is constructed based on the time step and the number of batteries. k×n :
[0025] In the formula, P t,j Let be the voltage of the j-th battery cell at time step t; n is the number of battery cells in the battery pack; and k is the total time step for a certain driving or charging segment of the vehicle.
[0026] Based on the voltage matrix R above k×n Calculate ten basic statistical characteristics, including maximum, minimum, mean, variance, median, absolute energy, kurtosis, mean absolute change, sample entropy, and skewness. The formulas for calculating each basic statistical characteristic are as follows: (1) Maximum value:
[0027] (2) Minimum value:
[0028] (3) Mean:
[0029] (4) Variance:
[0030] (5) Median:
[0031] (6) Absolute energy:
[0032] (7) Kurtosis:
[0033] (8) Average absolute change:
[0034] (9) Sample entropy:
[0035] In the formula, For (P) 1, j , ..., P t, j , ..., P k, j The probability that two sequences match at a points in (A). a (r) is (P) 1, j , ..., Pt, j , ..., P k, j The probability that two sequences match at a+1 points; r is the tolerance window, described as:
[0036] (10) Skewness:
[0037] Furthermore, the range and variance of each feature were calculated, resulting in twenty statistical features. All features were calculated separately for driving and charging states, using the following formulas:
[0038]
[0039] In the formula, F represents the 10 basic statistical characteristics in (1) to (10) above; s F,r The range of characteristic F; s F,v It is the variance of characteristic F.
[0040] Step 5: Screening of security characterization features.
[0041] Feature screening is performed on data from normal vehicles and vehicles that have experienced thermal runaway. Based on the multiple statistical features, an actual feature distribution is constructed using a multi-iteration Gaussian mixture model. Combined with hypothesis testing methods, the optimal battery safety characterization features are selected from three dimensions.
[0042] The three dimensions are as follows: Dimension 1: For each statistical feature, a multi-iteration Gaussian mixture model (MIGMM) is used to construct the actual distribution, and a Kolmogorov-Smirnov test is performed against the theoretical Gaussian distribution. Features with a p-value > 0.05 are selected. The theoretical distribution is a Gaussian distribution, and its probability density function can be expressed as:
[0043] In the formula, s represents the twenty constructed statistical features. This represents the average value of s for different driving or charging segments. The standard deviation of different driving / charging segments s is calculated using the following formula:
[0044]
[0045]
[0046] In the formula s (i) Let be the characteristic of the i-th driving segment or charging segment. m is the total number of driving segments or charging segments for a normal electric vehicle.
[0047] The actual distribution is constructed using a multi-iteration Gaussian mixture model (MIGMM), whose probability density function is described as:
[0048] In the formula, W is the number of Gaussian models, and α w It represents the probability that a feature belongs to the w-th sub-model. ; The Gaussian distribution density function of the w-th sub-model is calculated by the following formula:
[0049] In the formula, These are the parameters of the w-th sub-model; μ w It is the mean of the w-th sub-model; It is the variance of the w-th sub-model.
[0050] Maximum likelihood function estimation and :
[0051] Based on the above probability density function and Gaussian distribution density function, Kolmogorov-Smirnov analysis is generally performed by establishing the null hypothesis that the empirical distribution conforms to the theoretical distribution, and the alternative hypothesis that it does not:
[0052] In the formula, H0 is the null hypothesis and H1 is the alternative hypothesis. Its KS test statistic for:
[0053] The p-value is an indicator for determining whether to accept or reject the null hypothesis in a KS test. Its calculation formula is as follows:
[0054] In the formula, Let α be the p-value. If the p-value is greater than the significance level α, according to hypothesis testing theory, there is insufficient reason to reject the null hypothesis, i.e., the null hypothesis is accepted, and the actual characteristic distribution of a normal battery follows a Gaussian distribution, thus the condition of dimension 1 is met. Statistically, the significance level is usually set to 0.05.
[0055] Dimension Two: Perform a KS test on the characteristic distributions of normal and unsafe batteries, selecting features with a p-value < 0.05; generally, the null hypothesis is that the two empirical distributions are not significantly different, and the alternative hypothesis is that they are significantly different.
[0056] In the formula, H0 is the null hypothesis, H1 is the alternative hypothesis, and P... UNS (s) is the cumulative probability density function of the actual distribution of unsafe batteries.
[0057] The MIGMM method is also used to construct the actual distributions of normal and unsafe batteries, and its probability density function is described as follows:
[0058]
[0059] In the formula, The probability density of a normal battery; This represents the probability density of unsafe batteries.
[0060] Based on the actual distribution constructed above, perform the KS test:
[0061] In the formula, Here is the KS test statistic for dimension two; m and Total number of driving / charging segments for normal electric vehicles and electric vehicles with unsafe batteries.
[0062] The formula for calculating the P-value is:
[0063] In the formula, This is the P-value for dimension two.
[0064] If the p-value is less than the significance level α, according to hypothesis testing theory, there is sufficient reason to reject the null hypothesis, that is, the feature can significantly distinguish between normal batteries and unsafe batteries.
[0065] Dimension three: Based on the screening results under both driving and charging conditions, select features that are effective in both conditions.
[0066] Ultimately, the mean range performed best across all three dimensions and was selected as the battery safety characterization feature.
[0067] Step Six: Multidimensional Statistical Analysis Validation was performed using data from other normal vehicles and vehicles that experienced thermal runaway. To verify the robustness of the mean range under different operating conditions, the circuit equations of the constructed battery Rint equivalent circuit model are as follows:
[0068]
[0069] In the formula, U oc，j U j R0，j I and I represent the battery's open-circuit voltage, terminal voltage, internal resistance in ohms, and current, respectively.
[0070] The open-circuit voltage is mainly affected by the battery's state of charge (SOC).
[0071] The ohmic internal resistance is mainly affected by ambient temperature, state of charge (SOC), and state of ohm (SOH).
[0072] In the formula, T is the ambient temperature; H is the battery SOH.
[0073] Therefore, the battery's terminal voltage can be expressed as:
[0074] Safety characteristics mean range s ME,r It is a mathematical transformation of the terminal voltage, s ME,r It can be described as:
[0075] s ME,r Due to the influence of ambient temperature, State of Health (SOH), State of Charge (SOC), and total current, a statistical analysis of the mean range was conducted from four dimensions: Ambient temperature affects battery performance, so season was selected as one of the dimensions for analysis, with analysis and verification conducted according to spring, summer, autumn, and winter; State of Health (SOH) affects the selected characteristics, which is expressed as mileage in actual electric vehicle operation, so mileage was selected as one of the dimensions for analysis, divided into six intervals for analysis and verification; State of Charge (SOC) affects the selected characteristics, so SOC was selected as one of the dimensions for analysis, divided into nine intervals for analysis and verification; Vehicle status also affects the selected characteristics, which can be divided into driving, fast charging, and slow charging for analysis and verification.
[0076] By comparing distributions and calculating P-values, it was verified that the mean range can effectively distinguish between normal and unsafe batteries across all dimensions, and the following patterns were summarized: the feature differentiation ability is stronger in summer, at high SOC, and under charging conditions; a uniform threshold can be used under different seasons and mileage; and it is recommended to use differentiated thresholds under different SOC and vehicle conditions.
[0077] This invention proposes a method for extracting and selecting battery safety features based on real vehicle data. Through multi-dimensional screening and verification, the effectiveness and robustness of the selected features in practical applications are ensured, providing reliable technical support for online monitoring and early warning of electric vehicle battery safety.
Claims
1. A method for extracting battery safety features, characterized in that, Includes the following steps: Step 1: Obtain the raw vehicle data of the electric vehicle; Step 2: Data Preprocessing: Perform data preprocessing on the raw vehicle data of the electric vehicle; Step 3: State Division: Divide the preprocessed vehicle data into two time series segments: driving segment and charging segment; Step 4: Statistical Feature Construction: For each driving segment or charging segment, extract multiple statistical features to characterize the inconsistency of battery voltage; Step 5: Safety Characteristic Feature Screening: Based on the aforementioned statistical features, the actual feature distribution is constructed using a multi-iteration Gaussian mixture model, and the optimal battery safety characterization features are screened from three dimensions using hypothesis testing methods. Step 6: Multi-dimensional statistical analysis: Based on the optimal battery safety characterization features, establish a statistical dimension analysis model to perform feature verification and statistical analysis from multiple dimensions.
2. The method according to claim 1, characterized in that, The raw vehicle data in step S1 includes information on battery voltage, battery current, SOC, vehicle speed, and time.
3. The method according to claim 1, characterized in that, The preprocessing in step S2 includes: handling missing and outlier values, smoothing noise; and data integration, merging data from multiple data sources.
4. The method according to claim 1, characterized in that, In step S3, the time series data is divided into states based on vehicle speed and battery current: if the vehicle speed is greater than 0 or the current is greater than 0, it is determined to be in driving state; if the vehicle speed is less than or equal to 0 and the current is stable and negative, it is determined to be in charging state; the time steps of consecutive states are merged into "segments", and finally driving segments and charging segments are obtained. The thousands of driving segments and charging segments of each vehicle constitute driving dataset and charging dataset respectively.
5. The method according to claim 1, characterized in that, The extraction of multiple statistical features in step S4 includes: constructing a corresponding voltage matrix for each driving segment or charging segment, including time step and number of batteries; and calculating the following basic statistical features based on the voltage matrix: maximum value, minimum value, mean, variance, median, absolute energy, kurtosis, mean absolute change, sample entropy, and skewness. Calculate the statistical features of each of the basic statistical features, including the range and variance, to obtain a total of twenty statistical features.
6. The method according to claim 5, characterized in that, Step S4 specifically involves: For each driving segment or charging segment, a voltage matrix R is constructed based on the time step and the number of batteries. k×n : In the formula, P t,j Let be the voltage of the j-th battery cell at time step t; n is the number of battery cells in the battery pack; k is the total time step for a certain driving or charging segment of the vehicle; Based on the voltage matrix R above k×n Calculate ten basic statistical characteristics, including maximum, minimum, mean, variance, median, absolute energy, kurtosis, mean absolute change, sample entropy, and skewness. The formulas for calculating each basic statistical characteristic are as follows: (1) Maximum value: (2) Minimum value: (3) Mean: (4) Variance: (5) Median: (6) Absolute energy: (7) Kurtosis: (8) Average absolute change: (9) Sample entropy: In the formula, B a (r) is (P) 1, j , ..., P t, j , ..., P k, j The probability that two sequences match at a points in (A). a (r) is (P) 1, j , ..., P t, j , ..., P k, j The probability that two sequences match at a+1 points; r is the tolerance window, described as: (10) Skewness: Furthermore, the range and variance of each feature were calculated, resulting in twenty statistical features. All features were calculated separately for driving and charging states, using the following formulas: In the formula, F represents the 10 basic statistical characteristics in (1) to (10) above; s F,r The range of characteristic F; s F,v It is the variance of characteristic F.
7. The method according to claim 1, characterized in that, The three dimensions in step S5 include: Dimension 1: For each statistical feature, construct the actual distribution using a multi-iteration Gaussian mixture model (MIGMM) and perform a Kolmogorov-Smirnov test against the theoretical Gaussian distribution; Dimension 2: Perform the KS test on the feature distribution of normal and unsafe batteries to determine whether the features can significantly distinguish between normal and unsafe batteries; Dimension 3: Based on the screening results under both driving and charging conditions, select features that are effective in both conditions.
8. The method according to claim 7, characterized in that, The three dimensions in step S5 are specifically: Dimension 1: For each statistical feature, a multi-iteration Gaussian mixture model (MIGMM) is used to construct the actual distribution, and a Kolmogorov-Smirnov test is performed against the theoretical Gaussian distribution. Features with a p-value > 0.05 are selected. The theoretical distribution is a Gaussian distribution, and its probability density function can be expressed as: In the formula, s represents the twenty constructed statistical features. This represents the average value of s for different driving or charging segments. The standard deviation of different driving / charging segments s is calculated using the following formula: In the formula s (i) This represents the characteristics of the i-th driving segment or charging segment. m represents the total number of driving segments or charging segments for a normal electric vehicle. The actual distribution is constructed using a multi-iteration Gaussian mixture model (MIGMM), whose probability density function is described as: In the formula, W is the number of Gaussian models, and α w It represents the probability that a feature belongs to the w-th sub-model. ; The Gaussian distribution density function of the w-th sub-model is calculated by the following formula: In the formula, These are the parameters of the w-th sub-model; It is the mean of the w-th sub-model; It is the variance of the w-th sub-model; Maximum likelihood function estimation and : Based on the above probability density function and Gaussian distribution density function, Kolmogorov-Smirnov analysis is generally performed by establishing the null hypothesis that the empirical distribution conforms to the theoretical distribution, and the alternative hypothesis that it does not: In the formula, H0 is the null hypothesis and H1 is the alternative hypothesis. Its KS test statistic for: The p-value is an indicator for determining whether to accept or reject the null hypothesis in a KS test. Its calculation formula is as follows: In the formula, Let α be the p-value. If the p-value is greater than the significance level α, according to hypothesis testing theory, there is insufficient reason to reject the null hypothesis, i.e., the null hypothesis is accepted, and the actual characteristic distribution of a normal battery follows a Gaussian distribution, thus condition 1 holds. Statistically, the significance level is usually set to 0.
05. Dimension Two: Perform a KS test on the characteristic distributions of normal and unsafe batteries, selecting features with a p-value < 0.05; generally, the null hypothesis is that the two empirical distributions are not significantly different, and the alternative hypothesis is that they are significantly different. In the formula, H0 is the null hypothesis, H1 is the alternative hypothesis, and P... UNS (s) is the cumulative probability density function of the actual distribution of unsafe batteries; The MIGMM method is also used to construct the actual distributions of normal and unsafe batteries, and its probability density function is described as follows: In the formula, The probability density of a normal battery; The probability density of unsafe batteries; Based on the actual distribution constructed above, perform the KS test: In the formula, Here is the KS test statistic for dimension two; m and Total number of driving / charging segments for normal electric vehicles and electric vehicles with unsafe batteries; The formula for calculating the P-value is: In the formula, The p-value for dimension two; If the p-value is less than the significance level α, according to hypothesis testing theory, there is sufficient reason to reject the null hypothesis, that is, the feature can significantly distinguish between normal batteries and unsafe batteries; Dimension 3: Based on the screening results under both driving and charging conditions, select features that are effective in both conditions; Ultimately, the mean range performed best across all three dimensions and was selected as the battery safety characterization feature.
9. The method according to claim 1, characterized in that, In step S6, data from other normal vehicles and vehicles that have experienced thermal runaway are used for verification. Based on the selected mean range, feature verification and multi-dimensional statistical analysis are performed to establish a statistical dimensional analysis model. The multiple dimensions include: first, the seasonal dimension, which analyzes the distinguishing ability of features under different temperature conditions; second, the mileage dimension, which analyzes the impact of battery health status on features; third, the SOC dimension, which analyzes the performance of features under different states of charge; and fourth, the vehicle state dimension, which analyzes the applicability of features under driving and charging states. Through statistical distribution comparison and hypothesis testing, the robustness of the selected features under multi-dimensional conditions is verified, and common patterns are summarized to support the formulation of battery safety management strategies.
10. The method according to claim 1, characterized in that, Step S6 specifically involves: The system was validated using data from other normal vehicles and vehicles that experienced thermal runaway. To verify the robustness of the mean range under different operating conditions, the circuit equations of the constructed battery Rint equivalent circuit model are as follows: In the formula, U j R 0，j I and I represent the battery's open-circuit voltage, terminal voltage, internal resistance in ohms, and current, respectively. The open-circuit voltage is mainly affected by the battery's state of charge (SOC). The ohmic internal resistance is mainly affected by ambient temperature, state of charge (SOC), and state of ohm (SOH). In the formula, T is the ambient temperature; H is the battery SOH; Therefore, the battery's terminal voltage can be expressed as: Safety characteristics mean range It is a mathematical transformation of the terminal voltage. It can be described as: Due to the influence of ambient temperature, State of Health (SOH), State of Charge (SOC), and total current, a statistical analysis of the mean range was conducted from four dimensions: Ambient temperature affects battery performance, so season was selected as one of the dimensions for analysis, with analysis and verification conducted according to spring, summer, autumn, and winter; State of Health (SOH) affects the selected characteristics, which is expressed as mileage in actual electric vehicle operation, so mileage was selected as one of the dimensions for analysis, divided into six intervals for analysis and verification; State of Charge (SOC) affects the selected characteristics, so SOC was selected as one of the dimensions for analysis, divided into nine intervals for analysis and verification; Vehicle status also affects the selected characteristics, which can be divided into driving, fast charging, and slow charging for analysis and verification.