Battery thermal runaway cause investigation method combining data mining with expert knowledge

Abstract

A battery thermal runaway cause investigation method combining data mining with expert knowledge fully integrates the advantages of machine learning algorithms and expert experience, first uses machine learning algorithms to quickly mine and extract data features, and then uses expert experience to provide prior knowledge guidance for the extracted features for cause reasoning, which can obtain preliminary inference results of thermal runaway causes in the remote data analysis stage, providing an important reference direction for the in-depth investigation of thermal runaway causes, not only effectively improving the accuracy of thermal runaway cause analysis, but also significantly reducing the scope of subsequent investigation, saving the time cost of on-site investigation, and improving the efficiency of new energy vehicle thermal runaway investigation. The original data and features extracted in the method can be directly measured and obtained through the sensors of the new energy vehicle battery system, without the need for additional sensors or high-frequency sampling data, with relatively low equipment and operation costs.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis and analysis technology for power batteries used in new energy vehicles, and specifically relates to a method for investigating the causes of thermal runaway in new energy vehicle batteries by combining remote data analysis with expert knowledge. Background Technology

[0002] Currently, the failures that may occur in the power batteries used in new energy vehicles are quite diverse. Some failures, such as thermal runaway, can easily cause serious safety risks, affecting not only the safety and reliability of the power battery system and the entire vehicle, but also inevitably hindering the promotion and application of electric vehicles. Therefore, it is necessary to conduct objective and accurate investigations and analyses of the causes of some important battery failures and accidents. Existing technologies for inferring the causes of thermal runaway accidents can be mainly divided into two categories: data-driven anomaly feature extraction and analysis methods, and investigation and analysis methods based on expert experience. However, both of these existing technologies still have some drawbacks. For example, the technical solutions using the former approach, such as Chinese patent applications CN114240260A and CN115575830A, are both advantageous and disadvantageous due to the complexity and diversity of thermal runaway scenarios and causes of thermal runaway in new energy vehicle lithium-ion batteries, as well as the large amount of data. The safety features extracted from a large amount of real vehicle data are multi-sourced, heterogeneous, and partially redundant, and cannot effectively locate the accident cell or characterize the thermal runaway risk. Therefore, it is still difficult to establish a training set with a clear mapping relationship between thermal runaway features and thermal runaway cause labels when training machine learning models, and the actual performance of the trained model is also limited by the computing power of the platform. As for the technical solutions using the latter approach, such as Chinese patent application CN115790899A, although they can provide prior experience with a clear mapping relationship between thermal runaway features and thermal runaway causes of power batteries, some on-site investigations are often carried out without any knowledge of the accident situation, which consumes huge time and manpower costs and may result in too much useless work in the early stages. Summary of the Invention

[0003] In view of this, and to address the technical problems existing in this field, the present invention provides a method for investigating the causes of battery thermal runaway by combining data mining and expert knowledge, specifically including the following steps:

[0004] Step 1: The cloud-based new energy vehicle operation monitoring platform extracts raw operation monitoring data for each new energy vehicle that has experienced battery thermal runaway, including: vehicle speed, voltage, current, SOC, insulation resistance, battery probe temperature, date-time, vehicle operating status, mileage, etc. After parsing, the data is converted into a time series format of numerical type and the corresponding data items are named accordingly.

[0005] Step 2: Perform data cleaning processes on the parsed data in sequence, including timestamp sorting, missing value imputation, outlier removal, and duplicate value removal.

[0006] Step 3: After data cleaning, perform vehicle operation status segmentation, date-time segmentation, and data item segmentation on the corresponding data items;

[0007] Step 4: Based on the current characteristics, vehicle speed characteristics, vehicle state coding characteristics, and on-site images / videos (if available) at the moment of thermal runaway, determine the specific thermal runaway scenario.

[0008] Step 5: Use the data obtained in Steps 1 to 3 to establish databases corresponding to different thermal runaway scenarios;

[0009] Step Six: Using the database established in Step Five, perform thermal runaway risk feature extraction. Extract the individual cell voltage correlation rate, probe temperature correlation rate, and insulation resistance features for specific short-term dimensions before thermal runaway occurs. Also, automatically filter individual cell voltage correlation features that are strongly correlated with thermal runaway risk for the entire life cycle data. Use machine learning to select the corresponding thermal runaway risk evolution patterns from the selected individual cell voltage correlation features throughout the entire life cycle. Finally, use SOC, battery probe temperature, and total current data clustering to extract abnormal charging behavior features throughout the entire life cycle.

[0010] Step 7: Using the feature data obtained in Step 6, and integrating expert experience, perform inferences on the specific causes of thermal runaway.

[0011] Furthermore, the raw data extracted in Step 1 is specifically decoded according to the provisions of GB / T 32960-2016; the parsing of the raw data specifically includes: converting the vehicle speed, voltage, current, and probe temperature data items into values ​​in the International System of Units (SI); converting the date-time data into the standard "year-month-day hour-minute-second" format; and the standardized naming of each data item includes: "Time", "Vehicle Status", "Mileage (km)", "Total Voltage (V)", "Total Current (A)", "Probe Temperature List (°C)", "Individual Unit Voltage List (V)", "SOC (%)", and "Insulation Resistance (kΩ)".

[0012] Furthermore, the data cleaning process in step two specifically includes: First, for the original data extracted by the platform that still has disordered time order after parsing, sorting it in ascending order of timestamps according to the "Time" column; the next data filling involves detecting the existence of "null" time frames and filling the element at that position with the element in the previous row of the same column; the data outlier removal specifically targets the "Individual Unit Voltage List (V)" data item, checking whether the individual unit voltage value exceeds 5V, and replacing the outlier value with the median voltage of the other units at that moment; finally, the data duplication value is checked to see if all data items in multiple consecutive rows are completely identical. If they are identical, only the first row of data is retained, and all other duplicate items are directly deleted.

[0013] Furthermore, the vehicle operating status segmentation in step three specifically involves segmenting the entire lifecycle data into parking and charging data, driving status data, and stationary parking data; the date-time segmentation extracts data from the thermal runaway occurrence and the 7 days prior to the thermal runaway occurrence in reverse order of the "time" data item; and the data item segmentation extracts the data item list corresponding to each vehicle operating status and date-time from "total voltage (V)," "individual cell voltage list (V)," "probe temperature list (°C)," "SOC (%)," "total current (A)," and "insulation resistance (kΩ)."

[0014] Furthermore, the specific thermal runaway scenarios identified in step four include parking and charging scenarios, driving scenarios, stationary parking scenarios, collision scenarios, water wading scenarios, and external fire source scenarios.

[0015] Furthermore, step six, which involves feature extraction for specific short-term dimensions prior to thermal runaway, includes the following steps:

[0016] ① The battery system operation data for the day the battery thermal runaway occurred and the previous 7 days were segmented using the "Time" data item;

[0017] ② Obtain the median of all individual cell voltages at each time point through the “Individual Cell Voltage List (V)” data item and generate a reference individual cell voltage curve; use the reference individual cell voltage curve to calculate: the voltage difference between each individual cell voltage and the reference voltage, the rate of change of voltage over time, and the time when the voltage drop rate exceeds the threshold, and store them as “Individual Cell Voltage Correlation Rate Characteristics Before Thermal Runaway”.

[0018] ③ Obtain the highest temperature and temperature rise rate of all temperature probes at each time point through the "Probe Temperature List (°C)" data item, and record the time when the temperature rise rate exceeds the threshold, and store it as "Probe Temperature Correlation Rate Feature before Thermal Runaway";

[0019] ④ Extract the minimum insulation resistance value and the rate of decrease of insulation resistance value from the "Insulation Resistance (kΩ)" data item and store them as "Insulation Resistance Correlation Rate Characteristics Before Thermal Runaway".

[0020] Furthermore, step six, the feature extraction process for the full lifecycle dimension data, includes the following steps:

[0021] ① All data on the parking and charging status throughout the entire life cycle are divided into charging cycles according to the "time" data item;

[0022] ② The “Single Unit Voltage List (V)” data item is segmented, and the relevant statistical features of each single unit voltage time series are extracted through automated feature engineering. Then, combined with the safety / risk category label of the single unit, feature importance analysis is performed, and statistical features with positive feature importance scores are automatically selected. The single unit feature matrix of each single unit is obtained as a strong correlation feature of thermal runaway risk.

[0023] ③ Input the individual feature matrices obtained in step ② into the machine learning algorithm of Gaussian mixture model, fit the maximum likelihood value of each individual, take the individual with the highest maximum likelihood as the reference individual, calculate the Euclidean distance between the feature coordinates of each individual and the feature coordinates of the reference individual in a single charging cycle, and obtain the thermal runaway risk feature distance (RFD); accumulate the RFD of each charging cycle, and define the result as the cumulative thermal runaway risk feature distance (ARFD) of each individual after the end of each charging cycle. Then, perform maximum-minimum normalization on the ARFD in the interval of 0 to 1 to obtain the normalized cumulative thermal runaway risk feature distance (NARFD); store the above calculation results as "full life cycle individual voltage risk features".

[0024] ④ Based on the data items “SOC (%)”, “Probe Temperature List (°C)”, and “Total Current (A)”, each charging cycle is taken as a sample. First, the “SOC (%)” and “Total Current (A)” features of each charging cycle sample are clustered using a clustering algorithm to mark whether the charging cycle has “overcharging” abnormal features. Then, the “Probe Temperature List (°C)” and “Total Current (A)” data items of each charging cycle sample are clustered using a clustering algorithm to mark whether the charging cycle has “low temperature charging” abnormal features. The above two marks are stored as “full life cycle abnormal charging behavior features”.

[0025] Furthermore, step seven, which involves sequentially inferring the specific causes of thermal runaway by integrating expert experience with various feature data, includes the following steps:

[0026] ① By extracting the "individual voltage correlation rate characteristics before thermal runaway", "probe temperature correlation rate characteristics before thermal runaway", and "insulation resistance correlation rate characteristics before thermal runaway", we first determine whether the voltage change rate and the highest probe temperature exceed the threshold. Then, we determine the order of the moment when the individual voltage change rate exceeds the threshold and the moment when the probe temperature rise rate exceeds the threshold. If the individual voltage drop occurs after the temperature rise, the individual is considered to have been ignited. Otherwise, if the individual voltage drop occurs before the temperature rise, the individual is considered to have spontaneously combusted.

[0027] ② If a single cell is found to have been ignited, determine the cause of the ignition: First, calculate the number of probes in the battery pack whose temperature exceeds the maximum temperature threshold. If the number does not exceed the threshold, it indicates that the cell was ignited by a spontaneously combusting cell in the same module or by thermal runaway propagation after spontaneous combustion of a cell in another adjacent module. Conversely, if the number of probes in the battery pack that exceed the maximum temperature threshold is higher than the threshold, it indicates that the cell was ignited by an external fire source. Possible sources of this fire source include other electrical system components inside the vehicle besides the battery pack or fire sources outside the vehicle.

[0028] ③ If a single cell is spontaneously combusted, then based on the evolution of the NARFD value in the "Life Cycle Cell Voltage Risk Characteristics", determine whether it belongs to "latent thermal runaway" or "sudden thermal runaway".

[0029] ④ If the cell is determined to be a "latent thermal runaway", then based on the "abnormal charging behavior characteristics throughout the entire life cycle", determine whether there are abnormal charging behavior tags such as "overcharging" or "low temperature charging". If so, it is inferred that the spontaneous combustion is caused by the accumulation of thermal runaway risk due to electricity abuse.

[0030] ⑤ If it is determined that the single cell belongs to "sudden thermal runaway", it is necessary to use the on-site battery system image information to detect whether there are traces of mechanical force squeezing, bumping, foreign object puncture, and whether the location of the traces corresponds to the location of the burning single cell; if the judgment result is "yes", it is inferred that the thermal runaway was caused by mechanical abuse.

[0031] The battery thermal runaway cause investigation method combining data mining and expert knowledge provided by this invention fully integrates the advantages of machine learning algorithms and expert experience. It first utilizes machine learning algorithms to quickly extract data features, and then leverages expert experience to provide prior knowledge guidance for causal reasoning based on the extracted features. This allows for preliminary inferences about the cause of thermal runaway during the remote data analysis stage, providing important reference directions for in-depth investigations. This not only effectively improves the accuracy of thermal runaway cause analysis but also significantly narrows the scope of subsequent investigations, saves time costs associated with on-site investigations, and improves the efficiency of thermal runaway investigations in new energy vehicles. All the raw data and features extracted in this method can be directly measured and obtained through sensors in the new energy vehicle's onboard battery system, eliminating the need for additional sensors or high-frequency sampling, resulting in lower equipment and computational costs. After data feature mining, the vehicle is added to the thermal runaway scenario database as a sample. This invention can also continuously update and add database samples, thereby continuously improving the mapping relationship between thermal runaway cause labels and thermal runaway data features. Attached Figure Description

[0032] Figure 1 This provides the overall framework for the thermal runaway cause investigation method provided by the present invention;

[0033] Figure 2 This refers to the data processing and judgment process for thermal runaway scenarios in the method of this invention;

[0034] Figure 3 This is the process for extracting risk features using remote real-vehicle data in the method of the present invention;

[0035] Figure 4 This is the process of inferring the cause of thermal runaway by incorporating expert experience in the method of this invention. Detailed Implementation

[0036] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0037] The battery thermal runaway cause investigation method combining data mining and expert knowledge provided by this invention, such as... Figure 1 As shown, the specific steps include:

[0038] Step 1: The cloud-based new energy vehicle operation monitoring platform extracts raw operation monitoring data for each new energy vehicle that has experienced battery thermal runaway, including: vehicle speed, voltage, current, SOC, insulation resistance, battery probe temperature, date-time, vehicle operating status, mileage, etc. After parsing, the data is converted into a time series format of numerical type and the corresponding data items are named accordingly.

[0039] Step 2: Perform data cleaning processes on the parsed data in sequence, including timestamp sorting, missing value imputation, outlier removal, and duplicate value removal.

[0040] Step 3: After data cleaning, perform vehicle operation status segmentation, date-time segmentation, and data item segmentation on the corresponding data items;

[0041] Step 4: Based on the current characteristics, vehicle speed characteristics, vehicle state coding characteristics, and on-site images / videos (if available) at the moment of thermal runaway, determine the specific thermal runaway scenario.

[0042] Step 5: Use the data obtained in Steps 1 to 3 to establish databases corresponding to different thermal runaway scenarios;

[0043] Step Six: Using the database established in Step Five, perform thermal runaway risk feature extraction. Extract the individual cell voltage correlation rate, probe temperature correlation rate, and insulation resistance features for specific short-term dimensions before thermal runaway occurs. Also, automatically filter individual cell voltage correlation features that are strongly correlated with thermal runaway risk for the entire life cycle data. Use machine learning to select the corresponding thermal runaway risk evolution patterns from the selected individual cell voltage correlation features throughout the entire life cycle. Finally, use SOC, battery probe temperature, and total current data clustering to extract abnormal charging behavior features throughout the entire life cycle.

[0044] Step 7: Using the feature data obtained in Step 6, and integrating expert experience, perform inferences on the specific causes of thermal runaway.

[0045] In a preferred embodiment of the present invention, the raw data extracted in step one is specifically decoded according to the provisions of GB / T32960-2016; the parsing of the raw data specifically includes: converting the vehicle speed, voltage, current, and probe temperature data items into values ​​in the International System of Units (SI); converting the date-time data into the standard "year-month-day hour-minute-second" format; the standardized naming of each data item includes: "time", "vehicle status", "mileage (km)", "total voltage (V)", "total current (A)", "probe temperature list (°C)", "individual unit voltage list (V)", "SOC (%)", and "insulation resistance (kΩ)".

[0046] The data cleaning process in step two specifically includes: First, for the raw data extracted by the platform that still has disordered time sequence after parsing, sorting it in ascending order of timestamps according to the "Time" column; the next data filling involves detecting the existence of "null" time frames and filling the element at that position with the element in the previous row of the same column; the data outlier removal specifically targets the "Individual Unit Voltage List (V)" data item, checking whether the individual unit voltage value exceeds 5V, and replacing the outlier value with the median voltage of the other units at that moment; finally, the data duplicate removal involves checking whether all data items in multiple consecutive rows are completely identical, if they are identical, only the first row of data is retained, and all other duplicate items are directly deleted.

[0047] Step 3, vehicle operation status segmentation, specifically involves segmenting the entire lifecycle data into parking and charging data, driving status data, and stationary parking data; date-time segmentation involves extracting data from the thermal runaway occurrence and the 7 days prior to the thermal runaway occurrence in reverse order of the "time" data item; data item segmentation involves extracting data item lists corresponding to each vehicle operation status and date-time from "total voltage (V)," "individual cell voltage list (V)," "probe temperature list (°C)," "SOC (%)," "total current (A)," and "insulation resistance (kΩ)."

[0048] The specific thermal runaway scenarios identified in step four include parking and charging scenarios, driving scenarios, stationary parking scenarios, collision scenarios, water wading scenarios, and external fire source scenarios. Figure 2 The data processing and judgment process related to thermal runaway scenarios is shown.

[0049] like Figure 3 As shown, step six, the feature extraction process for a specific short period before thermal runaway, includes the following steps:

[0050] ① The battery system operation data for the day the battery thermal runaway occurred and the previous 7 days were segmented using the "Time" data item;

[0051] ② Obtain the median of all individual cell voltages at each time point through the “Individual Cell Voltage List (V)” data item and generate a reference individual cell voltage curve; use the reference individual cell voltage curve to calculate: the voltage difference between each individual cell voltage and the reference voltage, the rate of change of voltage over time, and the time when the voltage drop rate exceeds the threshold, and store them as “Individual Cell Voltage Correlation Rate Characteristics Before Thermal Runaway”.

[0052] ③ Obtain the highest temperature and temperature rise rate of all temperature probes at each time point through the "Probe Temperature List (°C)" data item, and record the time when the temperature rise rate exceeds the threshold, and store it as "Probe Temperature Correlation Rate Feature before Thermal Runaway";

[0053] ④ Extract the minimum insulation resistance value and the rate of decrease of insulation resistance value from the "Insulation Resistance (kΩ)" data item and store them as "Insulation Resistance Correlation Rate Characteristics Before Thermal Runaway".

[0054] The feature extraction process for full lifecycle data includes the following steps:

[0055] ① All data on the parking and charging status throughout the entire life cycle are divided into charging cycles according to the "time" data item;

[0056] ② The “Single Cell Voltage List (V)” data item is segmented, and the relevant statistical features of each single cell voltage time series are extracted through automated feature engineering. Then, combined with the safety / risk category label of the single cell, feature importance analysis is performed, and statistical features with positive feature importance scores are automatically selected as strong correlation features of thermal runaway risk. A single cell feature matrix is ​​built for each single cell using the strong correlation features of thermal runaway risk.

[0057] ③ Input the individual feature matrices obtained in step ② into the machine learning algorithm of Gaussian mixture model, fit the maximum likelihood value of each individual, take the individual with the highest maximum likelihood as the reference individual, calculate the Euclidean distance between the feature coordinates of each individual and the feature coordinates of the reference individual in a single charging cycle, and obtain the thermal runaway risk feature distance (RFD); accumulate the RFD of each charging cycle, and define the result as the cumulative thermal runaway risk feature distance (ARFD) of each individual after the end of each charging cycle. Then, perform maximum-minimum normalization on the ARFD in the interval of 0 to 1 to obtain the normalized cumulative thermal runaway risk feature distance (NARFD); store the above calculation results as "full life cycle individual voltage risk features".

[0058] ④ Based on the data items “SOC (%)”, “Probe Temperature List (°C)”, and “Total Current (A)”, each charging cycle is taken as a sample. First, a clustering algorithm is used to mark whether the “overcharge” abnormal feature exists in the “SOC (%)” and “Total Current (A)” features of each charging cycle sample. Then, a clustering algorithm is used to cluster the “Probe Temperature List (°C)” and “Total Current (A)” data items of each charging cycle sample to mark whether the “low temperature charging” abnormal feature exists in the charging cycle. The above two marks are stored as “Full Life Cycle Abnormal Charging Behavior Features”.

[0059] like Figure 4As shown, step seven, which involves sequentially inferring the specific causes of thermal runaway by integrating various feature data with expert experience, includes the following steps:

[0060] ① By extracting the "individual voltage correlation rate characteristics before thermal runaway", "probe temperature correlation rate characteristics before thermal runaway", and "insulation resistance correlation rate characteristics before thermal runaway", we first determine whether the voltage change rate and the highest probe temperature exceed the threshold. Then, we determine the order of the moment when the individual voltage change rate exceeds the threshold and the moment when the probe temperature rise rate exceeds the threshold. If the individual voltage drop occurs after the temperature rise, the individual is considered to have been ignited. Otherwise, if the individual voltage drop occurs before the temperature rise, the individual is considered to have spontaneously combusted.

[0061] ② If a single cell is found to have been ignited, determine the cause of the ignition: First, calculate the number of probes in the battery pack whose temperature exceeds the maximum temperature threshold. If the number does not exceed the threshold, it indicates that the cell was ignited by a spontaneously combusting cell in the same module or by thermal runaway propagation after spontaneous combustion of a cell in another adjacent module. Conversely, if the number of probes in the battery pack that exceed the maximum temperature threshold is higher than the threshold, it indicates that the cell was ignited by an external fire source. Possible sources of this fire source include other electrical system components inside the vehicle besides the battery pack or fire sources outside the vehicle.

[0062] ③ If a single cell is spontaneously combusted, then based on the evolution of the NARFD value in the "Life Cycle Cell Voltage Risk Characteristics", determine whether it belongs to "latent thermal runaway" or "sudden thermal runaway".

[0063] ④ If the cell is determined to be a "latent thermal runaway", then based on the "abnormal charging behavior characteristics throughout the entire life cycle", determine whether there are abnormal charging behavior tags such as "overcharging" or "low temperature charging". If so, it is inferred that the spontaneous combustion is caused by the accumulation of thermal runaway risk due to electricity abuse.

[0064] ⑤ If it is determined that the single cell belongs to "sudden thermal runaway", it is necessary to use the on-site battery system image information to detect whether there are traces of mechanical force squeezing, bumping, foreign object puncture, and whether the location of the traces corresponds to the location of the burning single cell; if the judgment result is "yes", it is inferred that the thermal runaway was caused by mechanical abuse.

[0065] It should be understood that the sequence number of each step in the embodiments of the present invention does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for investigating the causes of battery thermal runaway by combining data mining and expert knowledge, characterized by: Specifically, the following steps are included: Step 1: The cloud-based new energy vehicle operation monitoring platform extracts raw operation monitoring data for each new energy vehicle that has experienced battery thermal runaway, including: vehicle speed, voltage, current, SOC, insulation resistance, battery probe temperature, date-time, vehicle operating status, and mileage. After parsing, the data is converted into a time series format of numerical type, and the corresponding data items are named accordingly. Step 2: Perform data cleaning processes on the parsed data in sequence, including timestamp sorting, missing value imputation, outlier removal, and duplicate value removal. Step 3: After data cleaning, perform vehicle operation status segmentation, date-time segmentation, and data item segmentation on the corresponding data items; Step 4: Based on the current characteristics, vehicle speed characteristics, and vehicle status coding characteristics at the moment of thermal runaway, determine the specific thermal runaway scenario; if there are pictures / videos of the scene at the time of thermal runaway, they should also be used as factors in determining the thermal runaway scenario. Step 5: Use the data obtained in Steps 1 to 3 to establish databases corresponding to different thermal runaway scenarios; Step Six: Using the database established in Step Five, perform thermal runaway risk feature extraction. Extract the individual cell voltage correlation rate, probe temperature correlation rate, and insulation resistance features for specific short-term dimensions before thermal runaway occurs. Also, automatically filter individual cell voltage correlation features that are strongly correlated with thermal runaway risk for the entire life cycle data. Use machine learning to select the corresponding thermal runaway risk evolution patterns from the selected individual cell voltage correlation features throughout the entire life cycle. Finally, use SOC, battery probe temperature, and total current data clustering to extract abnormal charging behavior features throughout the entire life cycle. Step 7: Using the feature data obtained in Step 6, and integrating expert experience, infer the specific cause of thermal runaway; the inference process includes the following steps: ① By extracting the "individual voltage correlation rate characteristics before thermal runaway", "probe temperature correlation rate characteristics before thermal runaway", and "insulation resistance correlation rate characteristics before thermal runaway", we first determine whether the voltage change rate and the highest probe temperature exceed the threshold. Then, we determine the order of the moment when the individual voltage change rate exceeds the threshold and the moment when the probe temperature rise rate exceeds the threshold. If the individual voltage drop occurs after the temperature rise, the individual is considered to have been ignited. Otherwise, if the individual voltage drop occurs before the temperature rise, the individual is considered to have spontaneously combusted. ② If a single cell is found to have been ignited, determine the cause of the ignition: First, calculate the number of probes in the battery pack whose temperature exceeds the maximum temperature threshold. If the number does not exceed the threshold, it indicates that the cell was ignited by a spontaneously combusting cell in the same module or by thermal runaway propagation after spontaneous combustion of a cell in another adjacent module. Conversely, if the number of probes in the battery pack that exceed the maximum temperature threshold is higher than the threshold, it indicates that the cell was ignited by an external fire source. Possible sources of this fire source include other electrical system components inside the vehicle besides the battery pack or fire sources outside the vehicle. ③ If a single cell is spontaneously combusted, then based on the evolution of the NARFD value in the "Life Cycle Cell Voltage Risk Characteristics", determine whether it belongs to "latent thermal runaway" or "sudden thermal runaway". ④ If the cell is determined to be a "latent thermal runaway", then based on the "abnormal charging behavior characteristics throughout the entire life cycle", determine whether there are "overcharging" or "low temperature charging" abnormal charging behavior tags. If so, it is inferred that the spontaneous combustion is caused by the accumulation of thermal runaway risk due to battery abuse. ⑤ If it is determined that the single cell belongs to "sudden thermal runaway", then it is necessary to use the on-site battery system image information to detect whether there are traces of mechanical force squeezing, bumping, foreign object puncture, and whether the location of the traces corresponds to the location of the burning single cell; if the judgment result is "yes", then it is inferred that the thermal runaway was caused by mechanical abuse.

2. The method as described in claim 1, characterized in that: The raw data extracted in Step 1 is decoded according to the provisions of GB / T32960-2016. The analysis of the raw data includes converting the vehicle speed, voltage, current, and probe temperature data items into values ​​in the International System of Units (SI), and converting the date-time data into the standard "year-month-day hour-minute-second" format. The standardized naming of each data item includes: "Time", "Vehicle Status", "Mileage (km)", "Total Voltage (V)", "Total Current (A)", "Probe Temperature List (°C)", "Individual Unit Voltage List (V)", "SOC (%)", and "Insulation Resistance (kΩ)".

3. The method as described in claim 2, characterized in that: The data cleaning process in step two specifically includes: First, for the raw data extracted by the platform that still has disordered time order after parsing, sorting it in ascending order of timestamps according to the "Time" column; the next data filling involves detecting the existence of "null" time frames and filling the element at that position with the element in the previous row of the same column; the data outlier removal specifically targets the "Individual Unit Voltage List (V)" data item, checking whether the individual unit voltage value exceeds 5V, and replacing the outlier value with the median voltage of the other units at that moment; finally, the data duplication check checks whether all data items in multiple consecutive rows are completely identical. If they are identical, only the first row of data is retained, and all other duplicate items are directly deleted.

4. The method as described in claim 3, characterized in that: Step 3, vehicle operation status segmentation, specifically involves segmenting the entire lifecycle data into parking and charging data, driving status data, and stationary parking data; date-time segmentation involves extracting data from the thermal runaway occurrence and the 7 days prior to the thermal runaway occurrence in reverse order of the "time" data item; data item segmentation involves extracting data item lists corresponding to each vehicle operation status and date-time from "total voltage (V)," "individual cell voltage list (V)," "probe temperature list (°C)," "SOC (%)," "total current (A)," and "insulation resistance (kΩ)." 5. The method as described in claim 4, characterized in that: The specific thermal runaway scenarios identified in step four include parking and charging scenarios, driving scenarios, stationary parking scenarios, collision scenarios, water wading scenarios, and external fire source scenarios.

6. The method as described in claim 5, characterized in that: Step six, the feature extraction process for specific short-term dimensions before thermal runaway, includes the following steps: ① The battery system operation data for the day the battery thermal runaway occurred and the previous 7 days were segmented using the "Time" data item; ② Obtain the median of all individual cell voltages at each time point through the "Individual Cell Voltage List (V)" data item and generate a reference individual cell voltage curve; use the reference individual cell voltage curve to calculate: the voltage difference between each individual cell voltage and the reference voltage, the rate of change of voltage over time, and the time when the voltage drop rate exceeds the threshold, and store them as "Individual Cell Voltage Correlation Rate Characteristics Before Thermal Runaway"; ③ Obtain the highest temperature and temperature rise rate of all temperature probes at each time point through the "Probe Temperature List (°C)" data item, and record the time when the temperature rise rate exceeds the threshold, and store it as "Probe Temperature Correlation Rate Feature before Thermal Runaway"; ④ Extract the minimum insulation resistance value and the rate of decrease of insulation resistance value from the "insulation resistance value (kΩ)" data item and store them as "insulation resistance value correlation rate characteristics before thermal runaway".

7. The method as described in claim 6, characterized in that: Step six, the feature extraction process for the full lifecycle dimension data, includes the following steps: ① All data on the parking and charging status throughout the entire life cycle are divided into charging cycles according to the "time" data item; ② The data item "Single cell voltage list (V)" is segmented, and the relevant statistical features of each single cell voltage time series are extracted through automated feature engineering. Then, combined with the safety / risk category label of the single cell, feature importance analysis is performed, and statistical features with positive feature importance scores are automatically selected. The single cell feature matrix of each single cell is obtained as a strong correlation feature of thermal runaway risk. ③ Input the individual feature matrices obtained in step ② into the machine learning algorithm of Gaussian mixture model, fit the maximum likelihood value of each individual, take the individual with the highest maximum likelihood as the reference individual, calculate the Euclidean distance between the feature coordinates of each individual and the feature coordinates of the reference individual in a single charging cycle, and obtain the thermal runaway risk feature distance RFD; accumulate the RFD of each charging cycle, and define the result as the cumulative thermal runaway risk feature distance ARFD of each individual after the end of each charging cycle, and then perform maximum-minimum normalization on ARFD in the interval of 0 to 1 to obtain the standardized cumulative thermal runaway risk feature distance NARFD; store the above calculation results as "full life cycle individual voltage risk characteristics"; ④ Based on the data items "SOC (%)", "Probe Temperature List (°C)" and "Total Current (A)", each charging cycle is taken as a sample. First, the "SOC (%)" and "Total Current (A)" features of each charging cycle sample are clustered using a clustering algorithm to mark whether the charging cycle has "overcharging" abnormal features. Then, the "Probe Temperature List (°C)" and "Total Current (A)" data items of each charging cycle sample are clustered using a clustering algorithm to mark whether the charging cycle has "low temperature charging" abnormal features. The above two marks are stored as "Full Life Cycle Abnormal Charging Behavior Features".

Images