Battery thermal runaway prediction methods, devices, equipment, and media based on segmented characteristics

Abstract

This application belongs to the field of battery thermal runaway prediction technology, and discloses a method, apparatus, device, and medium for battery thermal runaway prediction based on segmented features. The method includes: acquiring the voltage and specified data of each cell in the battery; slicing the specified data according to a preset dimension segmentation rule to obtain slice datasets for each dimension; sorting the cell voltages of each frame of data in each slice dataset, and calculating the mode of the sorted values ​​of each cell per preset time unit under each slice dimension; extracting time-series features, including complexity features, approximate entropy features, average absolute change values, and coefficient of variation, based on the time-series data of the mode of the sorted values ​​of each cell; and predicting the probability of battery thermal runaway based on the time-series features. This application improves the accuracy of battery thermal runaway prediction through multi-dimensional slicing, mode statistics of sorted values, and extraction of time-series features, providing strong support for battery safety management.

Description

Technical Field

[0001] This application relates to the field of battery safety control technology, and in particular to a method, apparatus, device and medium for predicting battery thermal runaway based on segmented characteristics. Background Technology

[0002] Thermal runaway in power batteries refers to the phenomenon where, under abnormal conditions such as overcharging, over-discharging, short circuits, mechanical damage, or high temperatures, a chain reaction of exothermic reactions occurs inside the battery, leading to a rapid rise in temperature and potentially causing a fire or explosion. Thermal runaway is one of the most serious safety risks associated with power batteries, potentially causing personal injury, property damage, and environmental pollution. With the rapid development of new energy vehicles and energy storage systems, thermal runaway prediction has become a key focus of the industry.

[0003] Currently, the prediction of thermal runaway in power batteries mainly relies on the following methods: single feature analysis methods, such as threshold judgments based on single features like voltage, temperature, and current, which cannot comprehensively reflect the complex state of the battery; statistical models, such as statistical analysis based on historical data, which lack the ability to capture real-time dynamic changes; and machine learning methods, which, while capable of handling multi-dimensional data, typically ignore the correlation between individual battery cells and the complexity of temporal characteristics.

[0004] The aforementioned existing methods all suffer from problems such as incomplete feature extraction, neglect of individual correlations and temporal complexity, resulting in low prediction accuracy. Summary of the Invention

[0005] This application provides a method, apparatus, device, and medium for predicting battery thermal runaway based on segmented features, aiming to solve the technical problem of low accuracy in predicting battery thermal runaway in the prior art.

[0006] To achieve the aforementioned objectives, the first aspect of this application proposes a battery thermal runaway prediction method based on segmented features, the method comprising:

[0007] Obtain the voltage and specified data of each cell in the battery, wherein the specified data includes at least one or more of the following: current, temperature, and SOC (State of Charge).

[0008] The specified data is sliced ​​according to the preset segmentation rules to obtain sliced ​​datasets for each dimension;

[0009] Sort the individual voltages of each frame in each slice dataset, and count the mode of the sorted values ​​of each individual in each slice dimension per preset time unit. Individuals with the same voltage are assigned the same sorted value.

[0010] Based on the mode time series data of the sorted values ​​of each individual, extract time series features;

[0011] Based on the aforementioned timing characteristics, the probability of thermal runaway of the battery is predicted.

[0012] Furthermore, predicting the probability of thermal runaway of the battery based on the time-series characteristics includes:

[0013] The temporal features of each individual are reduced in dimensionality using a pre-defined dimensionality reduction method to obtain dimensionality-reduced features;

[0014] The reduced-dimensional features of all individual units of the same vehicle are flattened and merged to form a vehicle-level feature vector;

[0015] Based on the vehicle-level feature vector, the probability of thermal runaway of the battery is predicted using a machine learning algorithm.

[0016] Further, the step of slicing the specified data according to a preset dimension segmentation rule to obtain sliced ​​datasets for each dimension includes:

[0017] Slicing by charge / discharge state dimension: when current > 0, it is divided into charging state dataset; when current < 0, it is divided into discharging state dataset.

[0018] Slicing by SOC dimension: Divide the SOC into multiple continuous intervals from 0 to 100%, and form SOC slice datasets for each corresponding interval.

[0019] Slicing by current dimension: The current is divided into current intervals that are less than a first current threshold, current intervals that are greater than a second current threshold, and one or more current intervals between the first current threshold and the second current threshold, and current slice datasets for the corresponding intervals are formed respectively; wherein, the second current threshold is greater than the first current threshold.

[0020] Slicing by temperature dimension: The temperature is divided into temperature ranges below a first temperature threshold, temperature ranges above a second temperature threshold, and one or more temperature ranges between the first temperature threshold and the second temperature threshold, and temperature slice datasets are formed for each range; wherein, the second temperature threshold is greater than the first temperature threshold.

[0021] Furthermore, the temporal features of each individual are reduced in dimensionality using a preset dimensionality reduction method to obtain dimensionality-reduced features, including:

[0022] Principal component analysis is performed on the temporal features of each individual to reduce the feature dimension to a preset number of dimensions. During the dimensionality reduction process, principal components that can reflect the main information of the original features are retained.

[0023] Calculate the contribution rate of each principal component to ensure that the preset number of dimensions after dimensionality reduction can retain the main changing trends and differences of the temporal features of each individual.

[0024] Furthermore, the step of predicting the probability of thermal runaway of the battery based on the vehicle-level feature vector using a machine learning algorithm includes:

[0025] The vehicle-level feature vector is input into a pre-trained thermal runaway prediction model, which outputs the probability value of the battery thermal runaway. The thermal runaway prediction model is a model trained based on historical vehicle-level feature vectors, using a logistic regression model, a support vector classifier, or a lightweight gradient booster model.

[0026] When the probability value exceeds a preset threshold, a thermal runaway risk warning is triggered.

[0027] Furthermore, the time-series features include complexity features, approximate entropy features, the average value of absolute change, and the coefficient of variation; the extraction of time-series features based on the mode of the sorted values ​​of each individual unit includes:

[0028] For each individual unit in the mode of the sorted value time series under each slice dimension, its complexity feature is calculated. The complexity is quantified by the ratio of the minimum number of distinct substrings into which the sequence is decomposed to the sequence length, combined with the logarithm of the sequence length.

[0029] A dynamic time window related to sequence variance and length is introduced, and the similarity difference of sequence substrings under different embedding dimensions is quantified to obtain approximate entropy features;

[0030] The average value of the absolute change is obtained by solving the arithmetic mean of the absolute differences of the modes of adjacent sorted values ​​in the time series;

[0031] The coefficient of variation is obtained by calculating the ratio of the standard deviation to the mean of the time series.

[0032] Furthermore, the process of sorting the individual voltages of each frame in each slice dataset and calculating the mode of the sorted values ​​of each individual unit per preset time unit under each slice dimension includes:

[0033] Based on each frame of data in each slice dimension, the voltages of all cells in the battery are sorted in ascending order to generate a corresponding sort value sequence. Cells with the same voltage are assigned the same sort value. The sort value range is [1, M], where M is the total number of cells.

[0034] Using days as the preset time unit, based on the daily sort value sequence of each individual in each slice dimension, the sort value with the highest frequency is counted as the mode of the sort value of that individual on that day.

[0035] The sorting value is calculated according to the formula:

[0036] ,in, Let i be the voltage sorting value of a single cell i in slice dimension d. This is an indicator function, where K is the number of cells with equal voltage.

[0037] A second aspect of this application proposes a battery thermal runaway prediction device based on segmented characteristics, comprising:

[0038] An acquisition unit is used to acquire the voltage and specified data of each cell in the battery, wherein the specified data includes at least one or more of voltage, current, temperature and SOC data;

[0039] The segmentation unit is used to slice the specified data according to the segmentation rules of the preset dimensions to obtain the sliced ​​dataset of each dimension.

[0040] The sorting unit is used to sort the individual voltages of each frame of data in each slice dataset, and to count the mode of the sorted values ​​of each individual in each slice dimension for each preset time unit. Individuals with the same voltage are assigned the same sorted value.

[0041] The extraction unit is used to extract time series features based on the mode of the sorted values ​​of each individual data point.

[0042] The prediction unit is used to predict the probability of thermal runaway of the battery based on the time-series characteristics.

[0043] A third aspect of this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the battery thermal runaway prediction method based on segmented features as described in any of the preceding claims.

[0044] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the battery thermal runaway prediction method based on segmented features as described in any of the preceding claims.

[0045] Beneficial effects:

[0046] This application presents a battery thermal runaway prediction method, apparatus, device, and medium based on segmented features. By slicing specified data of individual battery cells according to preset dimensions such as charge / discharge state, SOC, current, and temperature, the battery operating data can be divided into different operating scenarios. Subsequent analysis can be carried out on battery characteristics under different scenarios, making the analysis more targeted and laying the foundation for accurate prediction of thermal runaway probability. The method sorts the individual cell voltages in each slice dataset and calculates the mode of the sorted values ​​for each preset time unit. Cells with the same voltage are assigned the same sorted value. The resulting mode of sorted values ​​reflects the typical sorting characteristics of individual cell voltages within specific dimensions and time units, providing stable and representative data for subsequent time-series feature extraction. Based on the time-series data of the mode of sorted values ​​for each cell, various time-series features are extracted, including complexity features, approximate entropy features, the average of absolute change values, and the coefficient of variation. These features characterize the change patterns of the time-series data of the mode of sorted values ​​from different perspectives, providing rich and effective feature inputs for thermal runaway probability prediction and improving prediction accuracy. By utilizing extracted time-series features and combining them with appropriate algorithm models, the probability of battery thermal runaway can be predicted. Compared with existing prediction methods that lack targeted segmentation and multi-feature fusion, this method can more accurately capture changes in battery state, thereby achieving accurate prediction of the probability of battery thermal runaway and providing a reliable basis for battery safety management. Attached Figure Description

[0047] Figure 1 A schematic flowchart illustrating a battery thermal runaway prediction method based on segmented features, according to an embodiment of the invention.

[0048] Figure 2 This is a schematic flowchart of a battery thermal runaway prediction device based on segmented features according to an embodiment of the invention.

[0049] Figure 3 This is a schematic block diagram of a computer device according to an embodiment of the invention.

[0050] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0053] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0054] Reference Figure 1 This application provides a battery thermal runaway prediction method based on segmented features, including the following steps:

[0055] S1: Obtain the voltage and specified data of each cell in the battery, wherein the specified data includes at least one or more of the following: current, temperature and SOC data.

[0056] The voltage mentioned above refers to the potential difference between the positive and negative electrodes of a single battery cell. The specified data refers to key parameters used for predicting battery thermal runaway, including current (the magnitude of the battery's charging and discharging current), temperature (the temperature of the battery cell), and SOC (State of Charge). Voltage, current, and temperature data for each cell are collected in real time from the electric vehicle's Battery Management System (BMS), and SOC data is calculated using the BMS's built-in algorithm. For example, taking a ternary lithium battery from a certain brand of electric vehicle as an example, the BMS collects the voltage (accuracy ±0.01V), current (accuracy ±0.1A), and temperature (accuracy ±1℃) of each cell at a frequency of 1Hz, and calculates the SOC every 10 seconds, ranging from 0-100%. Obtaining multi-dimensional battery operating data provides comprehensive and accurate basic data support for subsequent thermal runaway prediction, ensuring that subsequent analysis is based on information that truly reflects the battery's state. It should be noted that the specified data mentioned above can be one or more of current, temperature, and SOC data; in this embodiment, it includes all current, temperature, and SOC data. The single cell mentioned above can be a battery module within a battery or a single cell, depending on the actual situation. For example, in a moduleless battery, a single cell is a single battery cell. When the battery includes multiple battery modules, the single cell can be a battery module.

[0057] S2: Slice the specified data according to the preset segmentation rules to obtain sliced ​​datasets for each dimension.

[0058] The aforementioned slicing process refers to dividing the overall data into multiple subsets based on different preset dimensions (such as state of charge / discharge, SOC range, current range, and temperature range). According to the characteristics of different battery operating states, the acquired specified data is divided into slice datasets based on the dimensions of state of charge / discharge, SOC, current, and temperature, resulting in corresponding slice datasets. For example, in the state of charge / discharge dimension, when the current > 0A, it is classified as a charging state dataset; when the current < 0A, it is classified as a discharging state dataset. In the SOC dimension, SOC is divided into three ranges: 0-30%, 30-80%, and 80-100%, forming slice datasets for each range. In the current dimension, current is divided into three ranges: less than 30A, 30-100A, and greater than 100A, forming current slice datasets. In the temperature dimension, it is divided into six ranges: less than 20℃, 20-25℃, 25-30℃, 30-35℃, 35-40℃, and greater than 40℃, forming temperature slice datasets. By using multi-dimensional slicing, we can analyze the characteristics of batteries under different operating scenarios, making subsequent feature extraction and prediction based on slice data more targeted and improving the accuracy of prediction.

[0059] S3: Sort the individual voltages of each frame in each slice dataset, and calculate the mode of the sorted values ​​of each individual in each slice dimension per preset time unit. Individuals with the same voltage are assigned the same sorted value.

[0060] The mode of the sorted values ​​mentioned above refers to the value that appears most frequently in the sorted values ​​of individual unit voltages within a preset time unit. In each slice dataset, the individual unit voltages of each frame of data are sorted in ascending order, and units with the same voltage are assigned the same sorted value. Then, using days as the preset time unit, the sorted value that appears most frequently in the daily sorted value sequence of each unit under each slice dimension is counted. For example, in a slice dataset with charging status, SOC of 30-80%, current of 30-100A, and temperature of 25-30℃, a certain frame contains 12 cells with voltages of 3.7V, 3.7V, 3.8V, 3.6V, 3.9V, 3.5V, 3.7V, 3.8V, 3.6V, 3.9V, 3.5V, and 3.7V respectively. After ascending sorting, the sorted values ​​are 1 (3.5V, 2 cells), 3 (3.6V, 2 cells), 5 (3.7V, 4 cells), 7 (3.8V, 2 cells), and 9 (3.9V, 2 cells). The sorted value of each cell in this frame is determined. Then, the sorted values ​​of this cell in all frames on that day are counted. If the sorted value "5" appears most frequently, then the mode of the sorted value of this cell in that slice dimension on that day is 5. By sorting and mode statistics, we can capture the typical sorting characteristics of cell voltage in a specific dimension and time unit, providing a stable and representative data foundation for subsequent time-series feature extraction.

[0061] S4: Extract time series features based on the mode of the sorted values ​​of each individual data point.

[0062] The aforementioned time-series features refer to those extracted from time-series data that reflect the patterns of data change, including complexity features, approximate entropy features, the average of absolute changes, and the coefficient of variation. For the mode time series of ranked values ​​of each individual data point under different slice dimensions, time-series features such as complexity, approximate entropy, the average of absolute changes, and the coefficient of variation are calculated respectively. Extracting multiple time-series features characterizes the changing patterns of the mode time-series data from different perspectives, providing rich and effective feature inputs for subsequent prediction of thermal runaway probability.

[0063] S5: Based on the timing characteristics, predict the probability of thermal runaway of the battery.

[0064] The aforementioned thermal runaway probability refers to the likelihood of a battery experiencing a thermal runaway failure. Using extracted time-series features, a suitable algorithm model is employed to predict the probability of battery thermal runaway. The time-series features of each individual cell are dimensionality-reduced (e.g., through principal component analysis) and fused to form a vehicle-level feature vector. This vector is then input into a pre-trained thermal runaway prediction model based on logistic regression. The model outputs the probability of battery thermal runaway. If the probability exceeds a preset threshold (e.g., 0.8), a risk of thermal runaway is identified. This quantitative prediction of battery thermal runaway probability provides a basis for battery safety management, facilitating preventative measures and ensuring battery safety.

[0065] This embodiment focuses on electric vehicle power batteries. First, it acquires multi-dimensional individual cell operating data from the BMS (Battery Management System). Then, it slices the data according to dimensions such as charge / discharge state, SOC (State of Charge), current, and temperature. For each slice, the individual cell voltages are sorted, and the mode of the sorted values ​​is calculated. Based on the mode time-series data, various time-series features are extracted. Finally, these features are used to predict the probability of thermal runaway. The entire process, from data acquisition to prediction, is progressive and highly targeted. Specifically, through multi-dimensional slicing, mode statistics of sorted values, and time-series feature extraction, it can accurately capture the characteristic changes of the battery under different operating states. Combined with a suitable prediction model, it achieves accurate prediction of the probability of battery thermal runaway. Compared to existing prediction methods that lack targeted segmentation and multi-feature fusion, this method has higher prediction accuracy, providing more reliable technical support for battery thermal runaway early warning and safety management.

[0066] In one embodiment, predicting the probability of thermal runaway of the battery based on the timing characteristics includes:

[0067] S51: Use a preset dimensionality reduction method to reduce the dimensionality of the temporal features of each individual to obtain dimensionality-reduced features.

[0068] The aforementioned dimensionality reduction methods refer to mapping a high-dimensional feature space to a low-dimensional feature space, thereby reducing the number of features, such as Principal Component Analysis (PCA). Reducing the feature dimensionality decreases subsequent computational costs while preserving the key information of temporal features, avoiding the curse of dimensionality caused by high-dimensional features, and improving model training and prediction efficiency. In other embodiments, dimensionality reduction methods can also include t-SNE, UMAP, or autoencoders.

[0069] S52: Flatten and merge the dimensionality-reduced features of all individual units of the same vehicle to form a vehicle-level feature vector.

[0070] The aforementioned vehicle-level feature vector refers to a feature vector that integrates the dimensionality-reduced features of all individual cells in a vehicle, reflecting the overall state of the vehicle's battery pack. Assuming a vehicle has n individual cells, each with k-dimensional features after dimensionality reduction, the dimensionality-reduced features of all cells are sequentially concatenated to form an n×k-dimensional vehicle-level feature vector. For example, an electric vehicle might have 12 individual cells, each with 2-dimensional features after dimensionality reduction. These 12 2-dimensional features are then sequentially laid out to form a 12×2=24-dimensional vehicle-level feature vector. Integrating the features of all individual cells reflects the overall state of the vehicle's battery pack, allowing subsequent predictions to be based on the comprehensive characteristics of the entire battery pack, thus improving the comprehensiveness and accuracy of the predictions.

[0071] S53: Based on the vehicle-level feature vector, predict the probability of thermal runaway of the battery using a machine learning algorithm.

[0072] The aforementioned machine learning algorithm refers to an algorithm that learns patterns from training data for prediction; here, logistic regression is used. Vehicle-level feature vectors are input into a pre-trained logistic regression model, which outputs the probability of battery thermal runaway. For example, a large number of vehicle-level feature vectors of electric vehicle batteries and their corresponding thermal runaway labels (occurred or not occurred) are collected, and this data is used to train the logistic regression model. After training, the model outputs the probability value of battery thermal runaway for a given vehicle when inputting its vehicle-level feature vector. Leveraging the powerful learning and generalization capabilities of machine learning algorithms, based on the fused vehicle-level features, the probability of battery thermal runaway is accurately predicted, achieving higher prediction accuracy compared to traditional empirical formulas or simple statistical methods. In other embodiments, the aforementioned machine learning algorithm can employ SVC (Support Vector Classifier), LightGBM (Light Gradient Boosting Machine), Random Forest, XGBoost, or deep learning models (such as LSTM) for thermal runaway risk identification.

[0073] In this embodiment, principal component analysis is used to reduce the dimensionality of the temporal features of each individual cell. The reduced dimensionality features of all cells are then fused to form a vehicle-level feature vector. A logistic regression algorithm is then used to predict the probability of thermal runaway, further optimizing the prediction process and improving its efficiency and accuracy. Specifically, dimensionality reduction reduces computational complexity, and the fusion of all individual cell features to form a vehicle-level feature vector, combined with machine learning algorithms, fully utilizes information from all cells in the battery pack. This makes thermal runaway prediction more comprehensive and accurate, with higher computational efficiency, effectively solving the problem of insufficient accuracy caused by predictions based only on a subset of cells or simple features in existing technologies.

[0074] In one specific embodiment, the aforementioned preset dimensionality reduction method is used to reduce the dimensionality of the temporal features of each individual entity to obtain dimensionality-reduced features, including:

[0075] S511: Perform principal component analysis on the temporal features of each individual to reduce the feature dimension to a preset number of dimensions, while retaining the principal components that can reflect the main information of the original features during the dimensionality reduction process.

[0076] Principal Component Analysis (PCA) is a dimensionality reduction method that transforms a set of potentially correlated variables into a set of linearly uncorrelated variables (principal components) through orthogonal transformation. Principal components are linear combinations of the original variables and retain as much information as possible about the original variables. For each individual variable's time-series features (such as complexity features, approximate entropy features, average absolute change values, coefficient of variation, etc., forming a high-dimensional feature vector), PCA is performed to determine the number of principal components, reducing the high-dimensional features to a predetermined low dimension. For example, if a certain individual variable has four time-series features, forming a 4-dimensional feature vector [C]...LZ H approx M AC [CV]. After principal component analysis, four principal components were obtained, with variance contribution rates of 50%, 35%, 10%, and 5%, respectively. The default setting was to retain the principal components with a cumulative contribution rate of 85%, i.e., selecting the first two principal components to reduce the dimensionality of the 4-dimensional features to 2-dimensionality. Through principal component analysis, while reducing the number of features, the key information of the original time-series features is preserved to the greatest extent, avoiding overfitting and other problems caused by excessive feature dimensionality during subsequent model training, and improving computational efficiency.

[0077] S512: Calculate the contribution rate of each principal component to ensure that the preset number of dimensions after dimensionality reduction can retain the main changing trends and differences of the temporal features of each individual.

[0078] The principal component contribution rate refers to the proportion of the variance of a particular principal component to the total variance of all principal components, reflecting the degree to which that principal component contributes to the information of the original variable. The variance contribution rate of each principal component is calculated, and the number of principal components to be retained is determined by the cumulative contribution rate. This ensures that the retained principal components can cover the main trends of change in the original time-series features and the differences between individual components. For example, continuing the previous example, the variances of the four principal components are λ1, λ2, λ3, and λ4, and the total variance is λ1 + λ2 + λ3 + λ4. The contribution rate of the first principal component is λ1 / (λ1 + λ2 + λ3 + λ4) = 50%, and the contribution rate of the second principal component is λ2 / (λ1 + λ2 + λ3 + λ4) = 35%. The cumulative contribution rate of the first two principal components reaches 85%, indicating that these two principal components can retain most of the information of the original 4-dimensional features. By calculating the contribution rate, the number of principal components after dimensionality reduction is scientifically determined, ensuring that the dimensionality-reduced features can effectively reflect the main trends of change in the time-series features of individual components and the differences between different individuals, providing high-quality dimensionality-reduced features for subsequent vehicle-level feature fusion and thermal runaway prediction.

[0079] In this embodiment, principal component analysis (PCA) is used to reduce the dimensionality of the temporal features of each individual entity. The number of principal components to be retained is determined by calculating the contribution rate of the principal components, thus achieving the goal of reducing feature dimensionality while preserving the key information of the original features to the greatest extent. Specifically, PCA for dimensionality reduction, compared with simple feature selection or unfounded dimensionality reduction methods in the prior art, can more scientifically and efficiently reduce feature dimensionality while retaining the main information of the temporal features. This provides concise features rich in key information for subsequent vehicle-level feature fusion and thermal runaway prediction, improving the training efficiency and prediction accuracy of the prediction model, and effectively solving the problems of computational complexity and overfitting caused by high-dimensional features.

[0080] In one embodiment, the above-mentioned prediction of the probability of thermal runaway of the battery based on the vehicle-level feature vector using a machine learning algorithm includes:

[0081] S531: Input the vehicle-level feature vector into the pre-trained thermal runaway prediction model and output the probability value of the battery thermal runaway. The thermal runaway prediction model is a model trained based on historical vehicle-level feature vectors, using a logistic regression model, a support vector classifier, or a lightweight gradient booster model.

[0082] The aforementioned thermal runaway prediction model refers to a machine learning model that predicts the probability of thermal runaway by training and learning the relationship between vehicle-level feature vectors and thermal runaway labels (occurred or not occurred) in historical data. Logistic regression, Support Vector Classifier (SVC), or Lightweight Gradient Boosting Machine (LightGBM) models can be used here. A large number of historical vehicle-level feature vectors of electric vehicle batteries are collected, along with corresponding thermal runaway occurrence labels (1 for occurrence, 0 for non-occurrence). This data is used to train the thermal runaway prediction model. After training, the vehicle-level feature vector to be predicted is input into the model, and the model outputs the probability value of battery thermal runaway. For example, 1000 sets of historical vehicle-level feature vectors (e.g., 24-dimensional) of electric vehicle batteries and their corresponding thermal runaway labels are collected, of which 300 sets indicate thermal runaway occurred (label 1), and 700 sets indicate non-occurrence (label 0). This data is used to train the LightGBM model. After training, when a 24-dimensional vehicle-level feature vector of a certain vehicle is input, the model outputs a probability of 0.75 for the vehicle's battery to experience thermal runaway. By leveraging the powerful learning capabilities of machine learning models, the complex relationship between vehicle-level feature vectors and thermal runaway is extracted from historical data, enabling accurate prediction of the probability of thermal runaway. Compared with traditional empirical formulas, the prediction accuracy is higher.

[0083] S532: When the probability value exceeds the preset threshold, a thermal runaway risk warning is triggered.

[0084] The aforementioned preset threshold refers to a manually set probability threshold for judging whether a battery has a risk of thermal runaway. Based on battery safety requirements and historical data statistics, a preset threshold for the probability of thermal runaway is set (e.g., 0.8). When the probability value output by the model exceeds this threshold, a corresponding thermal runaway risk warning mechanism is triggered, such as sending a warning signal to the vehicle control system or displaying a notification on the in-vehicle display screen. For example, if the preset threshold is set to 0.8, and the predicted probability of thermal runaway for a vehicle battery is 0.85, exceeding the threshold, a thermal runaway risk warning is triggered, and the in-vehicle system issues an audible and visual alarm to alert the driver to battery safety. By triggering warnings through preset thresholds, timely alerts can be issued before battery thermal runaway occurs, buying time for preventative measures (such as reducing charging power or activating the cooling system) and effectively ensuring battery safety.

[0085] In this embodiment, a thermal runaway prediction model is trained using historical data. Vehicle-level feature vectors are input into the model to predict the probability of thermal runaway, and a thermal runaway warning is triggered by a preset threshold, forming a complete process from feature input to probability prediction and then to risk warning. Using a machine learning model for thermal runaway probability prediction, compared to existing prediction methods based on simple statistics or single models, can more accurately capture the relationship between vehicle-level features and thermal runaway, resulting in higher prediction accuracy. Combined with the preset threshold warning mechanism, timely alerts to thermal runaway risks are achieved, providing an effective means for battery safety management and greatly improving the safety of battery use.

[0086] In one embodiment, the specified data is sliced ​​according to a preset segmentation rule to obtain sliced ​​datasets for each dimension, including:

[0087] S21: Slice by charge / discharge state dimension: When the current > 0, it is divided into a charging state dataset, and when the current < 0, it is divided into a discharging state dataset.

[0088] The aforementioned charge / discharge states refer to whether the battery is in the charging or discharging process, determined by the direction of the current. Based on the sign of the current, the data is divided into two datasets: charging and discharging. For example, when an electric vehicle's power battery is charging, the current is >0A, and the data collected at this time is classified as a charging state dataset; when discharging while driving, the current is <0A, and the data is classified as a discharging state dataset. The chemical reactions and heat generation characteristics of the battery differ during charging and discharging; analyzing them separately allows for more accurate capture of battery characteristics under different states, providing more targeted data for subsequent thermal runaway prediction.

[0089] S22: Slicing by SOC dimension: Divide the SOC into multiple continuous intervals from 0 to 100%, and form SOC slice datasets for each interval.

[0090] SOC (State of Charge) refers to the battery's state of charge. SOC slice datasets are datasets divided into segments based on SOC intervals. Specifically, the SOC range of 0-100% is divided into multiple consecutive intervals, forming corresponding slice datasets. For example, if the SOC is divided into three intervals: 0-30%, 30-80%, and 80-100%, then when the battery's SOC is 45% at a certain moment, the data for that moment is assigned to the 30-80% SOC slice dataset. Within different SOC intervals, the battery's internal resistance, polarization, and other characteristics differ. SOC slicing allows for analysis of battery characteristics at different remaining charge stages, improving the accuracy of subsequent predictions.

[0091] S23: Slicing by current dimension: Divide the current into current intervals less than a first current threshold, current intervals greater than a second current threshold, and one or more current intervals between the first current threshold and the second current threshold, and form current slice datasets for the corresponding intervals; wherein, the second current threshold is greater than the first current threshold.

[0092] Current thresholds refer to the critical values ​​used to divide current ranges. The first and second current thresholds are used to define different current intensity ranges. Setting the first current threshold to 30A and the second to 100A, the current is divided into three ranges: less than 30A, 30-100A, and greater than 100A, forming corresponding current slice datasets. For example, when the charging current is 25A, the data is assigned to the less than 30A current slice dataset; when the current is 60A, it is assigned to the 30-100A current slice dataset; and when the current is 120A, it is assigned to the greater than 100A current slice dataset. Under different current intensities, the battery's heating and electrochemical reaction rates differ. Current slices allow for the analysis of battery characteristics under different current loads, providing a more accurate basis for predicting thermal runaway.

[0093] S24: Slicing by temperature dimension: Divide the temperature into temperature ranges below a first temperature threshold, temperature ranges above a second temperature threshold, and one or more temperature ranges between the first temperature threshold and the second temperature threshold, and form temperature slice datasets for the corresponding ranges; wherein the second temperature threshold is greater than the first temperature threshold.

[0094] The aforementioned temperature thresholds refer to the critical values ​​for dividing temperature ranges. The first and second temperature thresholds are used to define different temperature ranges. Setting the first temperature threshold to 20℃ and the second temperature threshold to 40℃, the temperature is divided into six ranges: below 20℃, 20-25℃, 25-30℃, 30-35℃, 35-40℃, and above 40℃, forming corresponding temperature slice datasets. For example, when the battery cell temperature is 18℃, the data is assigned to the temperature slice dataset below 20℃; when the temperature is 28℃, it is assigned to the temperature slice dataset between 25-30℃; and when the temperature is 42℃, it is assigned to the temperature slice dataset above 40℃. Temperature is a key factor affecting battery thermal runaway. The thermal stability and the degree of side reactions of batteries differ in different temperature ranges. Temperature slices allow for precise analysis of battery characteristics at different temperatures, greatly improving the accuracy of thermal runaway prediction.

[0095] This embodiment takes an electric vehicle power battery as an example, and performs segmentation processing based on four dimensions: state of charge / discharge, state of charge (SOC), current, and temperature, according to clearly defined thresholds or interval rules. The state of charge / discharge distinguishes between the charging and discharging processes; SOC is divided into three typical intervals; current is divided into three intervals with thresholds of 30A and 100A; and temperature is refined into six intervals with thresholds of 20℃ and 40℃, comprehensively covering data under different battery operating scenarios. Through multi-dimensional and refined segmentation processing, battery operating data is classified according to different states, charge levels, current intensities, and temperature intervals. Compared to the coarse, unsegmented, or single-dimensional segmentation methods in existing technologies, this approach allows for more precise subsequent analysis of battery characteristics under different operating conditions. It provides a more targeted data foundation that better reflects the true state of the battery for sorting value mode statistics, time-series feature extraction, and thermal runaway probability prediction, greatly improving the accuracy and reliability of subsequent thermal runaway prediction.

[0096] In one embodiment, the aforementioned time-series features include complexity features, approximate entropy features, the average value of absolute change, and the coefficient of variation; the extraction of time-series features based on the mode of the sorted values ​​of each individual entity includes:

[0097] S41: For each individual unit in the sorted value mode time series under each slice dimension, calculate its complexity feature. The complexity is quantified by the ratio of the minimum number of distinct substrings into which the sequence is decomposed to the sequence length, combined with the logarithm of the sequence length.

[0098] The aforementioned complexity is a metric for measuring the complexity of a sequence. It reflects the complexity of a sequence by calculating the minimum number of distinct substrings into which the sequence can be decomposed. For example, consider the mode time series of sorted values ​​of a single entity i in slice dimension d. The Lempel-Ziv algorithm is used to determine the minimum number K(N) of distinct substrings that can be decomposed. Combined with the sequence length N, the formula is used to... Computational complexity characteristics. For example, the mode time series sequence of the sorted values ​​of a certain single entity in the charging state and SOC 30-80% slice dimension is [5,5,6,5,7,6,5] (N=7). It is decomposed into [5],[5,6],[5,7],[6,5] by the Lempel-Ziv algorithm, which has a total of K(7)=4 distinct substrings. Then the complexity characteristics are as follows. Complexity features can reflect the degree of complexity of the mode time series of sorted values. Compared with existing technologies that only use simple statistics (such as mean and variance), they can more meticulously characterize the dynamic characteristics of the series and provide richer information for thermal runaway prediction.

[0099] S42: Introduce a dynamic time window related to sequence variance and length, and quantify the similarity difference of sequence substrings under different embedding dimensions to obtain approximate entropy features.

[0100] The aforementioned approximate entropy is an indicator used to measure the irregularity and complexity of time series. It reflects the complexity of the sequence by analyzing the similarity differences of sequence substrings under different embedding dimensions. The dynamic time window refers to a time range that is dynamically adjusted based on the sequence variance and length, and is used for approximate entropy calculation. This is related to the introduction of sequence variance. Dynamic time window related to length N Given an embedding dimension m and a similarity tolerance r, calculate the number of sequences whose distance to the first m elements of the sequence is less than r under different embedding dimensions. And thus obtain and Approximate entropy features are obtained by using the difference between the two. ,in , Let be the number of sequences whose distance to the first m elements of the sequence is less than r, where m is the embedding dimension and r is the similarity tolerance. Specifically, in calculating... Time is measured in dynamic time windows. Given a range, iterate through the substrings within the window, checking if their distance from the first m elements of the sequence is less than the similarity tolerance r, and count the number of substrings that meet the condition. For example, when the dynamic time window... When the value is 20, the system searches for substrings that meet the distance requirement within a time range of length 20. The selection of the embedding dimension *m* and the similarity tolerance *r* is based on statistical analysis of a large amount of battery thermal runaway experimental data. After testing time-series data of different battery types (such as ternary lithium batteries and lithium iron phosphate batteries) at different charge and discharge stages, it was found that when *m* is between 2 and 4 and *r* is 0.1 to 0.2 times the standard deviation of the sequence, the approximate entropy feature has the highest sensitivity to battery thermal runaway. Taking ternary lithium batteries as an example, the standard deviation of their voltage sequences is usually between 0.05V and 0.2V. In this case, selecting *r* = 0.15 × standard deviation can effectively distinguish the difference in sequence complexity between the normal state and the pre-thermal runaway state. For example, the variance of a sequence... If the length N=10, then the dynamic time window Set m=2, r=0.5, calculate... and Assuming =0.8, =0.6, then the approximate entropy characteristic =0.8-0.6=0.2. The approximate entropy feature can effectively capture the irregular changes in the sequence. The introduction of the dynamic time window makes it more adaptable to the characteristics of the sequence itself. Compared with the fixed window approximate entropy calculation in the existing technology, the approximate entropy calculated by this method can more accurately reflect the complexity of the mode time series of the sorted value and improve the accuracy of thermal runaway prediction.

[0101] S43: The average value of the absolute change value is obtained by solving the arithmetic mean of the absolute differences of the modes of adjacent sorted values ​​in the time series.

[0102] The average of absolute changes is the arithmetic mean of the absolute differences between adjacent elements in a time series, used to measure the degree of volatility of the series. For time series with the most significant changes in sorted values... Calculate the absolute difference between adjacent elements, and then calculate the arithmetic mean of these absolute differences. For example, a certain sequence =[5,6,5,7,6], the absolute differences between adjacent values ​​are |6-5|=1, |5-6|=1, |7-5|=2, |6-7|=1, and the arithmetic mean is (1+1+2+1) / (4)=1.25. This feature can intuitively reflect the fluctuation of the mode time series of the sorted values. Compared with using only variance, it focuses more on the changes between adjacent times, providing information on the short-term fluctuations of the series for thermal runaway prediction.

[0103] S44: The coefficient of variation is obtained by solving the ratio of the standard deviation to the mean of the time series.

[0104] The coefficient of variation mentioned above is the ratio of the standard deviation to the mean, used to measure the dispersion of the series and eliminate the influence of dimensions. First, calculate the mode of the sorted time series and the mean. and standard deviation Then calculate the ratio of the two to obtain the coefficient of variation. The coefficient of variation can eliminate the influence of dimensions and objectively reflect the degree of dispersion of the sequence. Compared with simply using the standard deviation, it can more accurately compare the fluctuation of the mode time series of sorted values ​​under different slice dimensions, providing more comparable characteristics for thermal runaway prediction.

[0105] This embodiment calculates four types of time-series features for the mode of the ranking values ​​of individual battery cells in electric vehicles across various slice dimensions: complexity features, approximate entropy features, average absolute change values, and coefficient of variation. Using the Lempel-Ziv complexity formula and the approximate entropy formula incorporating a dynamic time window, the complexity, irregularity, volatility, and dispersion of the sequence are characterized from different perspectives, providing rich feature inputs for subsequent thermal runaway prediction. Specifically, compared to existing technologies that only use single or simple features (such as threshold judgments for single parameters like voltage and temperature), the multi-dimensional time-series features extracted by this method can more comprehensively and deeply reflect the dynamic changes in the mode of the battery cell voltage ranking values. Complexity features capture the diversity of sequence patterns, approximate entropy features reflect sequence irregularity, the average absolute change value reflects volatility, and the coefficient of variation measures dispersion. These features complement each other, providing richer and more discriminative information for thermal runaway prediction, significantly improving the accuracy and reliability of thermal runaway prediction, and effectively solving the problems of incomplete feature extraction and inaccurate capture of dynamic changes in battery state in existing technologies.

[0106] In one specific embodiment, when training a thermal runaway prediction model (such as the LightGBM model), feature weights are dynamically optimized using a gradient boosting algorithm, enabling the model to adaptively learn the contribution of each feature to thermal runaway prediction. For example, in the training data of ternary lithium batteries, the approximate entropy feature (reflecting sequence irregularity) accounts for approximately 35% of the weight, complexity features (reflecting sequence pattern diversity) account for approximately 30%, the coefficient of variation (reflecting dispersion) accounts for approximately 25%, and the average absolute change value (reflecting volatility) accounts for approximately 10%. This weight distribution is automatically generated by the model based on the feature importance of thermal runaway cases in historical data.

[0107] To verify the necessity of multi-feature fusion, a comparative experiment was conducted: 100 sets of ternary lithium battery data (including 30 sets of thermal runaway precursor samples) from the same batch were selected, and predictions were performed using both "single feature" and "multi-feature fusion" methods. The experimental results are as follows:

[0108] Using only approximate entropy features: prediction accuracy was 75%, and false positive rate was 15%;

[0109] Using only complexity features: accuracy was 70%, false positive rate was 18%;

[0110] Using only the coefficient of variation: accuracy was 68%, and the false alarm rate was 20%.

[0111] Multi-feature fusion: accuracy improved to 92%, false positive rate reduced to 5%.

[0112] The experiments above demonstrate that a single feature can only reflect the battery state from a specific dimension (e.g., the approximate entropy feature is sensitive to sequence disorder in the early stages of thermal runaway, but insensitive to short-term voltage fluctuations), while multi-feature fusion can comprehensively capture battery anomalies through complementarity, significantly improving prediction accuracy. For example, when a micro-short circuit occurs inside the battery, the complexity feature increases due to the increased diversity of voltage sorting patterns, and the approximate entropy feature increases due to the enhanced sequence irregularity. Combining the two can identify risks earlier, providing an earlier warning compared to a single feature.

[0113] In one embodiment, the above-mentioned sorting of the individual voltages of each frame of data in each slice dataset and calculating the mode of the sorted values ​​of each individual unit per preset time unit under each slice dimension includes:

[0114] S31: Based on each frame of data in each slice dimension, sort the voltages of all cells in the battery in ascending order to generate a corresponding sorted value sequence. Cells with the same voltage are assigned the same sorted value. The sorted value range is [1, M], where M is the total number of cells. The sorted value calculation satisfies the formula:

[0115]

[0116] in, Let i be the voltage sorting value of a single cell i in slice dimension d. This is an indicator function, where K is the number of cells with equal voltage.

[0117] The ascending order mentioned above refers to arranging the data from smallest to largest. The sorted value sequence refers to the sequential numbering sequence corresponding to the voltage of each individual cell after sorting; cells with the same voltage are assigned the same sorted value. Within each slice dimension, for each frame of data, the voltages of all cells within the battery are arranged from smallest to largest to determine the sorting position of each cell's voltage. Cells with the same voltage are assigned the same sorted value, forming the sorted value sequence for that frame of data. For example, an electric vehicle battery has 12 cells. In a slice dataset with charging state, SOC 30-80%, current 30-100A, and temperature 25-30℃, the voltages of each cell in a certain frame are 3.5V, 3.6V, 3.7V, 3.5V, 3.8V, 3.7V, 3.9V, 3.7V, 3.6V, 3.8V, 3.5V, and 3.7V. The voltages, when sorted in ascending order, are 3.5V, 3.5V, 3.5V, 3.6V, 3.6V, 3.7V, 3.7V, 3.7V, 3.7V, 3.8V, 3.8V, and 3.9V. Calculate the ranking value for each voltage. For a single cell with a voltage of 3.5V, there are three identical voltages; the ranking value is calculated as follows. ,in Therefore, the sorting value for a 3.5V cell is 3+1=4. Similarly, the sorting value for a 3.6V cell is 6; the sorting value for a 3.7V cell is 6; the sorting value for a 3.8V cell is 10; and the sorting value for a 3.9V cell is 12. Therefore, the sorting value sequence for the 12 cells in this frame is 4, 6, 6, 4, 10, 6, 12, 6, 6, 10, 4, 6. By sorting the cell voltages and assigning the same sorting value to cells with the same voltage, the relative positional relationship of cell voltages within the battery can be effectively reflected, facilitating the detection of locally abnormal cells. Compared to existing methods that do not consider the relative sorting of cell voltages, this method can more sensitively capture abnormal distributions of cell voltages within the battery. (The above formula...) This method scientifically solves the problem of sorting individual cells with the same voltage, ensuring that cells with the same voltage have the same sorting value. This guarantees the rationality and consistency of the sorting. Compared with the simple ascending sorting in the existing technology that ignores the case of the same voltage, the sorting value calculated by this formula more accurately reflects the relative position of the individual cell voltages, providing a more reliable sorting data basis for subsequent analysis.

[0118] Using days as the preset time unit, based on the daily sorted value sequence of each individual unit under each slice dimension, the sorted value with the highest frequency is counted as the mode of the sorted value of that unit on that day.

[0119] Using days as the preset time unit, the sorted values ​​of all frames for each unit within each slice dimension are collected for that day, forming a daily sorted value sequence. The sorted value that appears most frequently in this sequence is taken as the mode of the sorted values ​​for that unit within that slice dimension for that day. For example, for a unit in the charging state, under the slice dimension of SOC 30-80%, there are 200 frames of data on a certain day. In its sorted value sequence, the sorted value "6" appears 120 times, "5" appears 50 times, "7" appears 30 times, and other sorted values ​​appear even less frequently. Therefore, the mode of the sorted values ​​for that unit within that slice dimension for that day is 6. Calculating the mode on a daily basis can filter out short-term fluctuations, retain the typical characteristics of unit voltage sorting, and provide stable and representative data points for subsequent time-series feature extraction. Compared to real-time sorted values, the mode better reflects the typical relative position of unit voltage within a day, improving the reliability of subsequent analysis.

[0120] In this embodiment, the individual cell voltages of each frame in each slice dataset are sorted in ascending order. A specific formula is used to calculate the sorting value (single cells with the same voltage are assigned the same sorting value). The mode of the daily sorting value sequence for each cell is counted on a daily basis, forming time-series data of the sorting value mode, which provides a foundation for subsequent time-series feature extraction. Specifically, by sorting the individual cell voltages and counting the mode of the sorting values, the relative distribution characteristics and typical features of the individual cell voltages in the battery over time can be effectively captured. The sorting value reflects the relative position of the individual cell voltages in the battery, making it easier to identify locally abnormal cells; the daily mode statistics filter out short-term fluctuations and retain typical features. Compared with existing methods that do not consider the relative sorting of individual cell voltages and typical features over time, this method provides more targeted and more reflective basic data for subsequent thermal runaway prediction, significantly improving the accuracy and reliability of thermal runaway prediction and effectively solving the problem of difficulty in detecting local anomalies inside the battery in existing technologies.

[0121] In one embodiment, obtaining the voltage and specified data of each cell in the battery includes:

[0122] The process involves acquiring raw data such as voltage, current, temperature, and SOC of each cell in the battery. Outliers are then filtered and processed using the 3σ principle or IQR (Interquartile Range) method. Specifically, the mean and standard deviation of the data are calculated, and data points exceeding the mean ± 3σ range are defined as outliers and corrected or deleted. Physically impossible values ​​in the raw data are checked and corrected; for example, if a reasonable voltage range is set to 2V to 5V, values ​​outside this range are replaced with boundary values ​​or deleted. Missing values ​​(NaN) in the raw data are detected and filled using linear interpolation or forward imputation methods.

[0123] The aforementioned outliers refer to data points that deviate from the normal distribution range of the data. The 3σ principle is an outlier detection method based on normal distribution, which considers data to be distributed within the range of mean ± 3 times the standard deviation; those exceeding this range are considered outliers. The mean and standard deviation of voltage or specified data (such as current) are calculated, and data points exceeding the mean ± 3σ range are identified as outliers, which are then corrected or deleted. For example, if the mean of a single battery cell voltage in an electric vehicle is 3.7V and the standard deviation is 0.1V, according to the 3σ principle, the voltage range should be 3.7 ± 3 × 0.1 = 3.4V to 4.0V. If a single cell voltage in a certain frame of data is 2.5V, significantly lower than 3.4V, it is identified as an outlier. It can be corrected using the voltage value at an adjacent time (e.g., 3.7V in the previous frame) or the outlier data point can be directly deleted. Effectively eliminating abnormal fluctuations in the data avoids interference from outliers in subsequent analyses (such as slicing and sorting), ensuring the reliability and accuracy of the data, and improving the quality of the basic data for thermal runaway prediction.

[0124] The aforementioned physically impossible values ​​refer to unreasonable data that do not conform to the physical characteristics of the battery, such as a voltage below 2V or above 5V (assuming the normal battery voltage range is 2-5V). Reasonable ranges are set for each data point, and data exceeding these ranges are identified as erroneous values ​​and replaced with boundary values ​​or deleted. For example, if a single cell voltage data point is recorded as 1.5V, which is below the lower limit of the normal range of 2V, it is determined to be an erroneous value and replaced with the lower limit of 2V; if the current data point is recorded as 200A, while the battery's maximum allowable current is 150A, it is determined to be an erroneous value and the data point is deleted. Correcting data that clearly violates the physical characteristics of the battery ensures that the data conforms to actual operating patterns, avoids deviations in subsequent analysis due to erroneous data, and provides reasonable and effective data support for thermal runaway prediction.

[0125] The missing values ​​mentioned above refer to blank data (NaN) that were not recorded during data acquisition. Linear interpolation is a method of filling in missing values ​​by linearly fitting adjacent known data points. It detects the location of missing values ​​in the original data, uses adjacent known data points, and calculates the estimated value of the missing value using the linear interpolation formula to fill in the missing value. For example, in a single-unit temperature data sequence [25℃, NaN, 28℃], the missing value is located in the middle. Through linear interpolation, it is calculated that: 25 + (28 - 25) × (1 / 2) = 26.5℃, thus filling in the missing value as 26.5℃. Filling in the missing parts during data acquisition ensures data integrity, avoids data gaps in subsequent slicing, sorting, and other analytical steps due to missing data, and improves data usability and analytical continuity.

[0126] In this embodiment, after acquiring the voltage, current, temperature, and SOC data of each cell, outlier processing, error value processing, and missing value processing are performed sequentially to ensure data quality. A standardized data cleaning process lays a reliable data foundation for subsequent multi-dimensional slicing, voltage sorting, and time-series feature extraction. Specifically, compared to existing methods that ignore data preprocessing or only perform simple cleaning, this embodiment significantly improves the accuracy, rationality, and completeness of the data through systematic outlier, error, and missing value processing. High-quality data avoids misjudgments caused by data noise, enabling subsequent feature extraction based on slicing and sorting to better reflect the true state of the battery, thereby improving the accuracy and reliability of thermal runaway prediction and effectively solving the problem of inaccurate prediction results due to poor data quality in existing technologies.

[0127] Reference Figure 2 This application also provides a battery thermal runaway prediction device based on segmented features, used to implement the battery thermal runaway prediction method based on segmented features described in any of the above embodiments, including:

[0128] The acquisition unit 10 is used to acquire the voltage and specified data of each cell in the battery, wherein the specified data includes at least one or more of current, temperature and SOC data;

[0129] Segmentation unit 20 is used to slice the specified data according to a preset segmentation rule to obtain sliced ​​datasets of each dimension.

[0130] The sorting unit 30 is used to sort the individual voltages of each frame of data in each slice dataset, and to count the mode of the sorting value of each individual in each slice dimension per preset time unit. Individuals with the same voltage are assigned the same sorting value.

[0131] Extraction unit 40 is used to extract time series features based on the mode time series data of the sorted values ​​of each individual unit;

[0132] The prediction unit 50 is used to predict the probability of thermal runaway of the battery based on the time-series characteristics.

[0133] Reference Figure 3 This application also provides a computer device whose internal structure can be as follows: Figure 3 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor is designed to provide computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores operating devices, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the aforementioned specified data. The network interface is used to communicate with external terminals via a network connection. Furthermore, the computer device may also include input devices and a display screen. When the computer program is executed by the processor, it implements the battery thermal runaway prediction method based on segmented features described in any of the above embodiments. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0134] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the battery thermal runaway prediction method based on segmented features described in any of the above embodiments. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0135] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0136] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0137] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting thermal runaway of a battery based on segmented features, the method comprising: The method comprises: acquiring the voltage and specified data of each single cell in the battery, the specified data at least including one or more of current, temperature and SOC data; slicing the specified data according to a preset segmentation rule of dimensions to obtain slice data sets of each dimension; sorting the single cell voltage of each frame of data in each slice data set, and counting the mode value of the sorting value of each single cell per preset time unit under each slice dimension, wherein single cells with the same voltage are assigned the same sorting value; extracting time sequence features based on the mode value time sequence data of each single cell; predicting the probability of thermal runaway of the battery based on the time sequence features; The time sequence features include complexity features, approximate entropy features, average value of absolute change value and coefficient of variation; the extraction of the time sequence features based on the mode value time sequence data of each single cell comprises: calculating the complexity features of the mode value time sequence sequence of each single cell under each slice dimension, wherein the complexity is quantified by the ratio of the minimum number of different substrings to the length of the sequence, and the logarithm of the length of the sequence is combined; introducing a dynamic time window related to the variance and length of the sequence, and quantifying the similarity difference of the sequence substring under different embedding dimensions to obtain the approximate entropy features; obtaining the average value of the absolute change value by solving the arithmetic mean of the absolute difference value of adjacent sorting values in the time sequence sequence; obtaining the coefficient of variation by solving the ratio of the standard deviation to the mean of the time sequence sequence.

2. The segment-feature-based battery thermal runaway prediction method according to claim 1, wherein, Predicting the probability of thermal runaway of the battery based on the time sequence features comprises: using a preset dimension reduction method to reduce the time sequence features of each single cell to obtain reduced features; flattening and fusing the reduced features of all single cells of the same vehicle to form a vehicle-level feature vector; predicting the probability of thermal runaway of the battery based on the vehicle-level feature vector through a machine learning algorithm.

3. The segmented feature based battery thermal runaway prediction method of claim 2, wherein, The use of a preset dimension reduction method to reduce the time sequence features of each single cell to obtain reduced features comprises: performing principal component analysis on the time sequence features of each single cell to reduce the feature dimension to a preset number of dimensions, wherein the principal components that can reflect the main information of the original features are retained during the dimension reduction process; calculating the contribution rate of each principal component to ensure that the preset number of dimensions of the reduced features can retain the main change trend and difference information of the time sequence features of each single cell. 4.The method of claim 2, wherein, The prediction of the probability of thermal runaway of the battery based on the vehicle-level feature vector through a machine learning algorithm comprises: inputting the vehicle-level feature vector into a pre-trained thermal runaway prediction model to output the probability value of the battery thermal runaway, wherein the thermal runaway prediction model is a model trained based on historical vehicle-level feature vectors based on a logistic regression model, a support vector classifier or a lightweight gradient boosting machine model; when the probability value exceeds a preset threshold, triggering a thermal runaway risk warning.

5. The segment-feature-based battery thermal runaway prediction method of claim 1, wherein, The slicing of the specified data according to a preset segmentation rule of dimensions to obtain slice data sets of each dimension comprises: charging and discharging state slicing: when the current is > 0, the charging state data set is divided, and when the current is < 0, the discharging state data set is divided; Slicing according to the SOC dimension: dividing the SOC into multiple continuous intervals according to 0-100%, and forming a SOC slice data set corresponding to each interval; Slicing according to the current dimension: dividing the current into a current interval less than a first current threshold, a current interval greater than a second current threshold, and one or more current intervals between the first current threshold and the second current threshold, and forming a current slice data set corresponding to each interval; wherein the second current threshold is greater than the first current threshold; Slicing according to the temperature dimension: dividing the temperature into a temperature interval less than a first temperature threshold, a temperature interval greater than a second temperature threshold, and one or more temperature intervals between the first temperature threshold and the second temperature threshold, and forming a temperature slice data set corresponding to each interval; wherein the second temperature threshold is greater than the first temperature threshold.

6. The segment-feature-based battery thermal runaway prediction method of claim 1, wherein, The sorting of the single cell voltage of each frame data in each slice data set, and the statistics of the mode value of the sorting value of each single cell per preset time unit under each slice dimension, include: Based on each frame data under each slice dimension, the voltages of all single cells in the battery are arranged in ascending order to generate a corresponding sorting value sequence, wherein single cells with the same voltage are assigned the same sorting value, and the sorting value range is [1, M], M is the total number of single cells; Taking a day as a preset time unit, based on the daily sorting value sequence of each single cell under each slice dimension, the sorting value with the highest occurrence frequency is counted as the mode value of the sorting value of the single cell on that day; The sorting value calculation satisfies the formula: wherein, is the voltage rank value of monomer i in slice dimension d, is the indicator function, K is the number of monomers with equal voltage.

7. A battery thermal runaway prediction apparatus based on segmented features, characterized by, A battery thermal runaway prediction method based on segmented features is provided, including: An acquisition unit configured to acquire voltages of single cells in the battery and specified data, the specified data including at least one or more of current, temperature, and SOC data; A segmentation unit configured to slice the specified data according to a segmentation rule of a preset dimension to obtain slice data sets of each dimension; An ordering unit configured to sort the single cell voltage of each frame data in each slice data set, and to count the mode value of the sorting value of each single cell per preset time unit under each slice dimension, wherein single cells with the same voltage are assigned the same sorting value; An extraction unit configured to extract time series features based on the time series data of the mode value of each single cell; A prediction unit configured to predict the probability of thermal runaway of the battery based on the time series features.

8. A computer device comprising a memory and a processor, the memory having stored therein a computer program, characterized in that, The processor executes the computer program to implement the steps of the battery thermal runaway prediction method based on segmented features as claimed in any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the battery thermal runaway prediction method based on segmented features as claimed in any one of claims 1 to 6.

Images