Battery thermal runaway prediction method and computer equipment based on cross-feature fusion

Abstract

This invention belongs to the field of battery safety control technology, and specifically relates to a battery thermal runaway prediction method and computer device based on cross-feature fusion. The method includes: acquiring battery operating condition features, time-series features, electrical features, and segmented features, wherein the operating condition features reflect the actual usage environment of the battery, the time-series features reflect the time-series changes in the battery state, the electrical features reflect the battery's electrical performance, and the segmented features reflect feature statistics or changes within a specific interval; calculating the cross-features of the battery's operating condition features, time-series features, electrical features, and segmented features to generate a fused cross-feature matrix; and predicting the probability of the battery experiencing thermal runaway based on the cross-feature matrix. This method overcomes the limitations of single features, improves prediction accuracy through multi-dimensional features and correlation mining, and is adaptable to multiple scenarios such as electric vehicles and energy storage systems, achieving automated and high-precision management of battery safety and effectively reducing the risk of thermal runaway.

Description

Technical Field

[0001] This invention relates to the field of battery safety control technology, and in particular to a battery thermal runaway prediction method and computer device based on cross-feature fusion. Background Technology

[0002] In battery thermal runaway prediction, existing technologies achieve risk assessment through nonlinear combination of multi-dimensional features and recurrent neural networks, but they do not incorporate actual operating conditions features such as overvoltage charging times and fast charging frequency, and lack refined feature extraction within the SOC (State of Charge) range and temperature range, resulting in weak local anomaly identification capabilities.

[0003] Battery thermal runaway refers to the phenomenon where, under abnormal operating conditions such as overcharging, over-discharging, high temperature, or mechanical damage, the internal chemical reactions of a battery become uncontrolled, leading to a rapid rise in temperature, combustion, or explosion. Thermal runaway not only threatens equipment safety but can also cause personal injury and property damage. Existing prediction systems mostly rely on single-dimensional features (such as voltage, temperature, and current), lacking in-depth analysis of combined features and operating condition characteristics, resulting in insufficient prediction accuracy.

[0004] In addition, although existing technologies can achieve risk assessment through nonlinear combination of multi-dimensional features and recurrent neural networks, they do not incorporate actual operating conditions features such as overvoltage charging times and fast charging frequency, and lack refined feature extraction within the SOC range and temperature range, resulting in weak local anomaly identification capabilities and insufficient prediction accuracy. Summary of the Invention

[0005] The main objective of this invention is to provide a battery thermal runaway prediction method and computer device based on cross-feature fusion, aiming to solve the technical problem of insufficient prediction accuracy of battery thermal runaway in existing technologies.

[0006] To achieve the aforementioned objectives, the first aspect of this invention proposes a battery thermal runaway prediction method based on cross-feature fusion, the method comprising:

[0007] The battery's operating condition characteristics, time-series characteristics, electrical characteristics, and segmented characteristics are obtained. The operating condition characteristics reflect the actual usage environment of the battery, the time-series characteristics reflect the time-series changes in the battery's state, the electrical characteristics reflect the battery's electrical performance, and the segmented characteristics reflect the characteristic statistics or changes within a specific interval.

[0008] Calculate the cross features of the battery's operating condition characteristics, timing characteristics, electrical characteristics, and segmentation characteristics, and generate a fused cross feature matrix;

[0009] Based on the cross-feature matrix, the probability of thermal runaway of the battery is predicted.

[0010] Furthermore, the operating condition characteristics include: overvoltage charging characteristics, ultra-low temperature charging characteristics, parking time characteristics, fast charging characteristics, and deep discharge characteristics. The overvoltage charging characteristics refer to the relevant statistics of the battery charging voltage exceeding a safety threshold, and the ultra-low temperature charging characteristics refer to the relevant statistics of the ambient temperature of the battery charging being lower than a preset temperature threshold.

[0011] Furthermore, the overvoltage charging characteristics include: the number of overvoltage charging cycles, the cumulative overvoltage charging time, and the average overvoltage amplitude;

[0012] The ultra-low temperature charging characteristics include: ultra-low temperature charging times, ultra-low temperature charging average temperature, and ultra-low temperature charging average current.

[0013] The parking time characteristics include: parking time and the number of long-duration parking sessions;

[0014] The fast charging characteristics include: number of fast charging cycles, average fast charging current, and fast charging temperature rise rate.

[0015] The deep discharge characteristics include: number of deep discharges, average deep discharge current, and SOC decline rate.

[0016] Furthermore, the time-series features include: sliding window statistical features, time-domain features, and frequency-domain features.

[0017] Furthermore, the sliding window statistical features include: the daily average voltage, the daily variance of current, and the daily maximum temperature;

[0018] The time-domain features include: voltage change rate and current fluctuation amplitude;

[0019] The frequency domain characteristics include: voltage main frequency and current harmonic component amplitude.

[0020] Furthermore, the electrical characteristics include: voltage characteristics, current characteristics, internal resistance characteristics, and SOC characteristics.

[0021] Furthermore, the voltage characteristics include: average voltage, voltage variance, and voltage difference;

[0022] The current characteristics include: average current, peak current, and current fluctuation amplitude;

[0023] The internal resistance characteristics include: internal resistance change rate and internal resistance-temperature correlation coefficient;

[0024] The SOC characteristics include: SOC change rate and SOC-voltage correlation coefficient.

[0025] Furthermore, the segmentation features include: segmentation features based on SOC intervals, segmentation features based on temperature intervals, segmentation features based on current intervals, segmentation features based on time intervals, and segmentation features based on operating condition intervals.

[0026] Furthermore, the segmented features based on the SOC interval include: SOC interval voltage features and SOC interval internal resistance features, wherein the SOC interval voltage features are the mean, variance, maximum, minimum, and rate of change of voltage within the SOC ∈ [30%, 90%] interval; and the SOC interval internal resistance features are the mean and rate of change of internal resistance within the SOC ∈ [10%, 20%] interval.

[0027] The segmented characteristics based on temperature ranges include: internal resistance characteristics in the high-temperature range and charging efficiency in the low-temperature range. The internal resistance characteristics in the high-temperature range are the mean, rate of change, and fluctuation amplitude of the internal resistance in the range where the temperature is >40°C. The charging efficiency in the low-temperature range is the hourly growth rate of the SOC in the range where the temperature is <0°C.

[0028] The segmented features based on the current range include: temperature features in the fast charging range and voltage features in the discharge range. The temperature features in the fast charging range are the average temperature, the rate of temperature increase, and the maximum temperature within the range where the current is >2C. The voltage features in the discharge range are the rate of voltage decrease within the range where the current is <-1C (discharge).

[0029] The segmented features based on time intervals include: parking time interval SOC features and long-term static internal resistance features, wherein the parking time interval SOC features are the rate of decrease of SOC calculated in the interval where the parking time is >12 hours; and the long-term static internal resistance features are the rate of change of internal resistance in the interval where the static time is >24 hours.

[0030] The segmented features based on operating condition ranges include: temperature features in the overvoltage charging range and voltage features in the deep discharge range. The temperature features in the overvoltage charging range are the calculated rate of temperature increase and fluctuation amplitude within the range where the voltage is >4.2V. The voltage features in the deep discharge range are the voltage fluctuation amplitude within the range where the SOC is <10%.

[0031] Further, the step of calculating the cross-features of the battery's operating condition characteristics, timing characteristics, electrical characteristics, and segmented characteristics to generate a fused cross-feature matrix includes:

[0032] The operating condition characteristics, timing characteristics, electrical characteristics, and segmentation characteristics are combined into an input feature set;

[0033] The input feature set is processed by a factorization machine. For any two different features in the input feature set, the latent vectors corresponding to the two features are obtained and the inner product operation is performed. Then, the values ​​of the two features are combined to calculate the results of multiple pairs of feature interactions. The results of multiple pairs of feature interactions are combined to form a second-order cross feature vector.

[0034] The input feature set is processed by a deep neural network. First, the various features in the input feature set are converted into a dense feature embedding matrix through an embedding layer. Then, the feature embedding matrix is ​​input into a network structure containing at least two fully connected layers. The ReLU activation function and the Sigmoid activation function are used in sequence for nonlinear transformation to obtain multiple sets of high-order interaction results of multiple features. The results of multiple sets of high-order interaction of multiple features are combined into a high-order cross feature vector.

[0035] The second-order cross feature vector and the higher-order cross feature vector are weighted and fused based on the fusion weight to obtain the fused cross feature vector, wherein the fusion weight is determined by maximizing the accuracy of the battery thermal runaway historical data verification set.

[0036] The fused cross feature vectors are aligned sequentially according to the timestamps of the battery operation data to generate a fused cross feature matrix. The rows of the cross feature matrix correspond to different timestamps, and the columns correspond to different cross features in the fused cross feature vectors. The matrix dimension is determined by the time series length, the dimension of the second-order cross feature vector, and the dimension of the higher-order cross feature vector.

[0037] A second aspect of the present invention 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 cross-feature fusion as described in any of the preceding claims.

[0038] Beneficial effects:

[0039] This invention presents a battery thermal runaway prediction method based on cross-feature fusion. By acquiring four types of features—operating condition, time series, electrical, and segmented features—it characterizes the actual usage scenario, dynamic change trend, core electrical attributes, and specific interval states of the battery, achieving a multi-dimensional and comprehensive description of the battery state and avoiding missed risk assessments due to incomplete features. By calculating the cross-features of the four types of features and generating a fusion matrix, it can automatically discover potential correlations between features. Compared to single-feature prediction, it effectively improves the sensitivity and accuracy of thermal runaway risk identification, providing more reliable data support for early warning. Operating condition features are adaptable to different usage environments (such as fast charging of electric vehicles and long-term static placement of energy storage systems), segmented features can accurately analyze key intervals (such as low temperature and high SOC), and the cross-feature matrix can be flexibly input into various prediction models such as XGBoost and LSTM. Whether for large-scale monitoring of electric vehicle fleets or customized early warning for energy storage power stations, it can meet the needs through flexible feature combinations and model adaptation, solving the problem of poor scenario adaptability of traditional methods. Through a systematic feature extraction and cross-feature generation logic, the entire process from raw data to risk prediction can be completed without human intervention, reducing human error while significantly improving the efficiency of large-scale battery safety management and reducing operation and maintenance costs. Attached Figure Description

[0040] Figure 1 A schematic flowchart illustrating a battery thermal runaway prediction method based on cross-feature fusion according to an embodiment of the invention;

[0041] Figure 2 This is a schematic block diagram of a computer device according to an embodiment of the present invention. The realization of the object, functional features, and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] Those skilled in the art will understand that, unless specifically 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 this specification 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.

[0044] It will be understood by those skilled in the art 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 invention 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.

[0045] Reference Figure 1 This invention provides a battery thermal runaway prediction method based on cross-feature fusion. The execution entity is a cloud-based server or an in-vehicle computer, etc. Taking the cloud as an example, the method includes the following steps:

[0046] S1. Obtain the battery's operating condition characteristics, time-series characteristics, electrical characteristics, and segmented characteristics, wherein the operating condition characteristics reflect the battery's actual usage environment, the time-series characteristics reflect the time-series changes in the battery's state, the electrical characteristics reflect the battery's electrical performance, and the segmented characteristics reflect the characteristic statistics or changes within a specific interval.

[0047] In this step, the aforementioned operating condition characteristics reflect the characteristics of the battery's actual usage environment, such as the number of overvoltage charging cycles and the number of ultra-low temperature charging cycles, which are recorded by the vehicle's BMS and uploaded to the cloud. Time-series characteristics reflect the battery's state changes over time, such as sliding window statistics and voltage change rate, extracted from the time-series data uploaded by the BMS. Electrical characteristics reflect the battery's electrical performance, such as voltage, current, internal resistance, and SOC-related parameters, directly derived from the electrical data collected in real-time by the BMS. Segmented characteristics are statistical features or changes extracted within specific intervals (such as SOC intervals or temperature intervals), also known as "slice features," generated by the cloud through interval analysis of the BMS data. The vehicle's BMS collects real-time data on battery temperature, voltage, current, internal resistance, SOC, and user behavior, such as charging voltage, ambient temperature, and parking status, and uploads it to the cloud via vehicle-to-everything (V2X) or mobile networks. After receiving the data, the cloud first performs outlier processing, format standardization, and numerical normalization (e.g., 3σ principle to remove outliers, Min-Max normalization to [0,1]), and then extracts the above four types of features from the preprocessed multidimensional dataset. For example, a certain brand of electric vehicle's BMS uploaded real-time data from 08:00:00 on August 24, 2025 to the cloud: temperature 25.3°C, single cell voltage 3.72V, charging / discharging current 2.1A, internal resistance 15.2mΩ, SOC 85%, overvoltage charging times 3, ultra-low temperature charging times 1, and parking time 12 hours. After cloud preprocessing, operating condition features (overvoltage charging times 3), time-series features (daily average voltage 3.72V), electrical features (internal resistance change rate 0.3mΩ / hour), and segmented features (average voltage 3.85V in the SOC ∈ [30%, 90%] interval). Uploading data directly through the BMS ensures the real-time nature and accuracy of the raw data; centralized cloud processing and extraction of four types of features cover the entire dimensions of battery "usage environment - time variation - electrical performance - interval state," overcoming the limitations of existing technologies that rely on only a single feature and laying the foundation for subsequent cross-feature generation.

[0048] S2. Calculate the cross features of the battery's operating condition characteristics, timing characteristics, electrical characteristics, and segmentation characteristics, and generate a fused cross feature matrix.

[0049] In this step, the aforementioned cross features are interactive combinations between multidimensional features (such as "SOC range voltage × overvoltage charging times"), which can uncover potential correlations between features; the fused cross feature matrix is ​​structured data that integrates all cross features and is used as input for subsequent risk prediction models.

[0050] The cloud-based system employs preset feature combination methods (such as Deep Factorization Machine (DeepFM), Genetic Algorithm (GA), or manual rules) to perform cross-calculation on four types of features. The DeepFM model is prioritized, with its FM part automatically learning second-order cross-features and its DNN part learning higher-order nonlinear interaction features, eliminating the need for manual design. For example, the cloud-based system inputs the extracted "SOC∈[30%,90%] interval voltage mean 3.85V" (segmented feature) and "overvoltage charging times 3 times" (operating condition feature) into the DeepFM model. The FM part calculates the second-order cross-feature value as 3.85×3÷10=1.155 (normalized to 1.2), while the DNN part combines "internal resistance change rate 0.3mΩ / hour" (electrical feature) to learn higher-order cross-features. Finally, a cross-feature matrix is ​​generated, containing dimensions such as "SOC interval voltage × overvoltage charging (1.2)" and "high-temperature internal resistance × fast charging frequency (0.45)". By using automated methods such as DeepFM to generate cross features, the inefficiency and omissions of manual design are avoided; the fusion matrix integrates multi-dimensional interactive information to capture risk signals that are difficult to detect by traditional methods, such as the synergistic risk of "overvoltage charging + voltage anomaly in a specific SOC range".

[0051] S3. Based on the cross-feature matrix, predict the probability of thermal runaway of the battery.

[0052] In this step, the aforementioned thermal runaway probability is the likelihood (value 0-1) of the battery experiencing thermal runaway within a future period, calculated by a cloud-based prediction model. This model is trained on historical thermal runaway label data (e.g., XGBoost, LSTM), and can further optimize the results using a rule engine and knowledge base. The cloud inputs the cross-feature matrix into a pre-set prediction model (XGBoost preferred), which outputs an initial thermal runaway probability. This probability is then further validated using a rule engine (e.g., "overvoltage charging × voltage fluctuation > 0.5 triggers a warning") and a knowledge base (e.g., typical cases of thermal runaway caused by overvoltage charging) to determine the final probability value. For example, the cloud inputs the cross-feature matrix into a trained XGBoost model, outputting an initial probability of 0.42. The rule engine validation finds "3 overvoltage charging cycles × voltage fluctuation 0.2 = 0.6 > 0.5," and the knowledge base matches cases of "overvoltage charging associated with thermal runaway," ultimately correcting the probability to 0.45. The predictive model combines a rule engine and a knowledge base, balancing the accuracy of data-driven approaches with the reliability of expert experience. It calculates and outputs probabilities in real time in the cloud, providing a quantitative basis for subsequent risk level classification and solving the problem of insufficient prediction accuracy in existing technologies.

[0053] This embodiment uses "predicting thermal runaway of an electric vehicle battery" as its application scenario. The execution entity is the cloud, and the data source is multi-dimensional data uploaded in real-time by the vehicle's Battery Management System (BMS). Through a three-step process of "full-dimensional feature extraction - automated cross-feature generation - multi-mechanism probability prediction," accurate assessment of thermal runaway risk is achieved. Compared with existing technologies: direct upload from the BMS ensures real-time data and avoids the latency of distributed processing; four types of features plus automatic cross-feature generation cover the full range of battery states and uncover potential risk correlations; and the collaboration of model, rules, and knowledge base improves prediction accuracy. The overall solution is suitable for large-scale electric vehicle fleet management, providing a feasible technical path for cloud-based battery safety monitoring.

[0054] In one embodiment, the above-mentioned operating condition characteristics include: overvoltage charging characteristics, ultra-low temperature charging characteristics, downtime characteristics, fast charging characteristics, and deep discharge characteristics.

[0055] The aforementioned overvoltage charging characteristics refer to the relevant statistics regarding battery charging voltage exceeding the safety threshold (typically 4.2V), including: overvoltage charging frequency, cumulative overvoltage charging time, and average overvoltage amplitude. Overvoltage charging frequency refers to the number of times the charging voltage exceeds 4.2V; cumulative overvoltage charging time refers to the total duration of all overvoltage charging processes; and average overvoltage amplitude refers to the average value of "actual voltage - 4.2V" during each overvoltage charge. The BMS monitors the charging voltage in real time. When the voltage exceeds 4.2V, it records the start / end time and real-time voltage of that charge and uploads the data to the cloud daily. The cloud system counts the number of charges, sums the duration of each charge to obtain the cumulative time, and calculates the average amplitude for each charge. For example, the BMS uploads data from three overvoltage charging events for a battery: the first time (voltage 4.5V, duration 10min, amplitude 0.3V), the second time (4.45V, 15min, 0.25V), and the third time (4.48V, 12min, 0.28V). Cloud calculations show: number of charging events = 3, cumulative time = 10 + 15 + 12 = 37min, average amplitude = (0.3 + 0.25 + 0.28) / 3 ≈ 0.28V. Refining the three-dimensional characteristics of overvoltage charging—frequency, duration, and intensity—reflects risk more effectively than a single "number of charging events" (e.g., the risk difference between multiple short-duration overvoltage events and a single long-duration overvoltage event), providing more accurate input for subsequent cross-feature analysis.

[0056] The aforementioned cryogenic charging characteristics refer to the relevant statistics of charging when the ambient temperature is below a preset temperature threshold (usually 0°C), including the number of cryogenic charging cycles, the average temperature during cryogenic charging, and the average current during cryogenic charging. The number of cryogenic charging cycles refers to the number of times charging occurs when the ambient temperature is <0°C; the average temperature during cryogenic charging refers to the average ambient temperature during all cryogenic charging cycles; and the average current during cryogenic charging refers to the average charging current during all cryogenic charging cycles. The BMS synchronously collects the ambient temperature and charging current during charging. When the temperature is <0°C, it records the data and uploads it to the cloud. The cloud then counts the number of cycles and calculates the average temperature and current. For example, if the BMS uploads two cryogenic charging data entries: the first (temperature -5°C, current 2A) and the second (-3°C, 2.5A), the cloud calculates: number of cycles = 2, average temperature = (-5-3) / 2 = -4°C, average current = (2+2.5) / 2 = 2.25A. By combining "temperature intensity" and "current magnitude", the damage to the battery caused by low-temperature charging can be quantified (e.g., the risk of -5°C +2.5A is higher than that of -3°C +2A), thus overcoming the one-sidedness of existing technologies that only count the number of times.

[0057] The aforementioned parking time characteristic refers to the statistical analysis of the time the battery is in a static state (e.g., current <0.1A), including parking time and the number of long-term parking instances. Parking time refers to the cumulative duration of the battery in a static state (current <0.1A); the number of long-term parking instances refers to the number of times the static duration exceeds 12 hours. The BMS monitors the current status, and when the current is <0.1A, it determines the battery to be parked, records the start / end time of the static period, and uploads it to the cloud. The cloud accumulates the static duration and counts the number of times it exceeds 12 hours. For example, the BMS uploads parking data for a certain day: 9:00-18:00 (9 hours, current 0.05A), 20:00-10:00 the next day (14 hours, current 0.08A); the cloud calculates: parking time = 9 + 14 = 23 hours, number of long-term parking instances = 1 (14 hours > 12 hours). Distinguishing between "normal parking" and "long-term parking," long-term parking may lead to abnormal battery self-discharge or increased internal resistance; this characteristic provides a basis for identifying "risks after static parking."

[0058] The aforementioned fast charging characteristics refer to the relevant statistics for charging current > 2C (C being the battery capacity ratio), including: number of fast charging cycles, average fast charging current, and fast charging temperature rise rate. The number of fast charging cycles refers to the number of times the charging current is > 2C; the average fast charging current is the average current across all fast charging processes; and the fast charging temperature rise rate is the average of the "temperature change / time" during fast charging (°C / minute). The BMS monitors the charging current, and when the current is > 2C, it is determined to be fast charging. Real-time current and temperature changes are recorded and uploaded to the cloud; the cloud then counts the number of cycles and calculates the average current and temperature rise rate. For example, the BMS uploads fast charging data three times: the first time (current 3C, temperature rises from 25°C to 30°C within 10 minutes, rate 0.5°C / min), the second time (2.5C, rises to 32°C in 15 minutes, rate 0.47°C / min), and the third time (3.2C, rises to 33°C in 8 minutes, rate 1.0°C / min). Cloud calculations show: number of times = 3, average current = (3 + 2.5 + 3.2) / 3 ≈ 2.9C, average temperature rise rate = (0.5 + 0.47 + 1.0) / 3 ≈ 0.66°C / min (normalized to 0.7°C / min). Combining "current intensity" and "temperature change" quantifies the thermal impact of fast charging on the battery (e.g., high current + high temperature rise rate carries a higher risk), providing key characteristics for identifying "fast charging-induced thermal runaway."

[0059] The aforementioned deep discharge characteristics refer to the statistical quantities related to discharge when the State of Charge (SOC) is below 10%, including: number of deep discharges, average deep discharge current, and SOC decay rate. The number of deep discharges refers to the number of discharges when the SOC is <10%; the average deep discharge current is the average current across all deep discharge processes; and the SOC decay rate is the average of the "SOC change / time" (% / hour) during deep discharges. The BMS monitors the SOC and discharge current, records data and uploads it to the cloud when the SOC is <10%; the cloud then counts the number of discharges and calculates the average current and SOC decay rate. For example, the BMS uploads data from two deep discharge cycles: the first (SOC from 10% to 5%, duration 2 hours, current 1.5A, rate 2.5% / h), and the second (8% to 3%, duration 1.5 hours, current 1.8A, rate 3.33% / h). Cloud calculations show: cycles = 2, average current = (1.5 + 1.8) / 2 = 1.65A, average SOC decrease rate = (2.5 + 3.33) / 2 ≈ 2.92% / h. This refined "frequency-current-rate" analysis of deep discharge reflects the degree of damage caused by over-discharge (e.g., high-rate deep discharge accelerates electrode aging), providing a basis for identifying "post-deep discharge risks."

[0060] The vehicle's BMS records real-time usage data such as battery charging voltage, ambient temperature, current status, and SOC changes, uploading it to the cloud at preset intervals (e.g., hourly). The cloud then filters and analyzes the uploaded data, extracting the five operating condition characteristics mentioned above. For example, an electric vehicle's BMS uploads daily data for August 24, 2025, to the cloud: charging voltage exceeding 4.2V 3 times, charging at ambient temperature <0°C 1 time, cumulative time of current <0.1A 12 hours (including 1 parking session >12 hours), charging current >2C 5 times, and discharging with SOC <10% 2 times. Based on this data, the cloud extracts overvoltage charging characteristics (3 times), ultra-low temperature charging characteristics (1 time), parking time characteristics (12 hours, including 1 long parking session), fast charging characteristics (5 times), and deep discharge characteristics (2 times).

[0061] Five types of operating condition characteristics directly reflect the actual usage environment and aging causes of the battery (such as overvoltage charging accelerating lithium dendrite growth), solving the problem that existing technologies ignore the impact of operating conditions; features are extracted from behavioral data uploaded by BMS in the cloud, ensuring that the features are strongly correlated with actual usage scenarios and improving the targeting of subsequent predictions.

[0062] In this embodiment, the specific types of operating condition characteristics are clearly defined, focusing on key behaviors throughout the entire battery lifecycle of "charging-stopping-discharging". During execution, the BMS is responsible for recording and uploading usage behavior data, while the cloud is responsible for feature extraction. The two work together to ensure the authenticity and completeness of the operating condition characteristics. Compared with existing technologies, this approach covers five core operating conditions: "overpressure, low temperature, static storage, fast charging, and deep discharging," comprehensively reflecting the battery's usage stress. Feature extraction relies on raw BMS behavior data, avoiding errors from manual annotation. It provides key inputs for subsequent cross-feature generation, including the "usage environment dimension" (e.g., "ultra-low temperature charging characteristics × fast charging characteristics" can reflect the dual risks of low-temperature fast charging). The overall solution makes thermal runaway prediction closer to the actual battery usage scenario, providing technical support for accurately identifying "risks caused by improper use". Furthermore, each operating condition characteristic includes three sub-parameters: "frequency, intensity, and dynamic change," comprehensively depicting the impact of the operating condition on the battery. This provides "fine-grained inputs" (e.g., "fast charging temperature rise rate × ultra-low temperature charging times") for subsequent DeepFM cross-feature generation, improving the risk differentiation of cross-features. The overall solution upgrades the operating condition characteristics from "qualitative description" to "quantitative analysis," providing data support for high-precision prediction of thermal runaway.

[0063] In one embodiment, the aforementioned time-series features include: sliding window statistical features, time-domain features, and frequency-domain features.

[0064] Sliding window statistical features refer to the characteristics obtained by statistically analyzing battery characteristics (voltage, current, temperature) within a fixed time window (such as 1 day or 1 hour), including the mean, maximum, minimum, and variance. These include the daily average voltage, daily variance of current, and daily maximum temperature. The daily average voltage is the arithmetic mean of the battery voltage over 24 hours, reflecting the daytime voltage stability. The daily variance of current is the variance of the current data over 24 hours, reflecting the degree of daytime current fluctuation. The daily maximum temperature is the highest battery temperature over 24 hours, reflecting the highest daytime heat load. The BMS uploads battery voltage, current, and temperature data (24 data points per day) to the cloud hourly. The cloud performs statistical calculations on these 24 data points: summing and averaging the voltage (daily average voltage), calculating the variance of current (daily variance of current), and taking the maximum temperature (daily maximum temperature). For example, the BMS uploaded hourly data for a battery on August 24, 2025: There were 24 voltage data points (3.7V, 3.72V, 3.68V, ..., 3.71V), which summed to 89.04V. The mean was 89.04 / 24 = 3.71V. The variance of the current data points (2.1A, 2.2A, 2.0A, ..., 2.1A) was calculated to be 0.04. The maximum value of the temperature data (25°C, 26°C, 28°C, ..., 27°C) is 28°C. The three statistical features correspond to "voltage stability, current fluctuation, and temperature limits," directly related to the battery's core safety indicators (e.g., a daily maximum temperature exceeding 40°C warrants attention for thermal runaway). Compared to multi-dimensional statistics, focusing on key indicators reduces computational complexity while ensuring no core risks are overlooked. A sliding window smooths out short-term fluctuations, reflecting the battery's daily stable state (e.g., whether the average voltage is within the normal range of 3.6-3.8V). Compared to real-time single-point data, statistical features better reflect long-term trends, avoiding misjudgments caused by instantaneous noise.

[0065] The aforementioned time-domain characteristics refer to the direct analysis of the dynamic changes in battery characteristics (such as rate of change and fluctuation amplitude) from a time dimension, reflecting the instantaneous trend of characteristic changes over time, including voltage change rate and current fluctuation amplitude. Voltage change rate refers to the amount of voltage change per unit time (usually measured in V / s or V / min), reflecting the dynamic speed of voltage change; current fluctuation amplitude refers to the standard deviation of current data, reflecting the stability of current over a short period. The BMS uploads voltage and current data to the cloud in 10-second increments; the cloud selects a short time window of 5 minutes (30 data points in total), calculates the voltage change rate (ΔV / 10s) of adjacent data points and takes the average, and calculates the standard deviation (fluctuation amplitude) of the current data. For example, the cloud selects 10-second data points for a battery from 14:00 to 14:05: voltage data (3.72V, 3.73V, 3.75V, ..., 3.74V) totaling 30 values, with the mean of adjacent differences being 0.002V / 10s = 0.0002V / s; current data (2.1A, 2.12A, 2.08A, ..., 2.1A) has a standard deviation of 0.02A. Voltage change rate can detect early overvoltage during charging or internal short circuits (such as voltage spikes / drops), while current fluctuation amplitude can identify abnormalities in the charging / discharging circuit (such as current fluctuations caused by loose contacts). These two features focus on "voltage dynamics - current stability," providing a quantitative basis for early warning of electrical faults. Time-domain features capture instantaneous changes in battery characteristics (such as voltage spikes / drops), enabling early detection of internal battery abnormalities (such as voltage spikes caused by separator damage). Compared to static statistics, dynamic change features are better able to reflect the "warning signals" before thermal runaway.

[0066] The aforementioned frequency domain characteristics refer to converting time-domain signals such as voltage and current to the frequency domain through Fourier transform, extracting features such as the dominant frequency and harmonic components to reflect the frequency distribution pattern of the signal, including the dominant voltage frequency and the amplitude of the current harmonic components. The dominant voltage frequency refers to the frequency component with the highest energy proportion after the voltage signal undergoes Fourier transform, reflecting the main periodicity of the voltage signal. The amplitude of the current harmonic components refers to the signal amplitude of frequencies that are integer multiples of the dominant frequency (such as the second and third harmonics) in the current signal, reflecting the degree of nonlinear distortion of the current (the larger the harmonics, the greater the electrical loss). The BMS uploads one hour of voltage and current data (3600 points in total) to the cloud at the second level; the cloud performs a Fast Fourier Transform (FFT) on the data to generate a frequency domain spectrum; the frequency with the highest energy is extracted as the dominant voltage frequency, and the amplitude of the second harmonic is extracted as the amplitude of the current harmonic components. For example, when performing a second-level FFT on the voltage data of a battery over one hour in the cloud, the energy proportion of the 50Hz frequency in the frequency domain spectrum reaches 90% (dominant frequency = 50Hz); in the FFT result of the current signal, the amplitude of the second harmonic (100Hz) at the dominant frequency of 50Hz is 0.015A (0.02A after normalization). Abnormal voltage dominant frequency (such as deviation from 50Hz±2Hz) may indicate grid interference or internal electrical faults, while current harmonic component amplitude exceeding the threshold (such as 0.03A) may lead to increased battery heating; these two features supplement electrical safety assessment from the frequency dimension and solve periodic problems that are difficult to detect in time domain analysis.

[0067] This embodiment clarifies three specific forms of time-series features, covering a full perspective across the time dimensions of "statistical trends, dynamic changes, and frequency distribution." During execution, the BMS provides high-frequency raw data (minute / second level), and the cloud extracts features through a three-level processing approach: "sliding window statistics, time domain analysis, and frequency domain transformation." Compared with existing technologies, the sliding window reflects long-term trends, the time domain reflects instantaneous changes, and the frequency domain reflects latent periodic anomalies. Based on high-frequency data from the BMS, the timeliness and accuracy of time-series features are ensured. It also provides a "time dimension" input (such as "voltage change rate × overvoltage charging times") for subsequent cross-feature generation, improving the ability to identify "dynamic risks." The overall solution upgrades time-series features from a "single time perspective" to "multi-dimensional time analysis," providing richer time-related information for thermal runaway prediction. Furthermore, each time-series feature category retains only 2-3 core parameters, reducing the cloud computing load and making it suitable for large-scale fleet real-time processing. Parameter selection is based on battery safety mechanisms (e.g., daily maximum temperature correlates with thermal load, harmonic amplitude correlates with electrical losses), ensuring a strong correlation between features and thermal runaway risk. This provides "key inputs in the time dimension" (e.g., "voltage change rate × overvoltage charging cycles") for subsequent cross-features, improving the accuracy of risk identification. The overall solution makes time-series feature extraction more targeted and operable, providing an efficient technical path for real-time cloud processing.

[0068] In one embodiment, the aforementioned electrical characteristics include: voltage characteristics, current characteristics, internal resistance characteristics, and SOC characteristics.

[0069] The aforementioned voltage characteristics, reflecting parameters of the battery voltage state (such as the average voltage, variance, and voltage difference of individual cells), are directly related to the battery's charging state and individual cell consistency. Specifically, they include the average voltage, voltage variance, and voltage difference. The average voltage refers to the arithmetic mean of the voltages of all individual cells in the battery pack, reflecting the overall voltage level; the voltage variance refers to the variance of the individual cell voltage data, reflecting the dispersion of individual cell voltages; and the voltage difference is the difference between the maximum and minimum values ​​of an individual cell voltage, directly reflecting individual cell consistency. The BMS collects the voltages of all individual cells in the battery pack (e.g., 100 voltage values ​​from 100 battery cells) and uploads them to the cloud every 5 minutes. The cloud performs statistical calculations on the individual cell voltage data to obtain the three voltage characteristics. For example, in a 100-cell battery pack, the BMS uploads individual cell voltage data of 3.70V, 3.72V, 3.71V, ..., 3.73V (100 cells in total); the cloud calculation shows: average value = (3.70 + 3.72 + ... + 3.73) / 100 = 3.715V ≈ 3.72V; variance = 0.001 The maximum value is 3.73V, the minimum value is 3.70V, and the voltage difference is 0.03V. These three features assess the voltage status from three dimensions: "overall, discrete, and differential." The average value reflects whether charging is normal (e.g., an average value > 3.8V indicates overcharging), the variance reflects the stability of the individual cell voltage, and the voltage difference reflects consistency (e.g., a difference > 0.1V requires balancing). Compared to the overall voltage, individual cell-level features are better able to detect local anomalies (e.g., a sudden increase in the voltage of a single cell indicates an internal short circuit).

[0070] The aforementioned current characteristics refer to parameters reflecting the state of battery charging and discharging current (such as mean, peak, and fluctuation amplitude), which are related to charging and discharging intensity and stability. Specifically, they include the average current, peak current, and current fluctuation amplitude. The average current refers to the arithmetic mean of the charging and discharging current over a period of time, reflecting the average charging and discharging intensity; the peak current refers to the maximum value of the charging and discharging current over a period of time, reflecting the instantaneous maximum load; and the current fluctuation amplitude refers to the standard deviation of the current data, reflecting current stability. The BMS uploads charging and discharging current data to the cloud at 1-second intervals; the cloud selects a 1-hour window (3600 data points) to calculate the average, peak, and fluctuation amplitude of the current. For example, the cloud analyzes 1 hour of current data for a certain battery: current range 1.9A-3.2A, average value = (1.9+2.1+…+3.2) / 3600 = 2.15A; peak value = 3.2A; standard deviation of current data = 0.15A (fluctuation amplitude). The average value reflects the charging and discharging intensity (e.g., average value > 2C indicates fast charging), the peak value reflects the instantaneous load (e.g., peak value > 3C requires limiting), and the fluctuation range reflects the current stability (e.g., > 0.2A indicates circuit abnormality). These three features cover "average-instantaneous-stable" to comprehensively assess the thermal shock and electrical damage to the battery caused by charging and discharging.

[0071] The aforementioned internal resistance characteristics refer to parameters reflecting the internal resistance state of the battery. Internal resistance growth is a significant precursor to battery aging and thermal runaway, specifically including the rate of change of internal resistance and the internal resistance-temperature correlation coefficient. The rate of change of internal resistance refers to the amount of change in internal resistance per unit time, reflecting the rate of increase in internal resistance. The internal resistance-temperature correlation coefficient is the Pearson correlation coefficient between internal resistance and temperature. In a normal battery, internal resistance decreases with increasing temperature (correlation coefficient is negative), but in abnormal situations (such as lithium dendrite growth), it may become positively correlated. The BMS measures the internal resistance (which can be calculated from current and voltage data in the cloud) and temperature every hour and uploads the data to the cloud. The cloud calculates the rate of change of internal resistance between two adjacent measurements (e.g., 15.2 mΩ at time t1, 15.5 mΩ at time t2, Δt = 1 h, rate of change = 0.3 mΩ / h). Twenty-four sets of internal resistance-temperature data from one day are selected to calculate the correlation coefficient. For example, the BMS uploads one day's internal resistance-temperature data for a battery: internal resistance 15.2mΩ (25°C), 15.1mΩ (26°C), ..., 15.0mΩ (28°C); cloud calculation shows: internal resistance change rate over 24 hours = (15.0-15.2) / 24 ≈ -0.008mΩ / h (normal degradation); internal resistance-temperature correlation coefficient = -0.8 (negative correlation, normal). The internal resistance change rate can quantify the aging speed (e.g., >0.3mΩ / h indicates accelerated aging), and the correlation coefficient can detect internal anomalies (e.g., a positive correlation coefficient indicates lithium dendrites); these two features are directly related to the battery's health status, providing core indicators for early warning of thermal runaway.

[0072] The aforementioned SOC characteristics refer to parameters reflecting the battery's State of Charge, which are related to the battery's remaining capacity and charge / discharge efficiency. Specifically, they include the SOC change rate and the SOC-voltage correlation coefficient. The SOC change rate refers to the change in SOC per unit time (formula: ΔSOC / Δt, unit % / hour), reflecting the charging / discharging speed. The SOC-voltage correlation coefficient is the Pearson correlation coefficient between SOC and voltage. In a normal battery, SOC and voltage are strongly positively correlated (correlation coefficient > 0.9), and the correlation decreases with capacity decay. The BMS uploads SOC and voltage data to the cloud every 5 minutes. The cloud calculates the SOC change rate within one hour (e.g., if SOC drops from 85% to 80%, Δt = 1h, change rate = -5% / h). Twenty-four sets of SOC-voltage data from one day are selected to calculate the correlation coefficient. For example, the BMS uploads one day's SOC-voltage data for a battery: SOC 85% (3.72V), 80% (3.70V), ..., 30% (3.55V); cloud calculation shows: SOC change rate during discharge = -5% / h, during charging = 10% / h; SOC-voltage correlation coefficient = 0.92 (strong positive correlation, normal). The SOC change rate can assess the charge / discharge speed (e.g., discharge rate > 10% / h indicates high current discharge), and the correlation coefficient can detect capacity decay (e.g., < 0.8 indicates capacity loss > 20%); these two features are linked to the battery's "state of charge - parameter consistency," providing a basis for charge / discharge safety and capacity management.

[0073] This embodiment defines four core types of electrical features, covering the four major electrical dimensions of the battery: voltage, current, internal resistance, and state of charge (SOC). During execution, the Battery Management System (BMS) collects and uploads various electrical data in real time. Features are extracted in the cloud using a "statistical-correlation" logic, ensuring a strong correlation between electrical features and battery safety and health status. Compared to existing technologies: the four types of electrical features comprehensively cover battery electrical performance, avoiding the limitations of single electrical parameters; each feature type includes a "dynamic change-parameter correlation" sub-dimension (e.g., internal resistance features include change rate and temperature correlation coefficient), reflecting battery status better than static parameters; based on high-precision electrical data from the BMS, the accuracy of the features is ensured, providing core input for subsequent cross-feature generation (e.g., "internal resistance change rate × high-temperature range internal resistance feature"). The overall solution makes electrical features the "core data pillar" for thermal runaway prediction, providing technical support for accurately identifying thermal runaway risks induced by electrical anomalies. Furthermore, each type includes 2-3 core parameters. During execution, the BMS provides unit-level and high-frequency electrical data, and the cloud calculates parameters according to a "statistical-correlation" logic to ensure the refinement and quantification of electrical characteristics. Each electrical parameter is based on a clear calculation formula (e.g., internal resistance change rate = ΔR / Δt), avoiding subjective judgment and ensuring feature reproducibility. Parameter selection focuses on the core dimensions of "safety-health-performance" (e.g., voltage difference correlates with consistency safety, internal resistance change rate correlates with health status). This provides "fine-grained electrical input" (e.g., "internal resistance change rate × high-temperature range internal resistance feature") for subsequent DeepFM cross-feature generation, improving the accuracy of risk identification. The overall solution upgrades electrical characteristics from "qualitative description" to "quantitative parameters," providing core data support for high-precision thermal runaway prediction, while meeting the high-efficiency requirements of large-scale cloud processing.

[0074] In one embodiment, the above-mentioned segmentation features include: segmentation features based on SOC intervals, segmentation features based on temperature intervals, segmentation features based on current intervals, segmentation features based on time intervals, and segmentation features based on operating condition intervals.

[0075] The aforementioned segmented features based on SOC intervals refer to the statistical features such as voltage and internal resistance extracted within a specific SOC interval, reflecting the state differences of the battery in different charge intervals. Specifically, these include SOC interval voltage features and SOC interval internal resistance features.

[0076] The voltage characteristics of the SOC interval are the mean, variance, maximum, minimum, and rate of change of the voltage within the SOC ∈ [30%, 90%] interval; the internal resistance characteristics of the SOC interval are the mean and rate of change of the internal resistance within the SOC ∈ [10%, 20%] interval.

[0077] Within the SOC ∈ [30%, 90%] (commonly used range), the formulas for the mean (μ_V), variance (σ_V), maximum (max_V), minimum (min_V), and rate of change (ΔV) of the voltage are as follows:

[0078] -μ_V=mean(voltage|SOC∈[30%,90%])

[0079] -σ_V=std(voltage|SOC∈[30%,90%])

[0080] -ΔV = max(voltage) - min(voltage) | SOC ∈ [30%, 90%].

[0081] The BMS uploads SOC and voltage data, and the cloud filters voltage data with SOC ∈ [30%, 90%], calculating the five parameters mentioned above. For example, the cloud filters voltage data (100 points) for a certain battery with SOC ∈ [30%, 90%]: mean μ_V = 3.85V, variance σ_V = 0.01. The maximum value max_V = 3.90V, the minimum value min_V = 3.80V, and the rate of change ΔV = 0.10V; generating SOC range voltage characteristics: 3.85V, 0.01V. 3.90V, 3.80V, 0.10V. SOC∈[30%,90%] is the high-frequency usage range of the battery. Abnormal voltage characteristics (such as average value >4.0V) directly indicate the risk of overcharging. The five parameters comprehensively characterize the stability of the voltage in the commonly used range, providing a basis for identifying "voltage anomalies under common operating conditions".

[0082] Within the SOC ∈ [10%, 20%] (low charge range), the mean (μ_R) and rate of change (ΔR) of the internal resistance are given by the following formulas:

[0083] -μ_R=mean(internal resistance|SOC∈[10%,20%])

[0084] -ΔR=(R_t-R_0) / R_0|SOC∈[10%,20%] (R_t is the internal resistance at time t, and R_0 is the initial internal resistance)

[0085] The BMS uploads SOC and internal resistance data, and the cloud filters internal resistance data for SOC ∈ [10%, 20%], calculating the two parameters mentioned above. For example, the cloud filters internal resistance data (50 points) for a certain battery with SOC ∈ [10%, 20%): mean μ_R = 15.8mΩ, initial internal resistance R_0 = 15.5mΩ, time t R_t = 15.7mΩ, rate of change ΔR = (15.7-15.5) / 15.5 ≈ 1.29%; generating internal resistance characteristics for the SOC range: 15.8mΩ, 1.29%. Internal resistance is prone to increase in the low charge range, and abnormal characteristics (such as rate of change > 2%) indicate accelerated battery aging; the two parameters focus on "mean - change", quantifying the internal resistance risk in the low charge range and supplementing the deficiencies of commonly used range characteristics.

[0086] The characteristics of batteries differ significantly within different SOC ranges (e.g., higher internal resistance in the low SOC range), and segmented features can capture anomalies within the range; compared with statistics across the entire SOC range, range features are better able to detect local anomalies and improve the targeting of risk identification.

[0087] The aforementioned segmented features based on temperature ranges refer to features such as internal resistance and charging efficiency extracted within a specific temperature range, reflecting the state of the battery under extreme temperatures. Specifically, these include internal resistance features in the high-temperature range and charging efficiency in the low-temperature range. The internal resistance features in the high-temperature range are the mean, rate of change, and fluctuation amplitude of the internal resistance within the temperature range >40°C. The charging efficiency in the low-temperature range is the hourly rate of increase of the State of Charge (SOC) within the temperature range <0°C.

[0088] Within the temperature range >40°C (high temperature range), the mean (μ_R_temp), rate of change (ΔR_temp), and fluctuation amplitude (σ_R_temp) of the internal resistance are given by the following formulas:

[0089] -μ_R_temp=mean(internal resistance|temperature>40°C)

[0090] -ΔR_temp=(R_t-R_{t-1}) / R_{t-1}|Temperature>40°C

[0091] -σ_R_temp=std(internal resistance|temperature>40°C)

[0092] The BMS uploads temperature and internal resistance data, and the cloud filters internal resistance data with temperatures >40°C, calculating the three parameters mentioned above. For example, the cloud filters internal resistance data (30 points) for a certain battery with temperatures >40°C: mean μ_R_temp = 16.2mΩ, rate of change ΔR_temp = 0.5% / h, fluctuation amplitude σ_R_temp = 0.2mΩ; generating high-temperature range internal resistance characteristics: 16.2mΩ, 0.5% / h, 0.2mΩ. High temperatures accelerate battery side reactions, and abnormal internal resistance characteristics (such as a rate of change >0.3% / h) indicate a risk of thermal runaway. These three parameters quantify the impact of high temperatures on internal resistance, providing a basis for identifying "high-temperature induced internal resistance anomalies".

[0093] Within the temperature range of <0°C (low temperature range), the hourly growth rate of SOC (η_SOC) is calculated using the following formula:

[0094] -η_SOC=(SOC_t-SOC_{t-1}) / Δt|Temperature<0°C (Δt is the time difference in hours)

[0095] The BMS uploads temperature and SOC data, and the cloud filters charging SOC data for temperatures below 0°C to calculate η_SOC. For example, the cloud filters charging data for a battery at temperatures below 0°C: SOC increases from 20% to 30%, Δt = 2h, η_SOC = (30-20) / 2 = 5% / h; generating a low-temperature range charging efficiency of 5% / h. Low temperatures reduce charging efficiency, and η_SOC < 3% / h indicates charging abnormalities (such as lithium plating); this parameter quantifies low-temperature charging performance and provides a basis for identifying "low-temperature charging risks".

[0096] Extreme temperatures are a major cause of thermal runaway. Segmented features can quantify the impact of temperature on the battery (e.g., a change in internal resistance at high temperatures >0.3 mΩ / h indicates a risk). Compared to statistics across the entire temperature range, extreme range features can better focus on high-risk scenarios and improve the timeliness of early warnings.

[0097] The aforementioned segmented features based on current range refer to the temperature, voltage, and other features extracted within a specific current range, reflecting the state of the battery under different charge and discharge intensities. These features include: temperature features in the fast charging range and voltage features in the discharge range. Specifically, the temperature features in the fast charging range are the average temperature, the rate of increase, and the maximum temperature within the range where the current is >2C. The voltage features in the discharge range are the rate of decrease of voltage within the range where the current is <-1C (discharge).

[0098] Within the current > 2C (fast charging range), the average temperature (μ_T_fast), the rate of temperature increase (ΔT_fast), and the maximum temperature (max_T_fast) are calculated using the following formulas:

[0099] -μ_T_fast=mean(temperature|current>2C)

[0100] -ΔT_fast=(T_t-T_{t-1}) / Δt|current>2C

[0101] -max_T_fast=max(temperature|current>2C)

[0102] The BMS uploads current and temperature data, and the cloud filters temperature data with current > 2C, calculating the three parameters mentioned above. For example, the cloud filters temperature data (40 points) for a battery with current > 2C: mean μ_T_fast = 32°C, rate of increase ΔT_fast = 0.8°C / min, maximum value max_T_fast = 35°C; generating fast charging range temperature characteristics: 32°C, 0.8°C / min, 35°C. Fast charging easily leads to a sudden temperature rise, and ΔT_fast > 1°C / min indicates overheating risk; these three parameters quantify the impact of fast charging on temperature, providing a basis for identifying "fast charging-induced thermal risks".

[0103] Within the current < -1C (high current discharge range), the voltage drop rate (ΔV_discharge) is given by the following formula:

[0104] -ΔV_discharge=(V_t-V_{t-1}) / Δt|current<-1C

[0105] The BMS uploads current and voltage data, and the cloud filters voltage data with current <-1C, calculating ΔV_discharge. For example, the cloud filters discharge data for a battery with current <-1C: the voltage drops from 3.72V to 3.70V, Δt=10min, ΔV_discharge=(3.70-3.72) / 10=-0.002V / min (the negative sign indicates discharge); generating the discharge range voltage characteristic: 0.002V / min (absolute value). High current discharge easily leads to a sudden voltage drop, ΔV_discharge>0.003V / min indicates abnormal discharge (such as electrode polarization); this parameter quantifies the stability of the discharge voltage, providing a basis for identifying "high current discharge risk".

[0106] High-current charging and discharging can easily cause the battery to heat up. Segmented characteristics can quantify the thermal impact of current intensity on the battery (e.g., a temperature rise rate of >1°C / min during fast charging indicates overheating). Compared with statistics across the entire current range, high-intensity current range characteristics are better able to detect risks induced by charging and discharging.

[0107] The aforementioned segmented features based on time intervals refer to features such as SOC and internal resistance extracted within a specific time interval, reflecting the state changes of the battery after a long period of rest; including: SOC features for parking time intervals and internal resistance features for long-term rest. The SOC features for parking time intervals are the rate of SOC decrease calculated within a parking time of >12 hours; the internal resistance features for long-term rest are the rate of change of internal resistance within a rest time of >24 hours.

[0108] For parking times > 12 hours (long-term parking), the rate of decrease in SOC (ΔSOC_park) is calculated using the following formula:

[0109] ΔSOC_park = (Initial SOC - Current SOC) / Parking Time | Parking > 12h

[0110] The BMS uploads parking status and SOC data, and the cloud filters data that has been parked for more than 12 hours, calculating ΔSOC_park. For example, if the cloud filters data for a battery that has been parked for more than 12 hours: SOC before parking = 80%, after parking = 75%, duration = 15 hours, ΔSOC_park = (80-75) / 15 ≈ 0.33% / h; generating the parking time interval SOC characteristic: 0.33% / h. Abnormal self-discharge during prolonged parking (ΔSOC_park > 0.5% / h) indicates an internal short circuit; this parameter quantifies the self-discharge rate, providing a basis for identifying "hidden risks after idling".

[0111] The rate of change of internal resistance (ΔR_standby) over a settling time > 24 hours (long-term settling) is given by the following formula:

[0112] ΔR_standby=(R_t-R_0) / R_0|Stationary > 24h

[0113] The BMS uploads data on the battery's resting state and internal resistance. The cloud then filters data that has been resting for more than 24 hours and calculates ΔR_standby. For example, the cloud filters data for a battery that has been resting for more than 24 hours: initial internal resistance R_0 = 15.2 mΩ, time t R_t = 15.6 mΩ, duration = 30 hours, ΔR_standby = (15.6 - 15.2) / 15.2 ≈ 2.63%; generating a long-term resting internal resistance characteristic: 2.63%. Abnormal internal resistance growth after long-term resting (ΔR_standby > 3%) indicates battery aging; this parameter quantifies the change in internal resistance after resting, supplementing the insufficient characteristics under dynamic operating conditions.

[0114] After a long period of inactivity, the self-discharge or changes in internal resistance of the battery are more likely to expose potential problems (such as a SOC drop rate of >0.5% / h when the battery is parked, indicating abnormal self-discharge); segmented features can capture the "hidden risks" after inactivity, supplementing the lack of features under dynamic operating conditions.

[0115] The aforementioned segmented features based on operating condition ranges refer to the temperature, voltage, and other features extracted within specific abnormal operating condition ranges, reflecting the battery's state under abnormal operating conditions. These features include: temperature features in the overvoltage charging range and voltage features in the deep discharge range. Specifically, the temperature features in the overvoltage charging range are the calculated rate of temperature increase and fluctuation amplitude within the range where the voltage is >4.2V; and the voltage features in the deep discharge range are the voltage fluctuation amplitude within the range where the SOC is <10%.

[0116] Within the voltage > 4.2V (overvoltage charging range), the rate of temperature rise (ΔT_overvoltage) and the fluctuation amplitude (σ_T_overvoltage) are expressed by the following formulas:

[0117] -ΔT_overvoltage=(T_t-T_{t-1}) / Δt|Voltage>4.2V (Δt is the time difference in minutes)

[0118] -σ_T_overvoltage=std(temperature|voltage>4.2V) (standard deviation of temperature data)

[0119] The BMS uploads charging voltage and temperature data to the cloud in real time. The cloud filters abnormal charging data with "voltage > 4.2V" and calculates the two parameters according to the above formula. If data is missing (e.g., temperature is not uploaded at a certain moment), linear interpolation is used to complete it. For example, an electric vehicle BMS uploads overvoltage charging data to the cloud: voltage 4.3V for 10 minutes, temperature rises from 25°C to 26°C, 28°C, 29°C, and 27°C (a total of 10 data points); the cloud calculates: ΔT_overvoltage=(29-25) / 10=0.4°C / min (mean), σ_T_overvoltage=std([25,26,28,29,27])=1.58°C (normalized to 0.3°C); and generates overvoltage charging range temperature characteristics: 0.4°C / min, 0.3°C. Overvoltage charging is the core cause of thermal runaway. This feature directly quantifies the temperature dynamics during overvoltage charging. If ΔT_overvoltage>0.5°C / min, it indicates that the internal side reactions of the battery are aggravated (such as electrolyte decomposition). Compared with existing technologies that only count the number of overvoltages, it can more accurately capture the synergistic risk of "overvoltage + temperature rise" and provide key signals for early warning of thermal runaway.

[0120] Within the SOC < 10% (deep discharge range), the voltage fluctuation range is calculated using the following formula:

[0121] σ_V_deepdischarge=std(voltage|SOC<10%) (standard deviation of voltage data)

[0122] The Battery Management System (BMS) uploads SOC and voltage data to the cloud in real time. The cloud filters discharge data with "SOC < 10%", removes outliers (such as noise points with voltage drops exceeding 0.1V, using the 3σ principle), and calculates the voltage standard deviation as the fluctuation range. For example, a battery BMS uploads deep discharge data (SOC 9%-5%): voltages 3.2V, 3.18V, 3.22V, 3.15V, 3.19V (a total of 5 valid data points); the cloud calculates: σ_V_deepdischarge=std([3.2,3.18,3.22,3.15,3.19])=0.026V (0.08V after normalization); generating the deep discharge range voltage characteristic: 0.08V. Deep discharge can easily damage the electrode structure. When the voltage fluctuation exceeds 0.1V, it indicates abnormal electrode polarization (such as the shedding of lithium dendrites from the negative electrode). This feature overcomes the limitation of existing technologies that only focus on the number of discharges. By quantifying the degree of damage to the battery caused by deep discharge through voltage stability, it improves the accuracy of identifying "deep discharge-induced thermal runaway".

[0123] Abnormal operating conditions are the direct cause of thermal runaway. Segmented features can quantify the impact of abnormal operating conditions on the battery (e.g., an overvoltage charging temperature rise rate > 0.5°C / min indicates risk). Compared with full-condition statistics, abnormal interval features can better focus on high-risk scenarios and improve the accuracy of early warning.

[0124] This embodiment defines five specific forms of segmented features, covering five major interval dimensions: "SOC-temperature-current-time-operating condition". The core is "dividing intervals according to risk scenarios and extracting features within each interval". During execution, the BMS provides multi-dimensional real-time data, and the cloud filters data and calculates features according to preset intervals, ensuring the "interval specificity" and "risk focus" of the segmented features. Compared with existing technologies: interval division is based on battery safety mechanisms (such as extreme temperatures and abnormal operating conditions being high-risk intervals), ensuring that features are strongly correlated with thermal runaway risk; each interval feature is a combination of "interval + target feature" (such as "high temperature interval + internal resistance change rate"), which is better at detecting local anomalies than full-range features; it provides "interval dimension" input (such as "high temperature interval internal resistance feature × fast charging frequency") for subsequent DeepFM cross-feature generation, improving the ability to identify the collaborative risks of "interval anomalies + operating condition anomalies". The overall solution makes segmented features a "high-risk scenario grasp" for thermal runaway prediction, solving the problem of insufficient identification of local anomalies in existing technologies, and providing a key basis for accurate early warning. Furthermore, the interval division is more risk-focused, with all intervals (e.g., SOC < 10%, voltage > 4.2V) classified as "high-risk thermal runaway scenarios," avoiding meaningless interval divisions and ensuring a strong correlation between features and risks. Parameter calculations are more quantitative, with each feature employing a clear formula to replace the qualitative descriptions of existing technologies, making features reproducible and comparable. Data sources are more reliable, relying on real-time data directly uploaded by the BMS to avoid errors from manual collection, while centralized cloud processing ensures consistency during large-scale fleet applications. Risk identification is more accurate, capturing "hidden risks" that are difficult to detect with existing technologies through a combination of "interval + dynamic features" (e.g., overpressure charging + temperature rise rate) (e.g., overpressure but no temperature rise → low risk, overpressure and rapid temperature rise → high risk), providing "high-discrimination input" (e.g., "overpressure temperature rise rate × number of overpressure charging cycles") for subsequent cross-feature generation.

[0125] In one embodiment, the cross-features of the battery's operating condition characteristics, timing characteristics, electrical characteristics, and segmentation characteristics are calculated to generate a fused cross-feature matrix, including:

[0126] The cross-features of the operating condition features, time sequence features, electrical features and segmentation features are calculated using a preset machine learning model, optimization algorithm or manual rules to generate a fused cross-feature matrix.

[0127] The aforementioned machine learning models refer to models that automatically learn the interaction relationships between features through algorithms, including Deep Factorization Machines (DeepFM) and Transformer-based attention mechanisms. The aforementioned optimization algorithms refer to algorithms that generate optimal feature combinations through iterative optimization, such as Genetic Algorithms (GA). The aforementioned manual rules refer to feature combination logic pre-set based on expert knowledge, such as "voltage features within a fixed SOC range" (e.g., SOC∈[30%, 90%] voltage × overvoltage charging times).

[0128] In this embodiment, three cross-feature generation methods are preset in the cloud and automatically selected according to the application scenario: large-scale fleet real-time prediction (e.g., 100,000 vehicles) → prioritize DeepFM model (high efficiency and automation); small batch of special vehicle models (e.g., energy storage station batteries) → can be combined with genetic algorithm to optimize feature combination; emergency fault diagnosis → enable manual rules (quickly locate known risk patterns); the input of all methods is the four types of features (operating condition, time series, electrical, segmentation) extracted in the above embodiment, and must first be normalized by Min-Max to be unified to the [0,1] interval. For example, a car manufacturer processes real-time data from 100,000 electric vehicles in the cloud, selecting the DeepFM model to generate cross-features: input features include "overvoltage charging times (operating condition, 0.6)," "daily average voltage (time series, 0.8)," "internal resistance change rate (electrical, 0.3)," and "SOC [30%, 90%] average voltage (segmented, 0.9)" (values ​​in parentheses are normalized). Simultaneously, a genetic algorithm is used for 50 battery groups in an energy storage power station to optimize the optimal combination weight of "high-temperature internal resistance change rate × fast charging frequency." Manual rules are used for 3 faulty vehicles to directly calculate "SOC < 10% voltage fluctuation × deep discharge times." Existing technologies rely on manual feature combination, resulting in low efficiency and incomplete coverage. This solution provides three complementary approaches: 1. Machine learning models achieve fully automatic cross-feature generation (no manual design required); 2. Optimized algorithms improve the optimality of the combination; 3. Manual rules ensure rapid response in emergency scenarios. Compared to the single approach of existing technologies, this solution balances "efficiency, accuracy, and flexibility," meeting the needs of different scenarios.

[0129] The aforementioned fused cross-feature matrix integrates two-dimensional structured data of all cross-features. Rows represent timestamps, and columns represent different cross-features. For example, the matrix may contain columns such as "SOC range voltage × overvoltage charging" or "high temperature internal resistance × fast charging frequency." The cloud performs structured integration on the cross-features generated in the selected manner: alignment by timestamp (e.g., generating one row per hour, corresponding to all cross-features of that hour); removal of invalid cross-features, such as null values ​​caused by missing features, which are filled in using the average of adjacent timestamps; and matrix dimensionality adaptation to subsequent prediction models (e.g., XGBoost requires an input dimension ≤ 100, requiring dimensionality reduction through principal component analysis, etc.). For example, the cloud integrates the cross-features of a certain electric vehicle from 08:00 to 09:00 on August 24, 2025: timestamp 2025-08-24 09:00, cross-features include "overvoltage charging times × SOC [30%, 90%] voltage (0.6 × 0.9 = 0.54)", "high temperature internal resistance change rate × fast charging frequency (0.5 × 0.8 = 0.4)", and "voltage change rate × deep discharge times (0.3 × 0.2 = 0.06)", generating a 1-row, 3-column matrix; if the "fast charging frequency" data is missing, it is filled in with 0.8 from the previous hour, and the final matrix has no empty values. The fused cross-feature matrix solves the problem of "scattered and unrelated features" in existing technologies, structuring the cross-information of the four types of features, which is convenient for batch processing by subsequent prediction models; at the same time, timestamp alignment ensures the consistency of data time sequence, and the completion mechanism ensures the integrity of the matrix, providing regular input data for high-precision risk prediction.

[0130] In this embodiment, DeepFM and other models automatically generate cross features, eliminating the need for manual design (such as high-order cross features like "voltage × SOC × temperature"), thus solving the problems of "high manual costs and incomplete coverage" in existing technologies. Cross features uncover potential correlations among four types of features (such as "operating condition × segmentation" reflecting the synergistic risk of "overvoltage charging + specific SOC range"), and can discover high-order nonlinear cross features that are difficult to detect using traditional methods. This matrix integrates these high-value features, laying the foundation for improving subsequent prediction accuracy. It provides a technical path for upgrading thermal runaway prediction from "single feature" to "multi-dimensional cross features."

[0131] Furthermore, the aforementioned machine learning model is a deep factorization machine model, wherein the factorization machine (FM) part of the deep factorization machine model automatically learns the second-order cross features of the input features, and the deep neural network (DNN) part learns the high-order nonlinear interaction features of the input features to generate a fused cross feature matrix.

[0132] The DeepFM model is deployed on the server in the cloud and initialized in two steps: Model training: Input historical data labeled with thermal runaway (e.g., 5000 sets of "normal battery" data + 500 sets of "thermal runaway battery" data), the loss function is binary cross-entropy, the optimizer is Adam, and iterates for 100 rounds until the accuracy on the validation set is stable; Input adaptation: The above four types of features, namely operating condition features, time series features, electrical features and segmentation features, are converted into dense vectors through the embedding layer and used as the common input of FM and DNN. For example, when training the DeepFM model in the cloud, the historical data includes "thermal runaway cases": a battery overcharged 3 times (x1=0.6), the average SOC [30%, 90%] voltage was 3.85V (x2=0.9), the internal resistance change rate was 0.5mΩ / h (x3=0.5), and the fast charging temperature rise rate was 0.8°C / min (x4=0.8). After the embedding layer converts x1-x4 into a 32-dimensional vector, the FM part calculates the second-order cross: such as ⟨v1,v2>x1x2=0.8×0.6×0.9=0.432), and the DNN part learns the third-order cross of x1-x4 through a 3-layer network (such as x1×x2×x3), and finally outputs the cross feature vector [0.432,0.36,0.24,...] (32 dimensions in total). Existing technologies "cannot automatically generate high-order cross features," but DeepFM's FM+DNN architecture solves this problem: FM automatically learns second-order crosses (such as "operating condition × segmentation") without manual design; DNN mines high-order nonlinear interactions (such as "operating condition × time series × electrical"), covering the blind spots of traditional methods; shared embedding layers reduce parameter redundancy and improve cloud training and inference efficiency (the number of parameters is reduced by 40% compared to independent embedding layers).

[0133] The trained DeepFM model performs cross-feature calculations on real-time data once per hour in the cloud (adapting to the BMS data upload cycle); the cross-features output by inference are structured into a matrix of "timestamp-feature name", including columns such as "SOC range voltage × overvoltage charging" and "high temperature internal resistance × fast charging frequency". During cloud-based inference, real-time data is processed according to the following process: Data Input: Receive four types of features uploaded by BMS at a certain moment (e.g., the number of overvoltage charging times of 2 and the SOC range voltage of 3.8V at 10:00 on August 24, 2025, etc.), and input them into DeepFM after preprocessing; Feature Crossing: The FM part outputs all second-order cross features (e.g., x1×x2, x1×x3, etc.), and the DNN part outputs 10 core high-order cross features (e.g., x1×x2×x3, x2×x3×x4, etc.); Matrix Integration: Align the second-order and high-order cross features according to the timestamp to generate a 1-row N-column matrix (N = number of second-order crosses + number of high-order crosses, e.g., 20 + 10 = 30 columns).

[0134] In this embodiment, DeepFM improves the prediction accuracy from 40% to over 70%. The core reason is that the matrix generated in this step covers second-order and higher-order cross-features, with an information density far exceeding that of single features in existing technologies. The cross-features are directly related to the risk of thermal runaway (e.g., x2×x4 reflects the synergistic risk of "common SOC range voltage + fast charging temperature rise"). The matrix structure facilitates subsequent batch processing of XGBoost models, and the cloud inference time is less than 1 second, meeting the real-time prediction requirements.

[0135] Compared with existing technologies, this embodiment uses FM+DNN to share an embedding layer, which retains the high efficiency of FM second-order crossover and has the high-order nonlinear modeling capability of DNN, thus solving the contradiction of existing technologies that "either can only perform second-order crossover or can only perform high-order crossover but with redundant parameters." There is no need to manually design crossover rules (such as "which SOC range should be matched with overvoltage charging"), the model automatically learns the optimal crossover combination, significantly improving development efficiency. Accuracy innovation: high-order crossover features capture "multi-dimensional collaborative risks" (such as "overvoltage + fast charging + high temperature"), which cannot be covered by single features or manual crossover in existing technologies, thus improving prediction accuracy.

[0136] In one specific embodiment, the calculation of the cross-features of the battery's operating condition characteristics, timing characteristics, electrical characteristics, and segmented characteristics to generate a fused cross-feature matrix includes:

[0137] The operating condition characteristics, timing characteristics, electrical characteristics, and segmentation characteristics are combined into an input feature set;

[0138] The input feature set is processed by a factorization machine. For any two different features in the input feature set, the latent vectors corresponding to the two features are obtained and the inner product operation is performed. Then, the values ​​of the two features are combined to calculate the results of multiple pairs of feature interactions. The results of multiple pairs of feature interactions are combined to form a second-order cross feature vector.

[0139] The input feature set is processed by a deep neural network. First, the various features in the input feature set are converted into a dense feature embedding matrix through an embedding layer. Then, the feature embedding matrix is ​​input into a network structure containing at least two fully connected layers. The ReLU activation function and the Sigmoid activation function are used in sequence for nonlinear transformation to obtain multiple sets of high-order interaction results of multiple features. The results of multiple sets of high-order interaction of multiple features are combined into a high-order cross feature vector.

[0140] The second-order cross feature vector and the higher-order cross feature vector are weighted and fused based on the fusion weight to obtain the fused cross feature vector. The value range of the fusion weight is [0.4, 0.6], and the fusion weight is determined by maximizing the accuracy of the battery thermal runaway historical data validation set.

[0141] The fused cross feature vectors are aligned sequentially according to the timestamps of the battery operation data to generate a fused cross feature matrix. The rows of the cross feature matrix correspond to different timestamps, and the columns correspond to different cross features in the fused cross feature vectors. The matrix dimension is determined by the time series length, the dimension of the second-order cross feature vector, and the dimension of the higher-order cross feature vector.

[0142] In this embodiment, to achieve quantitative capture of the nonlinear correlation between the above four types of features, this scheme uses an improved DeepFM model for cross-feature generation, specifically implemented through the following formula:

[0143] The formula for calculating the fused cross feature vector is:

[0144]

[0145] α is the fusion weight (value range [0.4, 0.6], in this embodiment it is 0.5, determined by maximizing the accuracy of the validation set);

[0146] The second-order cross feature vector output by the FM part is calculated as follows:

[0147] ( =1,2,...,m)

[0148] In the formula, m is the second-order cross feature dimension (in this embodiment) ), , Features , The latent vector (32-dimensional, optimized through model training). , which is the implicit vector inner product;

[0149] The higher-order cross feature vector output by the DNN part is calculated as follows:

[0150]

[0151] In the formula For the feature embedding matrix ( for Dense vector transformed by the embedding layer (32 dimensions). (32×64) (64×10) is the weight matrix of the fully connected layer. (64 dimensions) (10-dimensional) is the bias vector. , These are ReLU and Sigmoid activation functions, respectively (in this embodiment, the higher-order cross feature dimension is 10).

[0152] The above-calculated fusion cross feature vector (Dimensional (m+10=16)) Aligned by timestamps to generate a fused cross-feature matrix. Where T is the time series length (in this example, T=24, corresponding to 24 hours of data), and the matrix elements are... , (t=1,...,24, s=1,...,16).

[0153] For example, a battery has input features X = [0.6, 0.8, 0.3, 0.9] (0.6 overvoltage charging cycles, 0.8 average voltage per day, 0.3 rate of change of internal resistance, and 0.9 average voltage across the SOC range), and a latent vector... =[0.1,0.2,...,0.05]、 =[0.3,0.1,...,0.2] (only the first two dimensions are shown for illustration), then the second-order cross feature = (0.1×0.3 + 0.2×0.1) ×0.6×0.9 = 0.05×0.54=0.027; The DNN part is calculated through the embedding layer and the fully connected layer. =[0.12, 0.08, ..., 0.05] (10 dimensions); final fused feature vector = 0.5×[0.027,0.048,...]+ 0.5×[0.12,0.08,...]= [0.0735, 0.064, ..., 0.025], which are then combined into matrix rows by timestamp.

[0154] The improved DeepFM model in this embodiment features feature interactions that better align with the physical and risk characteristics of the battery field. General DeepFM is geared towards scenarios like advertising and recommendation, but its latent vectors and interaction logic lack industry-specificity. The improved version, through customized latent vector initialization and training objectives (using battery thermal runaway labels as supervision), allows feature latent vectors (such as those for "overvoltage charging" and "SOC range voltage") to more accurately encode electrochemical risk semantics. For example, the inner product of the latent vector for overvoltage charging and the latent vector for high temperature more accurately reflects the increase in thermal runaway probability when "overvoltage + high temperature" are combined, making the feature crossover results more closely match the actual mechanism of battery thermal runaway. The fusion weights are dynamically adapted to the characteristics of battery risk propagation. In general DeepFM, the FM (second-order cross) and DNN (higher-order cross) fusion weights are mostly fixed values ​​or trained based on general tasks. In the improved version, the fusion weight α is optimized through the validation set of historical battery thermal runaway data (such as α=0.5 in the paper). It can balance "low-order, interpretable pairwise feature correlations (such as 'overvoltage charging + SOC range anomaly')" and "high-order, complex multi-feature chain reactions (such as the chain correlation of 'overvoltage → internal resistance change → temperature anomaly')". It retains the risk logic that can be explained by experts and captures potential correlations that are difficult to summarize manually, thereby improving the accuracy and interpretability of prediction. Heterogeneous feature embedding is better suited to battery data types. Battery features include discrete types (such as "fast charging / slow charging" labels), continuous types (such as voltage values), and interval types (such as SOC segments). The improved embedding layer uses categorized feature-specific encoding and continuous feature normalization + smooth embedding to make different types of features more consistent with electrochemical laws when transformed into dense vectors (e.g., for interval features such as "low temperature range (-10℃~0℃)," the embedded vector can reflect the correlation between temperature range and thermal runaway). This avoids the embedding bias caused by feature type mismatch in general DeepFM, laying a more reliable feature representation foundation for subsequent cross-computation. Computational efficiency is adapted to large-scale real-time battery monitoring. Real-time monitoring of large fleets (such as 100,000 vehicles) requires high inference speed. The improved version simplifies the DNN network structure (e.g., using only two fully connected layers, with the dimension reduced from 32→64→10) and optimizes the hidden vector dimension (set to 32), significantly reducing the latency and computing power consumption of real-time cloud computing while ensuring cross-feature representation capabilities. Compared to the general DeepFM (which designs deeper networks for complex tasks), it is better able to meet the needs of battery monitoring scenarios that require "low latency and high concurrency".The improved version demonstrates stronger generalization ability in small sample scenarios. Thermal runaway exists in rare scenarios with "low probability and high risk" (such as fast charging at extreme low temperatures). The general DeepFM is prone to overfitting in small sample scenarios. The improved version incorporates prior knowledge in the battery field (such as high-risk feature combinations from the rule engine as regularization constraints), allowing the model to focus more on the correlation patterns that physically conform to the thermal runaway mechanism when learning cross features. Even in rare scenarios, it can make predictions based on "generalization of known high-risk combinations", thus improving the generalization ability in small sample scenarios.

[0155] In one embodiment, predicting the probability of thermal runaway of the battery based on the cross-feature matrix includes:

[0156] The cross-feature matrix is ​​input into a preset prediction model, and risk analysis is performed by combining expert rules in the rule engine and historical cases in the knowledge base to generate the probability of thermal runaway and the corresponding risk level.

[0157] The preset prediction model takes as input a cross-feature matrix generated by DeepFM and outputs a preliminary probability of thermal runaway risk (0-1). The training data for model training consists of a labeled historical cross-feature matrix (e.g., "normal" label probability 0, "thermal runaway" label probability 1), with a loss function of binary cross-entropy and an optimizer of Adam, iterating until the validation set AUC > 0.85. The cloud then inputs the cross-feature matrix (e.g., 1 row and 30 columns of cross-features per hour) into the trained prediction model, which calculates and outputs a preliminary probability. If the matrix contains missing values ​​(e.g., a certain cross-feature is not generated), the model's built-in mean imputation mechanism is used. For example, the cloud inputs the cross-feature matrix (containing 30 features such as "x1×x2=0.32" and "x2×x4=0.48") of a certain electric vehicle at 10:00 on August 24, 2025 into the XGBoost model. The model calculates that the initial probability of thermal runaway risk is 0.45. If "x3×x4=0.18" is missing, it is filled with 0.18 from the previous hour, and the final probability deviation is <0.05.

[0158] The aforementioned rule engine is a collection of expert rules stored in a cloud database, such as "Rule 1: SOC range voltage > 4.0V and overvoltage charging × voltage fluctuation > 0.5 → triggering a level 1 warning"; the aforementioned knowledge base refers to a case library, which stores typical thermal runaway cases (such as "Case 1: 3 overvoltage charges + high temperature internal resistance change rate of 0.5mΩ / h → thermal runaway") and expert knowledge (such as "lithium dendrite growth leads to a sudden increase in internal resistance, which easily triggers thermal runaway"), etc.

[0159] The cloud-based system performs secondary risk analysis according to the following process: Rule matching: The core features in the cross-feature matrix (such as "overvoltage charging × SOC range voltage") are compared with the warning rules in the rule engine to determine whether the rule is triggered; Case matching: The similarity between the real-time cross-features and the cases in the knowledge base is calculated (e.g., a cosine similarity > 0.8 is considered a match), and the risk level corresponding to the case is extracted; Probability correction: The initial probability output by the prediction model t is corrected by combining the rule triggering result and the case matching result (e.g., if a level 1 warning rule is triggered, the probability increases by 0.3; if a low-risk case is matched, the probability decreases by 0.1). For example, the cloud analyzes a battery with an initial probability of 0.45: Rule matching: The cross-feature "overvoltage charging × SOC range voltage = 0.55 > 0.5" triggers rule 1 (level 1 warning), and the probability is corrected to 0.45 + 0.3 = 0.75; Case matching: The real-time feature has a similarity of 0.85 with the knowledge base "Case 2: overvoltage charging twice + fast charging temperature rise rate 0.6°C / min → risk probability 0.7", and the match is successful, with the probability corrected to 0.75 - 0.05 = 0.7; The final risk probability is 0.7, corresponding to level 2 risk. Existing technologies "lack expert knowledge assistance". This solution uses a rule engine + knowledge base to: compensate for the "rare case omission" problem of pure data-driven models (such as the case of "sudden increase in internal resistance after long-term parking" in the knowledge base, which the model has not seen but can be identified through matching); correct model prediction bias (such as the model outputting a low probability due to data noise, but the rule triggers a high risk warning); improve the interpretability of the results (such as "high risk is triggered by rule 1 + matching case 2"), and solve the "black box prediction" problem of existing technologies.

[0160] The aforementioned risk levels are stored in the cloud, such as Level 1 (0.8-1.0, push notification), Level 2 (0.5-0.8, key monitoring), Level 3 (0.2-0.5, increased monitoring frequency), and Normal (0-0.2, normal operation). The cloud compares the corrected risk probability with the level threshold to determine the risk level and generates a report containing "probability-level-cause". The report is pushed to the vehicle manufacturer's management platform or user APP via the cloud API (e.g., for Level 2 risk, push notification "Current probability of battery thermal runaway is 0.7, it is recommended to reduce fast charging"). The clear risk level facilitates quick understanding and action by users. Compared to existing technologies that only output probabilities, this solution's "probability + level + cause" report reduces the user's understanding cost (no need to interpret probability values); provides action guidance (e.g., Level 2 risk recommends reducing fast charging); supports refined management by vehicle manufacturers (e.g., key monitoring of Level 2 risk vehicles); and improves the practicality of battery safety management.

[0161] In this embodiment, a pure data-driven model (predictive model, such as XGBoost) ensures accuracy, a rule engine ensures "hard constraints" (such as mandatory overpressure warnings), and a knowledge base ensures "coverage of rare cases." These three elements work together to address the "limitations of a single mechanism" in existing technologies. Rule triggering and case matching provide the causes of risks, addressing the pain point of "unexplainable prediction results" in existing technologies, making it easier for automakers and users to trace the source of risks. The rule engine and knowledge base can be updated as needed (such as adding "low-temperature fast charging risk rules") to meet the risk analysis needs of different scenarios.

[0162] Furthermore, the above prediction model is a model trained based on an XGBoost model, an LSTM model, or a Bayesian network model.

[0163] The XGBoost model is an ensemble learning model based on Gradient Boosting Decision Tree (GBDT), which excels at handling high-dimensional sparse features (such as cross-feature matrices). XGBoost improves prediction accuracy to over 70%, and its core advantages are strong resistance to overfitting; robustness to missing values, eliminating the need for additional handling of null values ​​in cross-features; and fast inference speed (<0.5 seconds per sample in the cloud), making it suitable for large-scale real-time fleet monitoring.

[0164] The LSTM model mentioned above is a long short-term memory network, which is good at processing time series data (such as cross feature matrices of multiple consecutive hours) and is suitable for dynamic time series prediction. Compared with XGBoost, which only uses features at a single time step, LSTM utilizes information from multiple time steps, which can detect trend risks earlier (such as features that continue to rise) and provide early warnings 12-24 hours in advance, solving the problem of "early warning lag" in existing technologies; however, the inference speed is relatively slow (single sequence < 2 seconds), making it suitable for scenarios with high requirements for early warning time.

[0165] The aforementioned Bayesian network model is a probabilistic graphical model that represents risk relationships through nodes (cross-features) and edges (probabilistic dependencies), outputting a probability distribution. Bayesian networks are easy to interpret probabilistically. Compared to the "black box" nature of XGBoost, its conditional probability table can clearly explain "the reason for the increased probability of thermal runaway when feature 1 = 0.6," making it suitable for expert troubleshooting. Furthermore, it is robust to small sample data (e.g., it can still be trained with only 1000 sets of historical data), solving the problem of "low accuracy in small sample scenarios" in existing technologies.

[0166] In one specific embodiment, the prediction model described above is a model trained based on the XGBoost model, and the output initial thermal runaway probability is: (Values ​​range from 0 to 1), incorporating two correction factors: rule matching degree and case similarity, to ultimately determine the probability of thermal runaway risk. The calculation formula is:

[0167]

[0168] in, , These are the rule correction weight and the case correction weight (with values ​​ranging from [0.2, 0.5], as specified in this embodiment). =0.3、 =0.2 (determined through verification using the historical case database).

[0169] The rule matching degree is calculated as follows: In the formula, P represents the total number of warning rules in the rule engine (P=3 in this embodiment, corresponding to level 1, level 2, and level 3 warning rules). The weight of rule p (the higher the risk level, the greater the weight, such as...). =0.6、 =0.3、 =0.1), To define the filtering range for the matching rules, This is an indicator function (1 if the p-th rule is triggered, 0 otherwise);

[0170] The formula for calculating case similarity is: In the formula, Q is the total number of typical thermal runaway cases in the knowledge base (Q=100 in this embodiment). Let be the fusion cross feature vector of the q-th case. For vector dot product, It is the L2 norm (similarity value ranges from 0 to 1, the closer to 1, the higher the case matching degree).

[0171] For example, the initial probability of a certain battery =0.45, rule matching result: Level 1 warning rule triggered (p=1, I=1), Level 2 / 3 rule not triggered (I=0), then =(0.6×1+0.3×0+0.1×0) / (0.6+0.3+0.1)=0.6; Case matching result: The highest similarity is with case number 25 (overvoltage charging + fast charging temperature rise). =0.85; final probability =(0.45+0.3×0.6+0.2×0.85) / (1+0.3+0.2)=(0.45+0.18+0.17) / 1.5=0.8 / 1.5≈0.53, corresponding to level three risk (0.2-0.5 is level three, but since the adjusted value is close to 0.5, it is actually monitored as level three).

[0172] The thermal runaway probability correction formula in this embodiment is customized for battery thermal runaway scenarios. It aligns with the risk mechanisms in the battery field, differing from general probability correction (purely data-driven). The formula incorporates battery-specific early warning rules (such as the high-risk combination of overvoltage charging and voltage fluctuations) and historical thermal runaway case characteristics, making the correction logic more closely match the actual mechanism of "electrochemical anomaly → thermal runaway." Dynamic balancing of multi-source risks is based on optimization through a battery case library. (Rule weight) (Case weighting) allows for flexible balancing of "model statistical prediction," "expert rule certainty risk," and "historical case similarity risk" based on the scenario. For example, it strengthens rule correction when high-risk rules are triggered and supplements historical experience when rare cases occur. It improves reliability in extreme / small-sample scenarios. Thermal runaway exists in low-probability extreme scenarios such as "ultra-low temperature fast charging," which are easily missed by single models. The formula captures the logic of extreme operating conditions known to experts through rule matching and calls upon historical similar scenario features through case similarity, solving the problem of insufficient generalization of the model in small-sample extreme scenarios. It enhances risk interpretability and traceability. Rule matching can clearly identify "which warning rules are triggered (e.g., a level 1 warning corresponds to overpressure + voltage fluctuation)," and case similarity can locate "which type of historical thermal runaway case is similar," facilitating technicians to trace the root cause of risks and assisting in battery improvement or maintenance decisions. It adapts to differentiated scenarios in large-scale fleets: through rules (covering common warnings for different batteries) and cases (covering historical issues for different vehicle models), it can flexibly adapt to the characteristics of "large differences in battery type and operating conditions" in large-scale fleets. Compared to general correction, it is more scenario-specific and improves the predictive stability across vehicle models and battery types.

[0173] In one embodiment, the acquisition of the battery's operating condition characteristics, timing characteristics, electrical characteristics, and segmentation characteristics includes:

[0174] Collect data on battery temperature, voltage, current, internal resistance, SOC, and user behavior to obtain a raw data set.

[0175] The aforementioned raw data collection includes multi-dimensional raw data, including temperature, voltage, current, internal resistance, SOC (electrical), and user behavior data (such as overvoltage charging times, parking time, and operating conditions), which are collected and uploaded in real time by the vehicle's BMS. The vehicle's BMS collects real-time data through onboard sensors (temperature sensors, voltage sensors, etc.) and uploads it to the cloud according to the following process: Data encapsulation: The data is encapsulated into JSON format according to "timestamp + parameter name + value + unit"; Network transmission: The data is uploaded to the cloud data receiving interface via 4G / 5G or V2X network; Data storage: The cloud stores the received raw data in a time-series database (such as InfluxDB), indexed by vehicle VIN code and timestamp.

[0176] The original data set is processed for outliers, format standardization, and numerical normalization to generate a multidimensional feature dataset.

[0177] In this step, outlier handling includes removing outliers using the 3σ principle (data exceeding μ ± 3σ is considered outlier) or the IQR method (data exceeding Q1 - 1.5IQR to Q3 + 1.5IQR is considered outlier), and filling in missing values ​​using linear interpolation. Format standardization refers to unifying the data format (JSON → CSV), timestamp format (YYYY-MM-DDHH:MM:SS), numerical precision (1 decimal place for temperature, 2 decimal places for voltage), and encoding (UTF-8). Numerical normalization refers to using Min-Max normalization (X_normalized = (X - X_min) / (X_max - X_min)) or Z-Score normalization (X_standardized = (X - μ) / σ) to unify the distribution to [0,1] or a standard normal distribution.

[0178] The cloud preprocesses the raw data according to the following process: Outlier handling: For voltage data (such as 3.72V, 10.0V (outlier), 3.71V), 10.0V is removed using the 3σ principle, and the missing 3.73V is filled in using linear interpolation ((3.72+3.71) / 2=3.715V); Format standardization: JSON is converted to CSV, the timestamp is unified as "2025-08-24 08:00:00", the temperature is retained to 1 decimal place (25.3°C), and the voltage is retained to 2 decimal places (3.72V); Numerical normalization: The temperature (range -20°C to 60°C) is normalized using Min-Max, 25.3°C→(25.3+20) / (60+20)=0.566. This embodiment ensures data quality and model input consistency. Outlier handling reduces the impact of noise on feature extraction (e.g., removing 50A current to avoid subsequent time-series feature deviations); format standardization facilitates batch processing in the cloud (e.g., CSV facilitates analysis using Python pandas); numerical normalization eliminates the influence of dimensions (e.g., temperature °C and voltage V are unified to [0,1]), avoiding overfitting of the model to large numerical features (e.g., small internal resistance mΩ values ​​are ignored), providing high-quality data for subsequent feature extraction and model training.

[0179] Based on the multidimensional feature dataset, operating condition features, time series features, electrical features, and segmentation features are extracted respectively.

[0180] The multidimensional feature dataset is preprocessed structured data containing normalized values ​​of electrical and behavioral data. Operating condition features are extracted from behavioral data (e.g., number of overvoltage charging cycles), time-series features from time series data (e.g., daily average voltage), electrical features from electrical data (e.g., rate of change of internal resistance), and segmented features from interval data (e.g., SOC ∈ [30%, 90%] voltage). Based on the multidimensional feature dataset, the cloud platform extracts features by category: Operating condition features: statistics on the number of overvoltage charging cycles (3 times), the number of ultra-low temperature charging cycles (1 time), etc.; Time-series features: calculation of the daily average voltage (3.72V → normalized 0.83), and the daily variance of current (0.05). →Normalization 0.2), etc.; Electrical characteristics: calculate the rate of change of internal resistance (0.3mΩ / h → normalization 0.3), SOC-voltage correlation coefficient (0.92 → normalization 0.92), etc.; Segmented characteristics: calculate the average voltage of SOC∈[30%,90%] (3.85V → normalization 0.86), the rate of change of internal resistance at high temperature (0.5mΩ / h → normalization 0.5), etc.

[0181] Multi-dimensional features cover the entire battery state. This step extracts features according to the categories of "operating condition-time sequence-electrical-segmentation" to ensure that no features are omitted. At the same time, based on the high-quality data after preprocessing, the feature accuracy is high (e.g., the error of internal resistance change rate is <0.05). Compared with the existing technology that only extracts 1-2 types of features, this solution uses four types of features in synergy to provide rich input for subsequent cross-feature generation and solve the problem of "insufficient feature coverage".

[0182] In this embodiment, data acquisition is comprehensive: the BMS simultaneously collects electrical and behavioral data, providing support for operating condition feature extraction and solving the problem of "insufficient utilization of operating condition features" in existing technologies; preprocessing is standardized: outlier handling, format standardization, and numerical normalization are implemented to ensure data quality and lay the foundation for feature extraction and model training; the feature extraction system covers four types of features: "usage environment - time variation - electrical performance - interval state", without omissions and with clear classification, improving the accuracy of subsequent risk identification.

[0183] Reference Figure 2 The present invention also provides a computer device, the internal structure of which can be as follows: Figure 2As shown, this 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 non-volatile storage media and internal memory. The non-volatile storage media stores operating devices, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores knowledge bases, rule bases, etc. The network interface is used for communication with external terminals via a network connection. Furthermore, the computer device may also include input devices and a display screen. When the aforementioned computer program is executed by a processor, it implements a battery thermal runaway prediction method based on cross-feature fusion. The method includes: acquiring battery operating condition characteristics, time-series characteristics, electrical characteristics, and segmented characteristics, wherein the operating condition characteristics reflect the actual usage environment of the battery, the time-series characteristics reflect the time-series changes in the battery state, the electrical characteristics reflect the battery's electrical performance, and the segmented characteristics reflect feature statistics or changes within a specific interval; calculating the cross-features of the battery's operating condition characteristics, time-series characteristics, electrical characteristics, and segmented characteristics to generate a fused cross-feature matrix; and predicting the probability of the battery experiencing thermal runaway based on the cross-feature matrix. Those skilled in the art will understand that... Figure 2 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0184] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A battery thermal runaway prediction method based on cross-feature fusion, characterized in that, The method comprises: obtaining the working condition characteristics, timing characteristics, electrical characteristics and segmented characteristics of the battery, wherein the working condition characteristics reflect the actual use environment of the battery, the timing characteristics reflect the time sequence change of the battery state, the electrical characteristics reflect the electrical performance of the battery, and the segmented characteristics reflect the feature statistics or change in a specific interval; calculating the cross characteristics of the working condition characteristics, timing characteristics, electrical characteristics and segmented characteristics of the battery to generate a fused cross characteristic matrix; based on the cross characteristic matrix, predicting the probability of thermal runaway of the battery; the calculation of the cross characteristics of the working condition characteristics, timing characteristics, electrical characteristics and segmented characteristics of the battery to generate a fused cross characteristic matrix comprises: grouping the working condition characteristics, timing characteristics, electrical characteristics and segmented characteristics into an input feature set; processing the input feature set through a factor decomposition machine part, for any two different characteristics in the input feature set, obtaining the hidden vectors corresponding to the two types of characteristics and performing inner product operation, and then combining the numerical values of the two types of characteristics to obtain the results of the interaction of the two types of characteristics, and grouping the results of the interaction of the two types of characteristics into a second-order cross characteristic vector; processing the input feature set through a deep neural network part, first converting each type of feature in the input feature set into a dense feature embedding matrix through an embedding layer, then inputting the feature embedding matrix into a network structure comprising at least two fully connected layers, and sequentially performing nonlinear transformation through a ReLU activation function and a Sigmoid activation function to obtain the results of the high-order interaction of multiple features, and grouping the results of the high-order interaction of multiple features into a high-order cross characteristic vector; based on the fusion weight, the second-order cross characteristic vector and the high-order cross characteristic vector are weighted and fused to obtain a fused cross characteristic vector, wherein the fusion weight is determined by maximizing the accuracy of the battery thermal runaway historical data validation set; aligning the fused cross characteristic vector according to the time stamp of the battery operation data in sequence to generate a fused cross characteristic matrix, wherein the rows of the cross characteristic matrix correspond to different time stamps, the columns correspond to different cross characteristics in the fused cross characteristic vector, and the matrix dimension is determined by the time sequence length, the dimension of the second-order cross characteristic vector and the dimension of the high-order cross characteristic vector. 2.The battery thermal runaway prediction method based on cross-feature fusion according to claim 1, wherein, The working condition characteristics include overvoltage charging characteristics, ultra-low temperature charging characteristics, parking time characteristics, fast charging characteristics and deep discharge characteristics, wherein the overvoltage charging characteristics refer to relevant statistical quantities when the battery charging voltage exceeds the safety threshold, and the ultra-low temperature charging characteristics refer to relevant statistical quantities when the battery charging environment temperature is lower than the preset temperature threshold. 3.The battery thermal runaway prediction method based on cross-feature fusion according to claim 2, characterized in that, The overvoltage charging characteristics include overvoltage charging times, overvoltage charging cumulative time and average overvoltage amplitude. The ultra-low temperature charging characteristics include ultra-low temperature charging times, ultra-low temperature charging average temperature and ultra-low temperature charging average current. The parking time characteristics include parking time and long-time parking times. The fast charging characteristics include fast charging times, fast charging average current and fast charging temperature rise rate. The deep discharge feature includes a deep discharge frequency, a deep discharge average current, and a SOC drop rate. 4.The battery thermal runaway prediction method based on cross-feature fusion according to claim 1, wherein, The timing feature includes a sliding window statistical feature, a time domain feature, and a frequency domain feature.

5. The battery thermal runaway prediction method based on cross-feature fusion according to claim 4, characterized in that, The sliding window statistical feature includes a voltage one-day mean value, a current one-day variance, and a temperature one-day maximum value. The time domain feature includes a voltage change rate and a current fluctuation amplitude. The frequency domain feature includes a voltage main frequency and a current harmonic component amplitude. 6.The battery thermal runaway prediction method based on cross-feature fusion according to claim 1, wherein, The electrical feature includes a voltage feature, a current feature, an internal resistance feature, and an SOC feature.

7. The battery thermal runaway prediction method based on cross-feature fusion according to claim 6, characterized in that, The voltage feature includes a voltage average value, a voltage variance, and a voltage difference. The current feature includes a current average value, a current peak value, and a current fluctuation amplitude. The internal resistance feature includes an internal resistance change rate and an internal resistance-temperature correlation coefficient. The SOC feature includes an SOC change rate and an SOC-voltage correlation coefficient. 8.The battery thermal runaway prediction method based on cross-feature fusion according to claim 1, wherein, The segmentation feature includes an SOC interval-based segmentation feature, a temperature interval-based segmentation feature, a current interval-based segmentation feature, a time interval-based segmentation feature, and a working condition interval-based segmentation feature.

9. The battery thermal runaway prediction method based on cross-feature fusion according to claim 8, characterized in that, The SOC interval-based segmentation feature includes an SOC interval voltage feature and an SOC interval internal resistance feature, wherein the SOC interval voltage feature is a mean value, a variance, a maximum value, a minimum value, and a change rate of the voltage in an SOC∈[30%, 90%] interval, and the SOC interval internal resistance feature is a mean value and a change rate of the internal resistance in an SOC∈[10%, 20%] interval. The temperature interval-based segmentation feature includes a high-temperature interval internal resistance feature and a low-temperature interval charging efficiency, wherein the high-temperature interval internal resistance feature is a mean value, a change rate, and a fluctuation amplitude of the internal resistance in a temperature>40℃ interval, and the low-temperature interval charging efficiency is an hourly growth rate of the SOC in a temperature<0℃ interval. The current interval-based segmentation feature includes a fast charging interval temperature feature and a discharge interval voltage feature, wherein the fast charging interval temperature feature is a mean value, a rise rate, and a maximum value of the temperature in a current>2C interval, and the discharge interval voltage feature is a drop rate of the voltage in a current<-1C interval. The time interval-based segmentation feature includes a parking time interval SOC feature and a long-time standing internal resistance feature, wherein the parking time interval SOC feature is a drop rate of the SOC in a parking time>12 hours interval, and the long-time standing internal resistance feature is a change rate of the internal resistance in a standing time>24 hours interval. The working condition interval-based segmentation feature includes an overvoltage charging interval temperature feature and a deep discharge interval voltage feature, wherein the overvoltage charging interval temperature feature is a rise rate and a fluctuation amplitude of the temperature in a voltage>4.2V interval, and the deep discharge interval voltage feature is a fluctuation amplitude of the voltage in an SOC<10% interval. 10.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-9. The processor executes the computer program to implement the steps of the battery thermal runaway prediction method based on cross-feature fusion according to any one of claims 1 to 9.

Images