Battery thermal runaway and safety threshold dynamic adjustment system and method

Abstract

The invention discloses a battery thermal runaway and safety threshold dynamic adjustment system, which comprises a data acquisition module used for acquiring battery state parameters, environmental parameters and vehicle use conditions; the data preprocessing module is used for preprocessing the collected data; the initial safety threshold setting module is used for calculating an initial risk threshold of the thermal runaway of the battery by utilizing a statistical rule in combination with a large amount of collected historical sample data, and taking the initial risk threshold as a judgment standard of an initial deployment stage; the dynamic threshold adjustment module is used for performing dynamic personalized adjustment on the initial risk threshold according to a Bayesian optimization algorithm; and the thermal runaway prediction model module is used for predicting the battery risk level in the future 24 hours and the battery temperature in the future 1 hour in combination with a traditional machine learning and deep learning model. The invention also discloses a battery thermal runaway and safety threshold dynamic adjustment method. According to the invention, an early warning mechanism with high accuracy and high recall rate is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy vehicle battery management, and specifically relates to a battery thermal runaway and safety threshold dynamic adjustment system and method. Background Art

[0002] In the application of new energy vehicles, the safety of power batteries has always been a key factor affecting the stability of the whole vehicle and the safety of users. Especially the phenomenon of battery thermal runaway may cause serious consequences such as combustion and explosion.

[0003] Existing methods usually use a fixed threshold setting method for thermal runaway monitoring, but they cannot cope with the non-linear change trends of batteries under complex conditions such as aging, operating conditions, and environmental changes, and are prone to false alarms or missed alarms, affecting the reliability of the early warning system.

[0004] Therefore, in view of the above problems, a battery thermal runaway and safety threshold dynamic adjustment system and method are provided. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a battery thermal runaway and safety threshold dynamic adjustment system and method, realizing an early warning mechanism with high accuracy and high recall rate.

[0006] To achieve the above object, the present invention provides the following technical solutions: A battery thermal runaway and safety threshold dynamic adjustment system according to one of the present inventions, comprising: A data acquisition module for collecting battery state parameters, environmental parameters, and vehicle operating conditions; A data preprocessing module for performing preprocessing operations on the collected data; An initial safety threshold setting module for calculating the initial risk threshold of battery thermal runaway by using statistical rules in combination with a large number of collected historical sample data as the judgment standard in the initial deployment stage; A dynamic threshold adjustment module for dynamically and individually adjusting the initial risk threshold according to the Bayesian optimization algorithm; A thermal runaway prediction model module for predicting the battery risk level in the next 24 hours and the battery temperature in the next 1 hour by combining traditional machine learning and deep learning models; An early warning strategy module for performing hierarchical early warning responses according to the prediction results and threshold judgments.

[0007] Preferably, the data acquisition module is deployed in the vehicle BMS (Battery Management System) system or a dedicated acquisition unit. The battery state parameters include cell / pack temperature, cell / pack voltage, charge / discharge current, SOC (State of Charge), SOH (State of Health of the battery), and the internal resistance of the battery; the environmental parameters include the temperature, humidity, air pressure, wind speed, road conditions, and traffic congestion of the external environment; the vehicle operating conditions include driving speed, acceleration, driving mode, charge / discharge power, and rate of change of battery power.

[0008] Preferably, in the data preprocessing module, the preprocessing operations include: Outlier detection and denoising, including: detection of over-limit temperature change rate, voltage or current drift, and SOC mutation. Among them, Over-limit temperature change rate: Calculate the change rate of temperature per second respectively, define a threshold, and the defined threshold is that the temperature change rate exceeding 2 degrees per second is abnormal data; Voltage or current drift: In the charging state or driving state, compare the current value with the data of the previous N frames. If the difference exceeds a certain threshold, it is considered abnormal data; SOC mutation detection: In the charging or driving state, the SOC changes continuously. If there may be a continuous frame change exceeding a certain threshold, it will be considered an abnormal situation.

[0009] Preferably, after outlier detection and denoising, data alignment and uniform sampling are performed: Since the sampling frequencies of each sensor are different, linear interpolation and spline resampling are used to unify it to one frame every 5 minutes, and a 24-hour historical sliding window is constructed for modeling; among them, The sampling frequency of the battery is one data per second; The sampling frequencies of external temperature, humidity, and wind speed are one data per 10 minutes; The sampling frequencies of speed, acceleration, and charge / discharge power are one data per second; For the data of one per second, only the data at the end of 5 and 10 minutes of the time point are taken. If there is no data at the time point, the data at the closest time point is taken; For the data of one per 10 minutes, linear interpolation is used for temperature and humidity, and the formula is as follows: ; In the formula, and respectively represent two data separated by 10 minutes, and respectively represent two time points; Normalization and standardization: All features are uniformly Min-Max normalized. Among them, for each feature column, independent calculations are performed, and The same feature ensures that the same normalization parameters are used in the training, validation, and inference phases. The original calculation formula is as follows: ; In the formula, and respectively represent the minimum and maximum values of this dimension.

[0010] Preferably, in the initial safety threshold setting module, the initial risk thresholds include: temperature threshold, voltage threshold, and SOC threshold. Among them, the calculation formulas are as follows: ; ; ; In the formula, , and are the temperature threshold, voltage threshold, and SOC threshold respectively, is the average value of the temperature, is the standard deviation of the temperature, is the value of the 99% quantile, is the value of the 98% quantile.

[0011] Preferably, in the dynamic threshold adjustment module, the dynamic personalized adjustment of the initial risk threshold includes: Objective function design: ; In the formula, is the set of threshold parameters, is the false alarm rate, is the miss rate, and are both weights; Optimization process: Using the Gaussian process GP as the surrogate model, using the historical data collected in the past month as the training set, and using the expected improvement sampling strategy to find the optimal threshold. The parameter space is the voltage difference threshold, temperature change rate threshold, and SOC fluctuation threshold. The specific process is as follows: Initialize the sampling points and sample the parameter space by random sampling; Train the GP model to fit to the objective function; Based on the expected improvement sampling strategy, find the next set of ; Evaluate the actual FP / FN and record the objective function; Update GP, repeat the above steps, and iterate until the maximum number of training rounds of 100 times is reached to complete the optimization; Personalized settings: Multiple sets of differentiated thresholds are generated according to vehicle ID, battery supplier, battery pack number, and climate zone conditions.

[0012] Preferably, in the thermal runaway prediction model module, the traditional machine learning model is an XGBoost (ensemble machine learning algorithm based on decision trees) model for static risk assessment. Based on the input temperature mean, maximum / minimum values, 24-hour voltage change rate, SOC stability, maximum temperature difference, SOC standard deviation, charge and discharge times, current standard deviation, temperature difference mean, temperature difference standard deviation, pressure difference mean, and pressure difference standard deviation, high, medium, and low three-level labels are obtained. Among them, high risk means severe fluctuations and overheating, medium risk means slight anomalies or trends towards risks, and low risk means stable and safe operation. The deep learning model is an LSTM (recurrent neural network) model for trend prediction. Based on the input serialized 24-hour data, the temperature prediction for the next 1 hour is obtained. If the future value approaches the current dynamic threshold, a secondary warning is triggered in advance.

[0013] Preferably, in the warning strategy module, the hierarchical warning response includes: Primary warning: The current value breaks through the threshold. Secondary warning: The LSTM predicted trend is about to break through. Tertiary warning: XGBoost determines it as a high risk. Combined trigger: When multiple warnings are met simultaneously, the expert system is linked to recommend operations, including battery swapping, cooling, and diagnosis.

[0014] A method for dynamically adjusting battery thermal runaway and safety thresholds according to the second aspect of the present invention includes: Step S1, collect battery state parameters, environmental parameters, and vehicle usage conditions. Step S2, perform preprocessing operations on the collected data. Step S3, combine a large number of collected historical sample data and use statistical rules to calculate the initial risk threshold of battery thermal runaway as the judgment standard in the initial deployment stage. Step S4, perform dynamic personalized adjustment on the initial risk threshold according to the Bayesian optimization algorithm. Step S5, combine traditional machine learning and deep learning models to predict the battery risk level for the next 24 hours and the battery temperature for the next 1 hour. Step S6, perform hierarchical warning response according to the prediction results and threshold judgment.

[0015] Preferably, in the step S1, the battery state parameters include the temperature of single cell / pack, the voltage of single cell / pack, the charge and discharge current, SOC, SOH, and the internal resistance of the battery; the environmental parameters include the temperature, humidity, air pressure, wind speed, road conditions, and traffic congestion of the external environment; the vehicle usage conditions include the driving speed, acceleration, driving mode, charge and discharge power, and the rate of change of the power.

[0016] Preferably, in the step S2, the preprocessing operations include: Outlier detection and denoising, including: detecting the overlimit of the temperature change rate, voltage or current drift, and SOC mutation. Among them, Overlimit of the temperature change rate: Calculate the change rate of the temperature per second respectively, define a threshold, and the defined threshold is that the data with the temperature change rate exceeding 2 degrees / second is abnormal data; Voltage or current drift: In the charging state or driving state, compare the current value with the data of the previous N frames. If the difference exceeds a certain threshold, it is considered abnormal data; SOC mutation detection: In the charging or driving state, the SOC changes continuously. If there may be a change exceeding a certain threshold in consecutive frames, it will be considered an abnormal situation.

[0017] Preferably, after the outlier detection and denoising, data alignment and unified sampling are performed: Since the sampling frequencies of the sensors are different, linear interpolation and spline resampling are used to unify them into one frame every 5 minutes, and a 24-hour historical sliding window is constructed for modeling. Among them, The sampling frequency of the battery is one data per second; The sampling frequencies of the external temperature, humidity, and wind speed are one data per 10 minutes; The sampling frequencies of the speed, acceleration, and charge and discharge power are one data per second; For the data of one per second, only the data at the end of 5 and 10 minutes of the time point are taken. If there is no data at the time point, the data at the closest time point is taken; For the data of one per 10 minutes, linear interpolation is used for the temperature and humidity, and the formula is as follows: ; In the formula, and represent two data separated by 10 minutes respectively, and represent two time points respectively; Normalization and standardization: All features are uniformly normalized by Min - Max. Among them, for each feature column, independent calculation is performed, and The same feature ensures the use of the same normalization parameters in the training, validation, and inference stages. The original calculation formula is as follows: ; In the formula, and respectively represent the minimum value and the maximum value of this dimension.

[0018] Preferably, in the step S3, the initial risk threshold includes: temperature threshold, voltage threshold and SOC threshold, and the calculation formulas are as follows respectively: ; ; ; In the formula, , and are the temperature threshold, voltage threshold and SOC threshold respectively, is the average value of temperature, is the standard deviation of temperature, is the value of the 99% quantile, is the value of the 98% quantile.

[0019] Preferably, in the step S4, the dynamic personalized adjustment of the initial risk threshold includes: Objective function design: ; In the formula, is the set of threshold parameters, is the false alarm rate, is the miss rate, and are both weights; Optimization process: Using the Gaussian process GP as the surrogate model, using the historical data collected in the past month as the training set, and using the expected improvement sampling strategy to find the optimal threshold. The parameter space is the voltage difference threshold, temperature change rate threshold, and SOC fluctuation threshold. The specific process is as follows: Initialize the sampling points and sample the parameter space by random sampling; Train the GP model to fit to the objective function; Based on the expected improvement sampling strategy, find the next set of ; Evaluate the actual FP / FN and record the objective function; Update GP, repeat the above steps, and iterate until the maximum number of training rounds of 100 times to complete the optimization; Personalized setting: Generate multiple sets of differentiated thresholds according to the vehicle ID, battery supplier, battery pack number, and climate zone conditions.

[0020] Preferably, in step S5, the traditional machine learning model is the XGBoost model, which is used for static risk assessment. Based on the input temperature mean, maximum / minimum value, 24-hour voltage change rate, SOC stability, maximum temperature difference, SOC standard deviation, charge and discharge times, current standard deviation, temperature difference mean, temperature difference standard deviation, pressure difference mean, and pressure difference standard deviation, high, medium, and low level labels are obtained. Among them, high risk means severe fluctuations and overheating, medium risk means slight anomalies or a trend towards risk, and low risk means stable and safe operation. The deep learning model is the LSTM model, which is used for trend prediction. Based on the input serialized 24-hour data, the temperature prediction for the next hour is obtained. If the future value approaches the current dynamic threshold, a secondary warning is triggered in advance.

[0021] Preferably, in step S6, the hierarchical warning response includes: Primary warning: The current value breaks through the threshold. Secondary warning: The LSTM predicted trend is about to break through. Tertiary warning: The XGBoost judges it as a high risk. Combined trigger: When multiple warnings are satisfied simultaneously, the expert system is linked to recommend operations, including battery swapping, cooling, and diagnosis.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: Through data collection, feature construction, intelligent prediction, and dynamic optimization, the present invention realizes the early judgment of the risk of battery thermal runaway and personalized warning, improving the vehicle safety and preventing major battery accidents. Description of the Drawings

[0023] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is a module diagram of a battery thermal runaway and safety threshold dynamic adjustment system of the present invention; Figure 2 is a flowchart of a battery thermal runaway and safety threshold dynamic adjustment method of the present invention. Detailed Embodiments

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0025] Such as Figure 1As shown in the figure, a system for dynamically adjusting the thermal runaway and safety threshold of a battery includes: a data acquisition module 1, a data preprocessing module 2, an initial safety threshold setting module 3, a dynamic threshold adjustment module 4, a thermal runaway prediction model module 5, and an early warning strategy module 6.

[0026] The data acquisition module 1 is used to collect battery state parameters, environmental parameters, and vehicle usage conditions.

[0027] In the embodiment, the data acquisition module 1 is deployed in the vehicle BMS system or a dedicated acquisition unit. The battery state parameters include single cell / pack temperature, single cell / pack voltage, charge and discharge current, SOC, SOH, and the internal resistance of the battery; the environmental parameters include the temperature, humidity, air pressure, wind speed, road conditions, and traffic congestion of the external environment; the vehicle usage conditions include driving speed, acceleration, driving mode, charge and discharge power, and the rate of change of battery power.

[0028] The data preprocessing module 2 is used to perform preprocessing operations on the collected data.

[0029] In the embodiment, the preprocessing operations include: Outlier detection and denoising, including: detecting over-limit temperature change rate, voltage or current drift, and SOC mutation. Among them, Over-limit temperature change rate: Calculate the change rate of temperature per second respectively, define a threshold, and the defined threshold is that the temperature change rate exceeding 2 degrees / second is abnormal data; Voltage or current drift: In the charging state or driving state, compare the current value with the data of the previous N frames. If the difference exceeds a certain threshold, it is considered abnormal data. In the present invention, it is defined to compare with the data of the previous 10 frames, and the difference exceeds 15% of the previous 10 frames; SOC mutation detection: Whether in the charging or driving state, the SOC changes continuously. If there may be a continuous frame change exceeding a certain threshold, it will be considered an abnormal situation. In the present invention, the amplitude of the continuous change frame is defined as not exceeding 2% in the charging state and not exceeding 1% in the driving state; After outlier detection and denoising, perform data alignment and uniform sampling: Since the sampling frequencies of each sensor are different, use linear interpolation plus spline resampling to unify to one frame every 5 minutes, and construct a 24-hour historical sliding window for modeling; among them, The sampling frequency of the battery is one data per second; The sampling frequencies of external temperature, humidity, wind speed and other data are one data per 10 minutes; The sampling frequencies of speed, acceleration, and charge and discharge power are one data per second; For the data of one data per second, only take the data at the end of 5 and 10 minutes of the time point. If there is no data at the time point, take the data at the closest time point; For data at 10 - minute intervals, linear interpolation is used for temperature and humidity, and the formula is as follows: ; In the formula, and represent two pieces of data separated by 10 minutes respectively, and represent two time points respectively. For data such as wind speed, they are processed in the same way that they are all the same within this 10 - minute period; Normalization and standardization: All features are uniformly subjected to Min - Max normalization. Among them, for each feature column, independent calculations are performed, and For the same feature, the same normalization parameters are ensured to be used in the training, validation, and inference stages. The original calculation formula is as follows: ; In the formula, and represent the minimum value and the maximum value of this dimension respectively, and the data of all dimensions are uniformly processed to be within the range of [0, 1].

[0030] The initial safety threshold setting module 3 is used to combine a large amount of collected historical sample data and calculate the initial risk threshold of battery thermal runaway using statistical rules as the judgment standard in the initial deployment stage.

[0031] In the embodiment, the initial risk threshold includes: temperature threshold, voltage threshold, and SOC threshold. Among them, the calculation formulas are as follows respectively: ; ; ; In the formula, , and are the temperature threshold, voltage threshold, and SOC threshold respectively, is the average value of temperature, is the standard deviation of temperature, is the value of the 99% quantile, is the value of the 98% quantile.

[0032] The dynamic threshold adjustment module 4 is used to perform dynamic personalized adjustment on the initial risk threshold according to the Bayesian optimization algorithm.

[0033] In the embodiment, the dynamic personalized adjustment of the initial risk threshold includes: Objective function design: ; In the formula, is a set of threshold parameters, is the false positive rate, is the false negative rate, and are both weights, both equal to 1, but adjustable; Optimization process: Using the Gaussian process GP as the surrogate model, using the collected historical data for 1 month as the training set, and using the expected improvement sampling strategy to find the optimal threshold. The parameter space is the voltage difference threshold, the temperature change rate threshold, and the SOC fluctuation threshold. The specific process is as follows: Initialize the sampling points, and sample the parameter space in a random sampling manner; Train the GP model to fit to the objective function; Based on the expected improvement sampling strategy, find the next set of ; Evaluate the actual FP / FN and record the objective function; Update the GP, repeat the above steps, iterate until reaching the maximum number of training rounds of 100 times, and complete the optimization; Personalized setting: Generate multiple sets of differentiated thresholds according to the vehicle ID, battery supplier, battery pack number, and climate zone conditions.

[0034] In the embodiment, the biggest reason for this step is that even for the same model of vehicle, different configurations and models correspond to different batteries, and under different environments, the dynamic thresholds will also be different. Therefore, it is necessary to consider that in different dimensions, each combination trains an independent threshold model.

[0035] The thermal runaway prediction model module 5 is used to combine traditional machine learning and deep learning models to predict the battery risk level for the next 24 hours and the battery temperature for the next 1 hour.

[0036] In the embodiment, the traditional machine learning model is the XGBoost model, which is used for static risk assessment. According to the input temperature mean, maximum / minimum value, 24h voltage change rate, SOC stability, maximum temperature difference, SOC standard deviation, charge and discharge times, current standard deviation, temperature difference mean, temperature difference standard deviation, pressure difference mean, pressure difference standard deviation, high, medium, and low three-level labels are obtained. Among them, high risk is severe fluctuation and overheating, medium risk is slight abnormality or trend towards risk, and low risk is stable and safe operation; In the historical record, starting the overheat protection or the data where thermal runaway occurs is defined as high risk; in the historical record, the data exceeding the threshold defined in the fourth step is defined as medium risk, and other data is defined as low risk; in this way, there are labels and input features, and then the present invention uses the XGBoost model for training, and the training process is the conventional steps of machine learning; The deep learning model is an LSTM model for trend prediction. Based on the input serialized 24-hour data, it obtains the temperature prediction for the next 1 hour. If the future value approaches the current dynamic threshold, a secondary warning is triggered in advance. For the model trained using LSTM, according to the serialized data obtained in the second step, with one data every five minutes, the label formed by combining the first 12 data is the temperature of the 24th data, the label formed by combining the 2nd to 13th data is the temperature of the 25th data, and so on; the feature of each combination corresponds to the temperature label one hour later; in this way, a time series of feature data can be obtained; by training with the LSTM time series model, the temperature one hour later can be predicted based on the features of this one hour; if any point in the prediction exceeds the dynamic threshold, a secondary warning will be triggered.

[0037] The warning strategy module 6 is used to perform hierarchical warning responses according to the prediction results and threshold judgments.

[0038] In the embodiment, the hierarchical warning responses include: Primary warning: The current value breaks through the threshold. Secondary warning: The LSTM prediction trend is about to break through. Tertiary warning: XGBoost determines it to be a high risk. Combined trigger: When multiple warnings are satisfied simultaneously, the expert system is linked to recommend operations, including battery swapping, cooling, and diagnosis.

[0039] As Figure 2 shown, a method for dynamically adjusting the battery thermal runaway and safety threshold includes: Step S1, collect battery state parameters, environmental parameters, and vehicle usage conditions.

[0040] In the embodiment, the battery state parameters include the temperature of single cells / pack, the voltage of single cells / pack, charge and discharge current, SOC, SOH, and the internal resistance of the battery; the environmental parameters include the temperature, humidity, air pressure, wind speed, road conditions, and traffic congestion conditions of the external environment; the vehicle usage conditions include driving speed, acceleration, driving mode, charge and discharge power, and the rate of change of the battery level.

[0041] Step S2, perform preprocessing operations on the collected data.

[0042] In the embodiment, the preprocessing operations include: Outlier detection and denoising, including: detection of over-limit temperature change rate, voltage or current drift, and SOC mutation. Among them, Over-limit temperature change rate: Calculate the change rate of temperature per second respectively, define a threshold, and the defined threshold is that the temperature change rate exceeding 2 degrees / second is abnormal data. Voltage or current drift: During the charging state or driving state, compare the current value with the data of the previous N frames. If the difference exceeds a certain threshold, it is considered abnormal data; SOC mutation detection: During the charging or driving state, the SOC changes continuously. If there may be a change exceeding a certain threshold in consecutive frames, it will be considered an abnormal situation; After detecting and denoising abnormal values, perform data alignment and unified sampling: Since the sampling frequencies of each sensor are different, linear interpolation and spline resampling are used to unify it to one frame every 5 minutes, and a 24-hour historical sliding window is constructed for modeling; among them, The sampling frequency of the battery is one data per second; The sampling frequencies of external temperature, humidity, wind speed and other data are one data per 10 minutes; The sampling frequencies of speed, acceleration, and charge-discharge power are one data per second; For the data of one per second, only the data at the end of 5 and 10 minutes of the time point are taken. If there is no data at the time point, the data at the closest time point is taken; For the data of one per 10 minutes, linear interpolation is used for temperature and humidity, and the formula is as follows: ; In the formula, and represent two pieces of data 10 minutes apart respectively, and represent two time points respectively; Normalization and standardization: All features are uniformly normalized by Min-Max, where each feature column is calculated independently, and The same feature ensures the use of the same normalization parameters in the training, validation, and inference stages. The original calculation formula is as follows: ; In the formula, and represent the minimum value and the maximum value of this dimension respectively.

[0043] Step S3, combined with a large number of collected historical sample data, use statistical rules to calculate the initial risk threshold of battery thermal runaway as the judgment standard in the initial deployment stage.

[0044] In the embodiment, the initial risk threshold includes: temperature threshold, voltage threshold and SOC threshold, where the calculation formulas are as follows respectively: ; ; ; In the formula, , and are the temperature threshold, voltage threshold, and SOC threshold respectively. is the average value of the temperature. is the standard deviation of the temperature. is the value of the 99% quantile. is the value of the 98% quantile.

[0045] Step S4: Dynamically and individually adjust the initial risk threshold according to the Bayesian optimization algorithm.

[0046] In the embodiment, dynamically and individually adjusting the initial risk threshold includes: Objective function design: ; In the formula, is the set of threshold parameters, is the false alarm rate, is the missed alarm rate, and are both weights; Optimization process: Using the Gaussian process GP as the surrogate model, using the historical data collected in the past month as the training set, and using the expected improvement sampling strategy to find the optimal threshold. The parameter space is the voltage difference threshold, temperature change rate threshold, and SOC fluctuation threshold. The specific process is as follows: Initialize the sampling points, and sample the parameter space by random sampling; Train the GP model to fit to the objective function; Based on the expected improvement sampling strategy, find the next set of ; Evaluate the actual FP / FN and record the objective function; Update the GP, repeat the above steps, and iterate until the maximum number of training rounds of 100 times to complete the optimization; Individualized setting: Generate multiple sets of differentiated thresholds according to the vehicle ID, battery supplier, battery pack number, and climate zone conditions.

[0047] Step S5: Combine traditional machine learning and deep learning models to predict the battery risk level in the next 24 hours and the battery temperature in the next 1 hour.

[0048] In the embodiment, the traditional machine learning model is the XGBoost model, which is used for static risk assessment. According to the input temperature mean, maximum / minimum value, 24h voltage change rate, SOC stability, maximum temperature difference, SOC standard deviation, charge and discharge times, current standard deviation, temperature difference mean, temperature difference standard deviation, pressure difference mean, and pressure difference standard deviation, high, medium, and low level labels are obtained. Among them, high risk means severe fluctuation and overheating, medium risk means slight abnormality or trend towards risk, and low risk means stable and safe operation. The deep learning model is the LSTM model, which is used for trend prediction. According to the input serialized 24-hour data, the temperature prediction for the next 1 hour is obtained. If the future value approaches the current dynamic threshold, a secondary warning is triggered in advance.

[0049] Step S6, according to the prediction result and threshold judgment, a hierarchical warning response is carried out.

[0050] In the embodiment, the hierarchical warning response includes: Primary warning: The current value breaks through the threshold. Secondary warning: The LSTM predicted trend is about to break through. Tertiary warning: The XGBoost judges it as a high risk. Combined trigger: When multiple warnings are satisfied simultaneously, the expert system is linked to recommend operations, including battery swapping, cooling, and diagnosis.

[0051] The technical details of the present invention will be described below through several embodiments. Embodiment 1: Overall structure of the battery thermal runaway prediction system This embodiment proposes a terminal-cloud collaborative battery thermal runaway risk prediction system, which mainly consists of the following modules: (1) Data acquisition module (terminal side) Installed in the BMS or an independent data acquisition unit.

[0052] Realtime collect battery data (such as temperature, voltage, current, SOC), vehicle status (vehicle speed, power), and environmental information (external temperature, humidity).

[0053] The data is uploaded to the cloud using the MQTT / HTTP protocol, and the period is configurable (default 1s).

[0054] (2) Data preprocessing module (cloud) Clean the original data, including: Anomaly detection: such as temperature jump (ΔT>30°C / s), voltage anomaly (V<2.5V or V>4.2V); Filtering and noise reduction: using moving average, Kalman filter; Time alignment: linearly interpolate data with different frequencies; Window processing: Extract historical data within 24 hours as input features.

[0055] (3)Initial safety threshold setting module Use historical data statistical features to calculate the initial threshold for thermal runaway determination: The initial risk thresholds include: temperature threshold, voltage threshold, and SOC threshold. Among them, the calculation formulas are as follows: ; ; ; (4)Dynamic threshold adjustment module Based on the Bayesian optimization algorithm: Construct the false alarm rate / missed alarm rate as the objective function; Use Gaussian process to model the current data distribution; Retrain the optimal threshold every 30 days; Support personalized settings (fine-tuning for dimensions such as vehicle models and battery pack numbers).

[0056] (5)Thermal runaway prediction model module XGBoost: The training data is the statistical features (maximum value, minimum value, mean value, slope, etc.) within a 24-hour window; Output: High, medium, and low risk levels.

[0057] LSTM: Input: Serialized historical temperature, voltage, and SOC data; Output: Prediction of future 1-hour temperature or voltage; If the predicted value approaches the threshold → Early warning.

[0058] (6)Early warning strategy module The triggering rules are as follows: The current value exceeds the dynamic threshold → Level 1 early warning; The LSTM prediction trend exceeds the threshold → Level 2 early warning; XGBoost determines high risk → High-level risk warning; Information can be fed back through the in-vehicle computer / APP / BMS panel; Link to the expert system and recommend maintenance measures (such as coolant inspection, battery pack replacement, etc.).

[0059] Example 2: Threshold dynamic adjustment process (Bayesian optimization) 1. Define the objective function design: ; In the formula, is a set of threshold parameters, is the false positive rate, is the false negative rate, and are both weights; 2. Collect the prediction results of the real alarm data model; 3. Statistically calculate FP and FN every day and fit the Gaussian process; 4. Use the EI (Expected Improvement) strategy to select the next threshold to be evaluated; 5. Perform multiple rounds of iteration and determine the optimal threshold after convergence.

[0060] Embodiment 3: Model training process 1. Data collection: Collect 12 months of data from 3000 vehicles; 2. Training set division: 90% of the data is used for training and 10% for testing; 3. LSTM configuration: Input: Data for the past 24 hours; Number of layers: 2; Hidden layer dimension: 128; Time step: 288 (once every 5 minutes, for a total of 24h); 4. XGBoost configuration: Maximum depth: 6; Learning rate: 0.1; Number of trees: 200; Input features: Mean value, variance, maximum change rate, etc.; 5. Integration logic: If the LSTM predicts that the future data approaches the threshold and XGBoost determines it to be a high risk, then trigger the highest level of warning.

[0061] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A battery thermal runaway and safety threshold dynamic adjustment system, characterized in that, Including: A data acquisition module, which is used to acquire battery state parameters, environmental parameters, and vehicle usage conditions; A data preprocessing module, which is used to perform preprocessing operations on the acquired data; An initial safety threshold setting module, which is used to calculate the initial risk threshold of battery thermal runaway by using statistical rules in combination with a large number of acquired historical sample data, and use it as the judgment standard in the initial deployment stage; A dynamic threshold adjustment module, which is used to dynamically and personalized adjust the initial risk threshold according to the Bayesian optimization algorithm; A thermal runaway prediction model module, which is used to predict the battery risk level in the next 24 hours and the battery temperature in the next 1 hour by combining traditional machine learning and deep learning models; An early warning strategy module, which is used to perform hierarchical early warning responses according to the prediction results and threshold judgments.

2. The dynamic adjustment system for battery thermal runaway and safety threshold according to claim 1, characterized in that The data acquisition module is deployed in the vehicle BMS system or a dedicated acquisition unit. The battery state parameters include single cell / pack temperature, single cell / pack voltage, charge and discharge current, SOC, SOH, and the internal resistance of the battery; the environmental parameters include the temperature, humidity, air pressure, wind speed, road conditions, and traffic congestion conditions of the external environment; the vehicle usage conditions include driving speed, acceleration, driving mode, charge and discharge power, and the rate of change of battery power.

3. A battery thermal runaway and safety threshold dynamic adjustment system according to claim 1, characterized in that In the data preprocessing module, the preprocessing operations include: Outlier detection and denoising, including: detection of over-limit temperature change rate, voltage or current drift, and SOC mutation. Among them, Over-limit temperature change rate: Calculate the change rate of temperature per second respectively, define a threshold, and the defined threshold is that the temperature change rate exceeds 2 degrees / second, which is regarded as abnormal data; Voltage or current drift: In the charging state or driving state, compare the current value with the data of the previous N frames. If the difference exceeds a certain threshold, it is considered abnormal data; SOC mutation detection: In the charging or driving state, the SOC changes continuously. If there may be a continuous frame change exceeding a certain threshold, it will be regarded as an abnormal situation.

4. A battery thermal runaway and safety threshold dynamic adjustment system according to claim 3, characterized in that After outlier detection and denoising, data alignment and unified sampling are performed: Since the sampling frequencies of each sensor are different, linear interpolation and spline resampling are used to unify it into one frame every 5 minutes, and a 24-hour historical sliding window is constructed for modeling; among them, The sampling frequency of the battery is one data per second; The sampling frequencies of external temperature, humidity, and wind speed are one data per 10 minutes; The sampling frequencies of speed, acceleration, and charge and discharge power are one data per second; For the data of one data per second, only the data at the end of 5 and 10 minutes of the time point are taken. If there is no data at the time point, the data at the closest time point is taken; For the data of one data per 10 minutes, linear interpolation is used for temperature and humidity. The formula is as follows: ； In the formula, and respectively represent two pieces of data separated by 10 minutes, and respectively represent two time points; Normalization and Standardization: All features are uniformly subjected to Min-Max normalization. Among them, each feature column is calculated independently, and For the same feature, ensure that the same normalization parameters are used in the training, validation, and inference stages. The original calculation formula is as follows: ； In the formula, and respectively represent the minimum value and the maximum value of this dimension.

5. A battery thermal runaway and safety threshold dynamic adjustment system according to claim 1, characterized in that, In the initial safety threshold setting module, the initial risk thresholds include: temperature threshold, voltage threshold, and SOC threshold. Among them, the calculation formulas are as follows: ； ； ； Wherein, , and are the temperature threshold, voltage threshold and SOC threshold respectively, is the average value of the temperature, is the standard deviation of the temperature, is the value of the 99% quantile, is the value of the 98% quantile.

6. The dynamic adjustment system for battery thermal runaway and safety threshold according to claim 1, characterized in that, In the dynamic threshold adjustment module, the dynamic and personalized adjustment of the initial risk threshold includes: Objective function design: ； In the formula, is a set of threshold parameters, is the false alarm rate, is the miss rate, and are both weights; Optimization process: Using the Gaussian process GP as the surrogate model, the historical data collected in the past month is used as the training set, and the expected improvement sampling strategy is used to find the optimal threshold. The parameter space is the voltage difference threshold, the temperature change rate threshold, and the SOC fluctuation threshold. The specific process is as follows: Initialize the sampling points and sample the parameter space by random sampling; Train the GP model to fit to the objective function; Based on the expected improvement sampling strategy, find the next set of ; Evaluate the actual FP / FN and record the objective function; Update the GP, repeat the above steps, and iterate up to the maximum number of training rounds of 100 times to complete the optimization; Personalized setting: Generate multiple groups of differentiated thresholds according to the vehicle ID, battery supplier, battery pack number, and climate zone conditions.

7. A battery thermal runaway and safety threshold dynamic adjustment system according to claim 1, characterized in that In the thermal runaway prediction model module, the traditional machine learning model is the XGBoost model, which is used for static risk assessment. According to the input temperature mean, maximum / minimum value, 24h voltage change rate, SOC stability, maximum temperature difference, SOC standard deviation, charge and discharge times, current standard deviation, temperature difference mean, temperature difference standard deviation, pressure difference mean, and pressure difference standard deviation, high, medium, and low three-level labels are obtained. Among them, high risk means severe fluctuations and overheating, medium risk means slight anomalies or trends towards risk, and low risk means stable and safe operation; The deep learning model is the LSTM model, which is used for trend prediction. According to the input serialized 24-hour data, the temperature prediction for the next 1 hour is obtained. If the future value approaches the current dynamic threshold, a secondary warning is triggered in advance.

8. A battery thermal runaway and safety threshold dynamic adjustment system according to claim 1, characterized in that, In the warning strategy module, the hierarchical warning response includes: Primary warning: The current value breaks through the threshold; Secondary warning: The LSTM predicted trend is about to break through; Tertiary warning: XGBoost judges as high risk; Combined trigger: When multiple warnings are satisfied at the same time, the expert system is linked to recommend operations, including battery swapping, cooling, and diagnosis.

9. A method for dynamically adjusting the battery thermal runaway and safety threshold, characterized in that, Including: Step S1, collect battery state parameters, environmental parameters, and vehicle usage conditions; Step S2, perform preprocessing operations on the collected data; Step S3, combine a large number of collected historical sample data and use statistical rules to calculate the initial risk threshold of battery thermal runaway as the judgment standard in the initial deployment stage; Step S4, dynamically adjust the initial risk threshold according to the Bayesian optimization algorithm; Step S5, combine traditional machine learning and deep learning models to predict the battery risk level for the next 24 hours and the battery temperature for the next 1 hour; Step S6, perform hierarchical warning response according to the prediction results and threshold judgment.

10. A method for dynamically adjusting the battery thermal runaway and safety threshold according to claim 9, characterized in that, In the said step S1, the battery state parameters include the temperature of single cell / entire pack, the voltage of single cell / entire pack, the charge and discharge current, SOC, SOH, and the internal resistance of the battery; the environmental parameters include the temperature, humidity, air pressure, wind speed, road conditions, and traffic congestion conditions of the external environment; the vehicle usage conditions include the driving speed, acceleration, driving mode, charge and discharge power, and the rate of change of the battery charge.

11. A method for dynamically adjusting the battery thermal runaway and safety threshold according to claim 9, characterized in that, In the said step S2, the preprocessing operations include: Outlier detection and denoising, including: detection of over-limit temperature change rate, voltage or current drift, and SOC mutation. Among them, Over-limit temperature change rate: Calculate the change rate of temperature per second respectively, define a threshold, and the defined threshold is that the temperature change rate exceeds 2 degrees / second, which is abnormal data; Voltage or current drift: During the charging state or driving state, compare the current value with the data of the previous N frames. If the difference exceeds a certain threshold, it is considered abnormal data. SOC mutation detection: During the charging or driving state, the SOC changes continuously. If there may be a change exceeding a certain threshold in consecutive frames, it will be considered an abnormal situation.

12. A method for dynamically adjusting the battery thermal runaway and safety threshold according to claim 11, characterized in that, In step S2, after detecting and denoising abnormal values, data alignment and uniform sampling are performed: Since the sampling frequencies of each sensor are different, linear interpolation and spline resampling are used to unify to one frame every 5 minutes, and a 24-hour historical sliding window is constructed for modeling; among them, The sampling frequency of the battery is one data per second; The sampling frequencies of external temperature, humidity, and wind speed are one data per 10 minutes; The sampling frequencies of speed, acceleration, and charge and discharge power are one data per second; For the data of one per second, only the data at the end of 5 and 10 minutes of the time point are taken. If there is no data at the time point, the data at the closest time point is taken; For the data of one per 10 minutes, linear interpolation is used for temperature and humidity, and the formula is as follows: ； In the formula, and respectively represent two pieces of data 10 minutes apart, and respectively represent two time points; Normalization and Standardization: All features are uniformly normalized using Min-Max normalization. Specifically, each feature column is calculated independently, and for the same feature, the same original normalization parameters are ensured for use in the training, validation, and inference phases. The calculation formula is as follows: ； In the formula, and respectively represent the minimum value and the maximum value of this dimension.

13. A method for dynamically adjusting the battery thermal runaway and safety threshold according to claim 9, characterized in that, In step S3, the initial risk thresholds include: temperature threshold, voltage threshold, and SOC threshold, and the calculation formulas are as follows respectively: ； ； ； Wherein, , and are the temperature threshold, voltage threshold and SOC threshold respectively, is the average value of the temperature, is the standard deviation of the temperature, is the value of the 99% quantile, is the value of the 98% quantile.

14. A method for dynamically adjusting battery thermal runaway and safety thresholds according to claim 9, characterized in that, In step S4, the dynamic personalized adjustment of the initial risk threshold includes: Objective function design: ； In the formula, is the set of threshold parameters, is the false alarm rate, is the miss rate, and are both weights; Optimization process: Using the Gaussian process GP as a surrogate model, using the collected historical data of one month as the training set, using the expected improvement sampling strategy to find the optimal threshold, and the parameter space is the voltage difference threshold, temperature change rate threshold, and SOC fluctuation threshold. The specific process is as follows: Initialize the sampling points and sample the parameter space by random sampling; Train the GP model to fit to the objective function; Based on the expected improvement sampling strategy, find the next set of ; Evaluate the actual FP / FN and record the objective function; Update the GP, repeat the above steps, and iterate until the maximum number of training rounds of 100 times is reached to complete the optimization; Personalized setting: According to the vehicle ID, battery supplier, battery pack number, and climate zone conditions, multiple groups of differentiated thresholds are generated.

15. A method for dynamically adjusting the battery thermal runaway and safety threshold according to claim 9, characterized in that, In step S5, the traditional machine learning model is the XGBoost model, which is used for static risk assessment. According to the input temperature mean, maximum / minimum value, 24-hour voltage change rate, SOC stability, maximum temperature difference, SOC standard deviation, charge and discharge times, current standard deviation, temperature difference mean, temperature difference standard deviation, pressure difference mean, and pressure difference standard deviation, high, medium, and low three-level labels are obtained. Among them, high risk is severe fluctuation and overheating, medium risk is slight abnormality or trend towards risk, and low risk is stable and safe operation; The deep learning model is the LSTM model, which is used for trend prediction. According to the input serialized 24-hour data, the temperature prediction for the next 1 hour is obtained. If the future value approaches the current dynamic threshold, a secondary warning is triggered in advance.

16. A method for dynamically adjusting the battery thermal runaway and safety threshold according to claim 9, characterized in that, In step S6, the hierarchical warning response includes: Primary warning: The current value breaks through the threshold; Secondary warning: The LSTM predicted trend is about to break through; Tertiary warning: The XGBoost judges as high risk; Combined trigger: When multiple warnings are satisfied at the same time, the expert system is linked to recommend operations, including battery replacement, cooling, and diagnosis.

Images