Virtual-real fusion cloud edge collaborative double-model thermal early warning method for automobile power battery

Abstract

The invention discloses a virtual-real fusion cloud edge collaborative double-model thermal early warning method for an automobile power battery. The method comprises six steps of hardware deployment and software integration, multi-modal data synchronous acquisition, data pre-processing, perception and diagnosis model building, model training and optimization and virtual-real fusion evolution, and thermal state analysis and fault processing. The method has the beneficial effects that double models of real-time judgment and trend pre-judgment are realized, the models can achieve millisecond-level response in case of an emergency fault, the development state of the fault is pre-judged in advance through LSTM time sequence analysis, and the problem of postmortem response is solved. An attention mechanism is introduced into the model, so that the decision-making process is transparent, fault occurrence reasons can be found, adaptive fusion and double-model collaboration are facilitated, and the classification accuracy of complex faults is improved. Bayesian reasoning is introduced into vehicle end SOH real-time estimation, a thermal safety threshold value is dynamically adjusted according to probability, and reference control over safety risk prevention is achieved.

Description

Technical Field

[0001] This invention relates to a fault early warning method, and more particularly to a dual-model thermal early warning method for automotive power batteries that integrates virtual and real data and coordinates cloud and edge computing, belonging to the field of power battery technology. Background Technology

[0002] With the development of new energy vehicles, power batteries, as the core power source of these vehicles, are prone to problems such as uneven temperature distribution and localized heat accumulation during charge-discharge cycles and long-term aging. If these battery problems are not detected in time in new energy vehicles, they may lead to thermal runaway, coolant failure, and cell short circuits, seriously endangering passenger safety.

[0003] Traditional methods cannot adapt to battery health degradation, which may lead to false alarms of battery failure. The lack of uncertainty in battery health estimation results in inaccurate decision-making and makes it difficult to achieve early warning and trend prediction of failure in a timely manner.

[0004] Chinese patent CN120481800A discloses a method, system, and device for monitoring fuel cells in vehicles. This solution analyzes and predicts fuel cell operating parameters, uses a 1D-CNN model to extract features from fuel cell alarm data, and establishes and trains a diagnostic model. It also uses an LSTM model to extract features from driving data, and establishes and trains an early warning model based on the extracted features. Real-time collected fuel cell operating data is input into the above models for analysis, and the fuel cell vehicle equipment is controlled in real-time based on the analysis results. This solution analyzes the overall operating parameters of the fuel cell, but fails to promptly detect and address issues such as uneven internal temperature distribution and localized heat accumulation within the power battery. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to address the problems existing in the prior art by proposing a dual-model thermal early warning method for automotive power batteries that integrates virtual and real data and coordinates cloud and edge computing. This application is based on a dual-model collaborative architecture of "1D CNN model + LSTM", combined with multi-source data acquisition and fusion, and attention mechanism and battery health status module, to build a complete closed-loop system from data acquisition to intelligent decision-making, which can realize adaptive battery aging and significantly improve the safety and reliability of battery system.

[0006] Technical Solution: A dual-model thermal early warning method for automotive power batteries, comprising the following steps:

[0007] Step 1: Hardware deployment and software integration, install sparse sensors and deploy soft sensing algorithms, and set data acquisition conditions;

[0008] Step 2: Synchronous acquisition of multimodal data. The sensor array is triggered to start data sampling according to the preset logic, and the acquired data is transmitted to the preprocessing module.

[0009] Step 3: Data preprocessing. The obtained raw data is preprocessed, including time synchronization, noise reduction, feature extraction, and normalization.

[0010] Step 4: Build a perception and diagnostic model, employing a collaborative approach of vehicle-side edge computing and cloud-based deep learning. A lightweight 1D CNN model is deployed on the vehicle for real-time screening, while an LSTM model and a Bayesian SOH estimation module are deployed in the cloud for deep diagnostics and parameter iteration.

[0011] The data obtained in step 3 is input into a 1D CNN model as a real-time classification model for hot states.

[0012] LSTM models are used as time series analyzers to capture temperature evolution patterns;

[0013] The attention mechanism module is used to couple the output of the LSTM, assign weights to the hidden states at different time steps, focus on the key time segments most relevant to the fault, and output the fault type, development trend and contribution of key features.

[0014] The LSTM outputs a hidden state vector, which is then normalized to obtain the attention weight distribution. The calculated attention weights are then used to sum the original hidden state sequence, generating a focus context vector. Finally, a Bayesian inference-based SOH estimation module and dynamic threshold adjustment are applied.

[0015] The 1D CNN model makes the judgment, and when an anomaly is detected, the LSTM model is triggered for deep analysis. The LSTM model integrates attention focus information and SOH context, and finally outputs the fault type, trend prediction, root cause analysis and SOH confidence interval.

[0016] Step 5: Model training and optimization, and the evolution of virtual-real fusion.

[0017] Data acquisition involves collecting data from normal operating conditions, faults, and the entire lifecycle of aging and real-vehicle historical data.

[0018] Divide the dataset into training, validation, and test sets in a 7:2:1 ratio;

[0019] Model training involves inputting the divided dataset into two models, training the model with the training set data, and testing the model with the validation set data. Training stops when the model's accuracy and recall reach the set values, thus completing the model training process.

[0020] Step Six: Thermal State Analysis and Fault Handling. Input the preprocessed data into the trained model.

[0021] First, the 1D CNN model completes the thermal state analysis of the power battery. If it is normal, it continues to monitor the next round of sampling data. If it is abnormal, the LSTM model starts time series analysis and combines it with recent historical data to match the fault cause of the abnormal features according to the temperature change trend curve. A solution is generated and sent to the vehicle terminal.

[0022] This invention implements a dual-model approach combining real-time judgment and trend prediction. The model achieves millisecond-level response in emergency fault situations and uses LSTM time-series analysis to predict the development of faults in advance, solving the problem of reactive response. The model incorporates an attention mechanism, making the decision-making process transparent and facilitating the discovery of fault causes. Adaptive fusion and dual-model collaboration improve the accuracy of classifying complex faults. Bayesian inference is introduced into the real-time estimation of vehicle-side SOH (State of Health) and the thermal safety threshold is dynamically adjusted based on probabilistics to achieve baseline control for safety risk prevention.

[0023] Preferably, the method for installing the sparse sensor in step one is as follows:

[0024] Sparse contact temperature measurement: NTC temperature sensors are arranged at key thermal characteristic points of the battery module, including the module center, the tab connection and the cooling inlet. The ratio of the number of sensors to the number of cells is not less than 1:10.

[0025] Global thermal field reconstruction: Based on sparse NTC sampling data, a graph neural network (GNN) is used to reconstruct the virtual temperature field of the area inside the battery pack where no sensors are placed, generating virtual temperature data of the entire cell.

[0026] Cooling system soft sensing: Eliminates physical intrusive sensors in the fluid loop, and estimates the coolant flow rate and pipeline pressure in real time based on the speed feedback, duty cycle signal and inlet / outlet temperature difference of the electronic water pump, using the Bernoulli equation of the fluid dynamics model.

[0027] Preferably, the data acquisition conditions in step one are a dual-mode sampling strategy combining routine monitoring and event triggering, as detailed below:

[0028] Normal safety monitoring mode: When the battery is charging, discharging or resting, the system maintains continuous real-time monitoring with a sampling frequency set from 1Hz to 10Hz to capture sudden abnormal fluctuations in temperature and voltage.

[0029] High-frequency diagnostic trigger mode: High-frequency synchronous sampling is triggered when any of the following conditions are met, the sampling frequency is switched to 100Hz, and the data is transmitted to the dual model for in-depth diagnostics:

[0030] Condition A (Abnormal Trigger): In normal monitoring data, if the temperature change rate of any cell exceeds the set threshold, the voltage difference abnormally increases, or an insulation fault alarm occurs, it will be triggered immediately.

[0031] Condition B (Periodic Aging Monitoring): Triggered once every 10% change in battery SOC;

[0032] Condition C (Timed Inspection): During continuous battery operation, it is forcibly triggered once at set intervals.

[0033] Preferably, the collected data includes the cell temperature collected by the NTC temperature sensor, the airflow temperature collected by the infrared sensor, the pipeline pressure collected by the pressure sensor, and the coolant flow rate collected by the flow sensor. All physical data are synchronously converted into 16-bit digital electrical signals.

[0034] The acquired digital signals are transmitted to the preprocessing module of the battery management system via shielded cables. During the transmission process, CRC cyclic redundancy check is used to ensure data integrity.

[0035] Preferably, the data preprocessing method in step three is as follows:

[0036] Time synchronization: The IEEE 1588-2008 protocol is used to achieve accurate time alignment of multi-source data;

[0037] Noise reduction processing: Kalman filtering algorithm is used for coolant pressure and flow data;

[0038] Feature extraction targets time-domain and frequency-domain features:

[0039] Time-domain characteristics: Calculate the standard deviation of the temperature of all individual cells, the integral value of the temperature difference between the inlet and outlet of the battery pack, with the integration interval being the current 10% SOC cycle, in °C·min, and the difference between the maximum and minimum values ​​of the coolant pressure fluctuation.

[0040] Missing value fault handling: When individual sensor data loss or drift is detected, the spatiotemporal correlation filling algorithm is activated. Using the valid value of the sensor at the previous moment and the current reading of the adjacent sensors in the physical location, a replacement value is generated by weighted averaging to ensure the integrity of the feature vector of the input model.

[0041] Internal resistance feature extraction: Recursive least squares with forgetting factor (FF-RLS) is used; based on driving segment data uploaded from the vehicle, voltage V and current I, the battery equivalent circuit equation U=OCV-I×R is constructed. meas Real-time iterative identification of the observed ohmic internal resistance R of the battery meas ;

[0042] Calculate the standard deviation of the temperature of all cells:

[0043] ,

[0044] Where Ti is the temperature of the i-th cell. The average temperature of the battery cell;

[0045] Calculate the integral value of the temperature difference between the inlet and outlet of the battery pack:

[0046] ,

[0047] Where t0 and t1 are the start and end times of the current 10% SOC cycle, and ΔT(t) is the real-time temperature difference;

[0048] Calculate the coolant pressure fluctuation range:

[0049] ,

[0050] P max P min These are the maximum and minimum pressure values ​​within the sampling period;

[0051] Frequency domain characteristics: The discrete Fourier transform of the cell temperature fluctuation signal for each window is given by the following formula:

[0052] ,

[0053] Where x(n) is a discrete temperature fluctuation signal sequence, n=0,1,...,N−1 is the sampling point index, N is the number of sampling points, k=0,1,...,N−1 is the frequency point index, and j is the imaginary unit;

[0054] Bandwidth energy calculation: In a certain frequency band [f a f b The corresponding energy is ,k belongs to [f a f b ], |X(k)| is the magnitude of the Fourier transform result;

[0055] Normalization: The mean-variance normalization method is used.

[0056] ,

[0057] in, X represents the normalized feature value, μ represents the original feature data in the preprocessing stage, σ represents the feature mean of the bench test statistics, and σ represents the standard deviation. Outlier data points exceeding μ±5σ are deleted, and the data scale is unified to the interval [-1,1].

[0058] In the preferred embodiment, the method for building the sensing and diagnosis model in step four, based on dual-model collaboration and Bayesian SOH estimation for thermal state sensing and diagnosis, is as follows:

[0059] S4.1 Real-time initial screening on the vehicle: The pruned and quantized 1D CNN model is deployed in the vehicle BMS chip; the model only takes into account the real-time features at the current moment and completes the binary classification of "normal / suspected abnormal" within 50ms; if it is judged to be normal, high-frequency data is discarded; if it is judged to be suspected abnormal, the data upload mechanism is triggered.

[0060] It consists of three convolutional layers, employing 3×1, 5×1, and 7×1 convolutional kernels, max pooling, two pooling layers with a stride of 2, and one fully connected layer. This is used to capture spatial and transient thermal anomaly features at multiple scales. The activation function is ReLU, enabling rapid determination of cell temperature anomaly or normality within 50ms.

[0061] ,

[0062] in, This is the output feature map of the current 1D CNN model. W is the activation function for the 1D CNN model. (l) For convolution kernel weights, For 1D CNN The output feature map, b (l) The bias is *, which represents the convolution operation; the final fully connected layer outputs a binary classification probability P. normal With P abnormal ;

[0063] S4.2, Cloud-based Deep Diagnostics: The LSTM model and attention mechanism module are deployed on a cloud server to receive abnormal data packets uploaded from the vehicle and capture long-term temperature evolution patterns.

[0064] The LSTM model, used as a time series analyzer, has three hidden layers with 128 neurons per layer. It captures the time series changes in battery temperature within one hour through forget gates, input gates, output gates, and cell state structures. It adopts a "sliding window + overlapping sampling" approach, dividing one hour into 11 overlapping 5-minute windows, with each window overlapping for 2.5 minutes and each window corresponding to 500 sampling points. This multi-window fusion method ensures continuous coverage of the entire one-hour time series and captures the temperature evolution pattern.

[0065] The attention mechanism module is used to couple the output of the LSTM, assign weights to the hidden states at different time steps, focus on the key time segments most relevant to the fault, and output the fault type, development trend and contribution of key features.

[0066] The LSTM outputs hidden state vectors, h1, h2, ..., h3. n Each h nBoth encode the sequence information up to time t; a training vector u and a weight matrix are introduced into the attention mechanism to compute the hidden state h at each step t. t Relevance score e to the current query t

[0067] ,

[0068] in, It is the transpose of the learnable parameter (u) in the attention mechanism;

[0069] Then, the attention weight distribution is obtained by normalization using the softmax function:

[0070] ,

[0071] Weight α t It intuitively reflects the relative importance of the information at time t to the current fault diagnosis. Its value is between 0 and 1, and the sum of the weights of all time steps is 1.

[0072] The calculated attention weights are used to sum the original hidden state sequence to generate a focusing context vector c:

[0073] ,

[0074] Where, α t It is the attention weight, h t These are the hidden states of the LSTM at each time step;

[0075] S4.3 Cloud-based SOH update: The Bayesian SOH estimation module runs in the cloud, uses vehicle nighttime charging data to update the SOH, and sends the calculated "dynamic threshold adjustment parameters" back to the vehicle via OTA to update the judgment criteria of the 1D CNN;

[0076] An Arrhenius temperature compensation mechanism is introduced, utilizing the synchronously acquired cell temperature T to adjust the identified real-time internal resistance R. meas Normalized internal resistance converted to standard reference temperature (25℃),

[0077] ,

[0078] Among them, R meas The measured internal resistance is given by Ea, the activation energy of the battery material is given by k, and T is given by T. meas T represents the current temperature. ref The reference temperature is 25℃.

[0079] Prior distribution: ~

[0080] in, The probability distribution representing the battery health status is: μ0 describes the average level of SOH as the initial health baseline; σ0² describes the dispersion of SOH as the initial health difference.

[0081] The prior experience distribution is the probability distribution estimate of SOH based on historical data or industry experience when acquiring data;

[0082] Bayesian update, when a new R is obtained ohm Update the empirical distribution of SOH:

[0083] ,

[0084] Among them, E new E represents new observational data. old Historically collected observational data;

[0085] Constructing a unified aging model for the entire life cycle: Establishing a globally continuous and differentiable nonlinear model describing the change of internal resistance with health status; this model adopts the form of a "linear-exponential composite function", and its specific expression is as follows:

[0086]

[0087] Among them, R norm The normalized internal resistance is after temperature compensation; SOH represents the battery health state (value range 0 to 1). Characterizes the initial internal resistance intercept of the battery at the time of manufacture; The linear decay coefficient characterizes the linear increase component of internal resistance caused by the steady-state growth of the SEI film. and The exponential acceleration factor characterizes the component of accelerated increase in internal resistance at the end of battery life due to nonlinear degradation mechanism.

[0088] parameter[ The vectors are obtained through offline training throughout their entire lifecycle and are fine-tuned and updated online using recursive least squares (RLS) running in the cloud.

[0089] Constructing a Bayesian network: The health state SOH is defined as a hidden variable, and the normalized internal resistance R is set. norm With cumulative ampere-hour throughput Ah total For observational evidence;

[0090] Posterior probability update: When the cloud receives a new valid R norm When processing data, the likelihood function P(R) is calculated based on a unified aging model. norm |SOH), combined with the prior distribution P(SOH) from the previous time step. t-1The posterior probability distribution P(SOH|R) of SOH at the current time is updated using Bayes' theorem. norm Ah total Output the expected value of SOH and its confidence interval. ;

[0091] Calculation of aging factor: Calculation of dynamic aging factor based on expected SOH value.

[0092] ;

[0093] in, This indicates the degree of battery aging. This represents the mean of the updated SOH;

[0094] Generate dynamic thresholds: via formula Calculate the threshold adjustment factor, where k adj This is the preset tolerance coefficient;

[0095] coefficients Over-the-air (OTA) updates are sent to the vehicle's Battery Management System (BMS) via over-the-air (OTA) technology to adjust the thermal fault detection threshold of the 1D CNN model in real time. Achieve adaptive and precise early warning for batteries at different stages of aging;

[0096] 1D CNN quickly determines the fault, and LSTM deep analysis is triggered when an anomaly occurs. LSTM integrates attention focus information and SOH context, and finally outputs fault type, trend prediction, root cause analysis and SOH confidence interval.

[0097] Preferably, the data collection in step five includes the collection of three main categories of data:

[0098] The first normal operating condition data collection covers the discharge rate of 1C, 2C, 3C, and 4C, the ambient temperature from -20℃ to 60℃, and the combination of battery charge from 0% to 100% SOC, collecting no less than 100,000 sets of data.

[0099] 1C to 4C represent different discharge intensities. "C" stands for rate, which is the ratio of the current to the battery's rated capacity. 1C: The current equals the battery capacity, theoretically charging / discharging completely in 1 hour. 4C: The current is 4 times the battery capacity, theoretically charging / discharging completely in 15 minutes.

[0100] The second type of fault data covers thermal runaway, cooling failure, cell short circuit, and sensor failure, with no less than 5,000 sets of data for each category;

[0101] The third part involves collecting historical vehicle failure data and extracting valid historical data, based on the entire lifecycle aging and real-vehicle historical data.

[0102] The preferred option is that the data types, judgment criteria, and handling solutions for the actual vehicle historical faults are as follows:

[0103] A. Early risks of thermal runaway:

[0104] The criteria for judgment must meet the following characteristics:

[0105] 1. The temperature change rate of the core battery cell continues to exceed the standard, and this state lasts for more than 30 seconds.

[0106] 2. The temperature difference between battery cells continues to widen, reaching the highest cell temperature, and this difference shows a monotonically increasing trend.

[0107] 3. Abnormal voltage plateau drop: The voltage of the corresponding abnormal cell drops by more than 5% of the nominal voltage during the stable range.

[0108] 4. The energy of the abnormal cell temperature fluctuation signal at medium and high frequencies is more than 100% higher than the historical baseline value;

[0109] Handling plan: Immediately trigger the highest level alarm, execute emergency commands through the battery management system: forcibly disconnect the main circuit high voltage contactor, request the vehicle controller to enter emergency power-down mode, start all cooling units to maximum power, and send a remote alarm signal containing vehicle location information through the vehicle remote communication module;

[0110] B. Decline in cooling system performance:

[0111] The criteria for judgment must meet the following characteristics:

[0112] 1. Integral value of temperature difference between battery pack inlet and outlet I ΔT Exceeding 150% of the dynamic safety threshold calculated based on the current SOH value.

[0113] 2. The coolant flow sensor reading remains below 70% of the rated design value.

[0114] 3. System pressure fluctuation amplitude A P It exceeds twice the standard deviation of the historical average under normal operating conditions.

[0115] 4. The coolant flow or pressure signal exhibits significant distortion or disappearance of characteristic peaks in the low-frequency range, which characterizes the system's circulation characteristics.

[0116] Handling plan: Trigger an advanced warning. The control strategy is as follows: immediately limit the charging rate to below 0.3C and force the cooling pump speed to run at the highest speed. The user will be prompted through the vehicle's human-machine interface that "the cooling system efficiency has decreased. Please check it in time." If the flow rate fails to return to the normal range within the next 5 minutes, the driver is advised to find a safe place to stop and contact maintenance.

[0117] C. Severe degradation of cell consistency or connection failure:

[0118] The criteria for judgment must meet the following characteristics:

[0119] 1. The standard deviation of temperature for all battery cells shows a continuous upward trend.

[0120] 2. The voltage difference ΔV between the maximum and minimum cell voltages within the battery pack exceeds a preset threshold.

[0121] 3. Internal resistance exhibits a step increase while capacity does not decrease accordingly: The overall ohmic internal resistance R of the battery pack is estimated using the DC pulse method. ohm ,

[0122] 4. A single step increase exceeding 10% occurs, but the Bayesian SOH estimation module does not report a corresponding capacity jump signal during the same period.

[0123] Solution: This triggers a low-level warning. It is recommended that the user perform battery equalization maintenance or check the electrical and mechanical connection status between battery modules during the next parking session. In subsequent operation, it is recommended that the vehicle control system reduce the continuous discharge rate and avoid the battery state of charge (SOC) value falling below 20%.

[0124] D. Natural aging in a healthy state:

[0125] The criteria for judgment must meet the following characteristics:

[0126] 1. The SOH estimate shows a continuous and slow decreasing trend: the expected health status value μ output by the Bayesian SOH estimation module. SOH <85%, of which And its output 95% confidence region [μ SOH −2σ,μ SOH The convergence of the +2σ width indicates high estimation certainty.

[0127] 2. The internal resistance curve matches the aging curve and the prediction curve based on the aging model, and no step jump as seen in Class C faults is observed.

[0128] The proposed solution is as follows: No fault alarm will be triggered, but the system will automatically generate a battery health status report. Simultaneously, the system will adjust the report based on the calculated dynamic aging factor η. aging It automatically and individually adjusts various thermal protection thresholds and provides gentle information prompts on the vehicle's infotainment system interface.

[0129] The preferred option, the model training method in step five, is as follows:

[0130] The labeled data is input into the dual model, and the loss value is calculated using the cross-entropy loss function. The Adam optimizer is used with an initial learning rate of 0.001, which decays by 10% every 100 rounds. Backpropagation is performed to adjust the convolutional kernel weights and neural network connection weights. The model is tested on a validation set every 50 rounds. Training is stopped when the model accuracy is ≥96% and the recall is ≥92% to ensure the model's generalization ability.

[0131] Beneficial Effects: This invention achieves a dual-model approach combining real-time judgment and trend prediction. The model can respond in milliseconds during emergency faults, and LSTM time-series analysis predicts the development of faults in advance, solving the problem of reactive response. The model introduces an attention mechanism, making the decision-making process transparent and facilitating the discovery of fault causes. Adaptive fusion and dual-model collaboration improve the accuracy of classifying complex faults. Bayesian inference is introduced into the real-time SOH estimation at the vehicle end, and the thermal safety threshold is dynamically adjusted based on probabilistics to achieve baseline control for safety risk prevention. Attached Figure Description

[0132] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0133] Figure 1 This is a system flowchart of the present invention;

[0134] Figure 2 This is a schematic diagram illustrating the data processing principle of the present invention. Detailed Implementation

[0135] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0136] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0137] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0138] like Figure 1 As shown, a dual-model thermal early warning method for automotive power batteries, which integrates virtual and real-world data and cloud-edge collaboration, includes the following steps:

[0139] Step 1: Hardware deployment and software integration, install sparse sensors and deploy soft sensing algorithms, and set data acquisition conditions;

[0140] The method for installing a sparse sensor is as follows:

[0141] Sparse contact temperature measurement: NTC temperature sensors are arranged at key thermal characteristic points of the battery module, including the module center, the tab connection and the cooling inlet. The ratio of the number of sensors to the number of cells is not less than 1:10.

[0142] Global thermal field reconstruction: Based on sparse NTC sampling data, a graph neural network (GNN) is used to reconstruct the virtual temperature field of the area inside the battery pack where no sensors are placed, generating virtual temperature data of the entire cell.

[0143] Cooling system soft sensing: Eliminates physical intrusive sensors in the fluid loop, and estimates the coolant flow rate and pipeline pressure in real time based on the speed feedback, duty cycle signal and inlet / outlet temperature difference of the electronic water pump, using the Bernoulli equation of the fluid dynamics model.

[0144] The data collection conditions employ a dual-mode sampling strategy that combines routine monitoring with event-triggered data, as detailed below:

[0145] Normal safety monitoring mode: When the battery is charging, discharging or resting, the system maintains continuous real-time monitoring with a sampling frequency set from 1Hz to 10Hz to capture sudden abnormal fluctuations in temperature and voltage.

[0146] High-frequency diagnostic trigger mode: High-frequency synchronous sampling is triggered when any of the following conditions are met, the sampling frequency is switched to 100Hz, and the data is transmitted to the dual model for in-depth diagnostics:

[0147] Condition A (Abnormal Trigger): In normal monitoring data, if the temperature change rate (dT / dt) of any cell exceeds the set threshold, the voltage difference abnormally increases, or an insulation fault alarm occurs, it will be triggered immediately;

[0148] Condition B (Periodic Aging Monitoring): Triggered once every 10% change in battery SOC (used to collect full life cycle aging data, i.e., the data required for SOH estimation).

[0149] Condition C (Timed Inspection): During continuous battery operation, it is forcibly triggered once every set time interval (e.g., 15 minutes).

[0150] Step 2: Synchronous acquisition of multimodal data. The sensor array is triggered to start data sampling according to the preset logic, and the acquired data is transmitted to the preprocessing module.

[0151] The collected data includes the cell temperature collected by the NTC temperature sensor, the airflow temperature collected by the infrared sensor, the pipeline pressure collected by the pressure sensor, and the coolant flow rate collected by the flow sensor. All physical data are synchronously converted into 16-bit digital electrical signals.

[0152] The acquired digital signals are transmitted to the preprocessing module of the battery management system via shielded cables. During the transmission process, CRC cyclic redundancy check is used to ensure data integrity.

[0153] Step 3: Data preprocessing. The obtained raw data is preprocessed, including time synchronization, noise reduction, feature extraction, and normalization.

[0154] The data preprocessing methods are as follows:

[0155] Time synchronization: The IEEE 1588-2008 protocol is used to achieve accurate time alignment of multi-source data;

[0156] Noise reduction processing: Kalman filtering algorithm is used for coolant pressure and flow data;

[0157] Feature extraction targets time-domain and frequency-domain features:

[0158] Time-domain characteristics: Calculate the standard deviation of the temperature of all individual cells, the integral value of the temperature difference between the inlet and outlet of the battery pack, with the integration interval being the current 10% SOC cycle, in °C·min, and the difference between the maximum and minimum values ​​of the coolant pressure fluctuation.

[0159] Missing value fault handling: When individual sensor data loss or drift is detected, the spatiotemporal correlation filling algorithm is activated. Using the valid value of the sensor at the previous moment and the current reading of the adjacent sensors in the physical location, a replacement value is generated by weighted averaging to ensure the integrity of the feature vector of the input model.

[0160] Internal resistance feature extraction: Recursive least squares with forgetting factor (FF-RLS) is used; based on driving segment data uploaded from the vehicle, voltage V and current I, the battery equivalent circuit equation U=OCV-I×R is constructed. meas Real-time iterative identification of the observed ohmic internal resistance R of the battery meas ;

[0161] Calculate the standard deviation of the temperature of all cells:

[0162] ,

[0163] Where Ti is the temperature of the i-th cell. The average temperature of the battery cell;

[0164] Calculate the integral value of the temperature difference between the inlet and outlet of the battery pack:

[0165] ,

[0166] Where t0 and t1 are the start and end times of the current 10% SOC cycle, and ΔT(t) is the real-time temperature difference;

[0167] Calculate the coolant pressure fluctuation range:

[0168] ,

[0169] P max P min These are the maximum and minimum pressure values ​​within the sampling period;

[0170] Frequency domain characteristics: The discrete Fourier transform of the cell temperature fluctuation signal for each window is given by the following formula:

[0171] ,

[0172] Where x(n) is a discrete temperature fluctuation signal sequence, n=0,1,...,N−1 is the sampling point index, N is the number of sampling points, k=0,1,...,N−1 is the frequency point index, and j is the imaginary unit;

[0173] Bandwidth energy calculation: In a certain frequency band [f a f b The corresponding energy is ,k belongs to [f a f b ], |X(k)| is the magnitude of the Fourier transform result;

[0174] Normalization: The mean-variance normalization method is used.

[0175] ,

[0176] in, X represents the normalized feature value, μ represents the original feature data in the preprocessing stage, σ represents the feature mean of the bench test statistics, and σ represents the standard deviation. Outlier data points exceeding μ±5σ are deleted, and the data scale is unified to the interval [-1,1].

[0177] Step 4: Build a perception and diagnostic model, employing a collaborative approach of vehicle-side edge computing and cloud-based deep learning. A lightweight 1D CNN model is deployed on the vehicle for real-time screening, while an LSTM model and a Bayesian SOH estimation module are deployed in the cloud for deep diagnostics and parameter iteration.

[0178] The data obtained in step 3 is input into a 1D CNN model as a real-time classification model for hot states.

[0179] LSTM models are used as time series analyzers to capture temperature evolution patterns;

[0180] The attention mechanism module is used to couple the output of the LSTM, assign weights to the hidden states at different time steps, focus on the key time segments most relevant to the fault, and output the fault type, development trend and contribution of key features.

[0181] The LSTM outputs a hidden state vector, which is then normalized to obtain the attention weight distribution. The calculated attention weights are then used to sum the original hidden state sequence, generating a focus context vector. Finally, a Bayesian inference-based SOH estimation module and dynamic threshold adjustment are applied.

[0182] The 1D CNN model makes the judgment, and when an anomaly is detected, the LSTM model is triggered for deep analysis. The LSTM model integrates attention focus information and SOH context, and finally outputs the fault type, trend prediction, root cause analysis and SOH confidence interval.

[0183] The following method is used to construct a sensing and diagnostic model for thermal state sensing and diagnosis based on dual-model collaboration and Bayesian SOH estimation:

[0184] S4.1 Real-time initial screening on the vehicle: The pruned and quantized 1D CNN model is deployed in the vehicle BMS chip; the model only takes into account the real-time features at the current moment and completes the binary classification of "normal / suspected abnormal" within 50ms; if it is judged to be normal, high-frequency data is discarded; if it is judged to be suspected abnormal, the data upload mechanism is triggered.

[0185] It consists of three convolutional layers, employing 3×1, 5×1, and 7×1 convolutional kernels, max pooling, two pooling layers with a stride of 2, and one fully connected layer. This is used to capture spatial and transient thermal anomaly features at multiple scales. The activation function is ReLU, enabling rapid determination of cell temperature anomaly or normality within 50ms.

[0186] ,

[0187] in, This is the output feature map of the current 1D CNN model. W is the activation function for the 1D CNN model. (l) For convolution kernel weights, For 1D CNN The output feature map, b (l) The bias is *, which represents the convolution operation; the final fully connected layer outputs a binary classification probability P. normal With P abnormal ;

[0188] S4.2, Cloud-based Deep Diagnostics: The LSTM model and attention mechanism module are deployed on a cloud server to receive abnormal data packets uploaded from the vehicle and capture long-term temperature evolution patterns.

[0189] The LSTM model, used as a time series analyzer, has three hidden layers with 128 neurons per layer. It captures the time series changes in battery temperature within one hour through forget gates, input gates, output gates, and cell state structures. It adopts a "sliding window + overlapping sampling" approach, dividing one hour into 11 overlapping 5-minute windows, with each window overlapping for 2.5 minutes and each window corresponding to 500 sampling points. This multi-window fusion method ensures continuous coverage of the entire one-hour time series and captures the temperature evolution pattern.

[0190] The attention mechanism module is used to couple the output of the LSTM, assign weights to the hidden states at different time steps, focus on the key time segments most relevant to the fault, and output the fault type, development trend and contribution of key features.

[0191] The LSTM outputs hidden state vectors, h1, h2, ..., h3. n Each h n Both encode the sequence information up to time t; a training vector u and a weight matrix are introduced into the attention mechanism to compute the hidden state h at each step t. t Relevance score e to the current query t

[0192] ,

[0193] in, It is the transpose of the learnable parameter (u) in the attention mechanism;

[0194] Then, the attention weight distribution is obtained by normalization using the softmax function:

[0195] ,

[0196] Weight α t It intuitively reflects the relative importance of the information at time t to the current fault diagnosis. Its value is between 0 and 1, and the sum of the weights of all time steps is 1.

[0197] The calculated attention weights are used to sum the original hidden state sequence to generate a focusing context vector c:

[0198] ,

[0199] Where, α t It is the attention weight, h t These are the hidden states of the LSTM at each time step;

[0200] S4.3 Cloud-based SOH update: The Bayesian SOH estimation module runs in the cloud, uses vehicle nighttime charging data to update the SOH, and sends the calculated "dynamic threshold adjustment parameters" back to the vehicle via OTA to update the judgment criteria of the 1D CNN;

[0201] An Arrhenius temperature compensation mechanism is introduced, utilizing the synchronously acquired cell temperature T to adjust the identified real-time internal resistance R. meas Normalized internal resistance converted to standard reference temperature (25℃),

[0202] ,

[0203] Among them, R meas The measured internal resistance is given by Ea, the activation energy of the battery material is given by k, and T is given by T. meas T represents the current temperature. ref The reference temperature is 25℃.

[0204] Prior distribution: ~

[0205] in, The probability distribution representing the battery health status is: μ0 describes the average level of SOH (initial health baseline); σ0² describes the dispersion of SOH (initial health difference).

[0206] The prior experience distribution is the probability distribution estimate of SOH based on historical data or industry experience when acquiring data.

[0207] Bayesian update, when a new R is obtained ohm Update the empirical distribution of SOH:

[0208] ,

[0209] Among them, E newE represents new observational data. old Historically collected observational data;

[0210] Constructing a unified aging model for the entire life cycle: Establishing a globally continuous and differentiable nonlinear model describing the change of internal resistance with health status; this model adopts the form of a "linear-exponential composite function", and its specific expression is as follows:

[0211]

[0212] Among them, R norm The normalized internal resistance is after temperature compensation; SOH represents the battery health state (value range 0 to 1). Characterizes the initial internal resistance intercept of the battery at the time of manufacture; The linear decay coefficient characterizes the linear increase component of internal resistance caused by the steady-state growth of the SEI film. and The exponential acceleration factor characterizes the component of accelerated increase in internal resistance at the end of battery life due to nonlinear degradation mechanism.

[0213] parameter[ The vectors are obtained through offline training throughout their entire lifecycle and are fine-tuned and updated online using recursive least squares (RLS) running in the cloud.

[0214] Constructing a Bayesian network: The health state SOH is defined as a hidden variable, and the normalized internal resistance R is set. norm With cumulative ampere-hour throughput Ah total For observational evidence;

[0215] Posterior probability update: When the cloud receives a new valid R norm When processing data, the likelihood function P(R) is calculated based on a unified aging model. norm |SOH), combined with the prior distribution P(SOH) from the previous time step. t-1 The posterior probability distribution P(SOH|R) of SOH at the current time is updated using Bayes' theorem. norm Ah total Output the expected value of SOH and its confidence interval. ;

[0216] Calculation of aging factor: Calculation of dynamic aging factor based on expected SOH value.

[0217] ;

[0218] in, This indicates the degree of battery aging. This represents the mean of the updated SOH.

[0219] Generate dynamic thresholds: via formula Calculate the threshold adjustment factor, where k adj This is the preset tolerance coefficient;

[0220] coefficients Over-the-air (OTA) updates are sent to the vehicle's Battery Management System (BMS) via over-the-air (OTA) technology to adjust the thermal fault detection threshold of the 1D CNN model in real time. Achieve adaptive and precise early warning for batteries at different stages of aging;

[0221] 1D CNN quickly determines the fault, and LSTM deep analysis is triggered when an anomaly occurs. LSTM integrates attention focus information and SOH context, and finally outputs fault type, trend prediction, root cause analysis and SOH confidence interval.

[0222] Step 5: Model training and optimization, and the evolution of virtual-real fusion.

[0223] Data acquisition involves collecting data from normal operating conditions, faults, and the entire lifecycle of aging and real-vehicle historical data.

[0224] Data collection includes the collection of three main categories of data:

[0225] The first normal operating condition data collection covers the discharge rate of 1C, 2C, 3C, and 4C, the ambient temperature from -20℃ to 60℃, and the combination of battery charge from 0% to 100% SOC, collecting no less than 100,000 sets of data.

[0226] 1C to 4C represent different discharge intensities. "C" stands for rate, which is the ratio of the current to the battery's rated capacity. 1C: The current equals the battery capacity, theoretically charging / discharging completely in 1 hour. 4C: The current is 4 times the battery capacity, theoretically charging / discharging completely in 15 minutes.

[0227] The second type of fault data covers thermal runaway, cooling failure, cell short circuit, and sensor failure, with no less than 5,000 sets of data for each category;

[0228] The third part involves collecting historical vehicle failure data and extracting valid historical data, based on the entire lifecycle aging and real-vehicle historical data.

[0229] The types of historical fault data, judgment criteria, and handling solutions for the actual vehicles are as follows:

[0230] A. Early risks of thermal runaway:

[0231] The criteria for judgment must meet the following characteristics:

[0232] 1. The temperature change rate of the core battery cell continues to exceed the standard, and this state lasts for more than 30 seconds.

[0233] 2. The temperature difference between battery cells continues to widen, reaching the highest cell temperature, and this difference shows a monotonically increasing trend.

[0234] 3. Abnormal voltage plateau drop: The voltage of the corresponding abnormal cell drops by more than 5% of the nominal voltage during the stable range.

[0235] 4. The energy of the abnormal cell temperature fluctuation signal at medium and high frequencies is more than 100% higher than the historical baseline value;

[0236] Handling plan: Immediately trigger the highest level alarm, execute emergency commands through the battery management system: forcibly disconnect the main circuit high voltage contactor, request the vehicle controller to enter emergency power-down mode, start all cooling units to maximum power, and send a remote alarm signal containing vehicle location information through the vehicle remote communication module;

[0237] B. Decline in cooling system performance:

[0238] The criteria for judgment must meet the following characteristics:

[0239] 1. Integral value of temperature difference between battery pack inlet and outlet I ΔT Exceeding 150% of the dynamic safety threshold calculated based on the current SOH value.

[0240] 2. The coolant flow sensor reading remains below 70% of the rated design value.

[0241] 3. System pressure fluctuation amplitude A P It exceeds twice the standard deviation of the historical average under normal operating conditions.

[0242] 4. The coolant flow or pressure signal exhibits significant distortion or disappearance of characteristic peaks in the low-frequency range, which characterizes the system's circulation characteristics.

[0243] Handling plan: Trigger an advanced warning. The control strategy is as follows: immediately limit the charging rate to below 0.3C and force the cooling pump speed to run at the highest speed. The user will be prompted through the vehicle's human-machine interface that "the cooling system efficiency has decreased. Please check it in time." If the flow rate fails to return to the normal range within the next 5 minutes, the driver is advised to find a safe place to stop and contact maintenance.

[0244] C. Severe degradation of cell consistency or connection failure:

[0245] The criteria for judgment must meet the following characteristics:

[0246] 1. The standard deviation of temperature for all battery cells shows a continuous upward trend.

[0247] 2. The voltage difference ΔV between the maximum and minimum cell voltages within the battery pack exceeds a preset threshold.

[0248] 3. Internal resistance exhibits a step increase while capacity does not decrease accordingly: The overall ohmic internal resistance R of the battery pack is estimated using the DC pulse method. ohm ,

[0249] 4. A single step increase exceeding 10% occurs, but the Bayesian SOH estimation module does not report a corresponding capacity jump signal during the same period.

[0250] Solution: This triggers a low-level warning. It is recommended that the user perform battery equalization maintenance or check the electrical and mechanical connection status between battery modules during the next parking session. In subsequent operation, it is recommended that the vehicle control system reduce the continuous discharge rate and avoid the battery state of charge (SOC) value falling below 20%.

[0251] D. Natural aging in a healthy state:

[0252] The criteria for judgment must meet the following characteristics:

[0253] 1. The SOH estimate shows a continuous and slow decreasing trend: the expected health status value μ output by the Bayesian SOH estimation module. SOH <85%, of which And its output 95% confidence region [μ SOH −2σ,μ SOH The convergence of the +2σ width indicates high estimation certainty.

[0254] 2. The internal resistance curve matches the aging curve and the prediction curve based on the aging model, and no step jump as seen in Class C faults is observed.

[0255] The proposed solution is as follows: No fault alarm will be triggered, but the system will automatically generate a battery health status report. Simultaneously, the system will adjust the report based on the calculated dynamic aging factor η. aging It automatically and individually adjusts various thermal protection thresholds and provides gentle information prompts on the vehicle's infotainment system interface.

[0256] Divide the dataset into training, validation, and test sets in a 7:2:1 ratio;

[0257] Model training involves inputting the divided dataset into two models, training the model with the training set data, and testing the model with the validation set data. Training stops when the model's accuracy and recall reach the set values, thus completing the model training process.

[0258] The model training method is as follows:

[0259] The labeled data is input into the dual model, and the loss value is calculated using the cross-entropy loss function. The Adam optimizer is used with an initial learning rate of 0.001, which decays by 10% every 100 rounds. Backpropagation is performed to adjust the convolutional kernel weights and neural network connection weights. The model is tested on a validation set every 50 rounds. Training is stopped when the model accuracy is ≥96% and the recall is ≥92% to ensure the model's generalization ability.

[0260] Step Six: Thermal State Analysis and Fault Handling. Input the preprocessed data into the trained model.

[0261] First, the 1D CNN model completes the thermal state analysis of the power battery. If it is normal, it continues to monitor the next round of sampling data. If it is abnormal, the LSTM model starts time series analysis and combines it with recent historical data to match the fault cause of the abnormal features according to the temperature change trend curve. A solution is generated and sent to the vehicle terminal.

[0262] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0263] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dual-model thermal early warning method for automotive power batteries, characterized in that, Includes the following steps: Step 1: Hardware deployment and software integration, install sparse sensors and deploy soft sensing algorithms, and set data acquisition conditions; Step 2: Synchronous acquisition of multimodal data. The sensor array is triggered to start data sampling according to the preset logic, and the acquired data is transmitted to the preprocessing module. Step 3: Data preprocessing. The obtained raw data is preprocessed, including time synchronization, noise reduction, feature extraction, and normalization. Step 4: Build a perception and diagnostic model, employing a collaborative approach of vehicle-side edge computing and cloud-based deep learning. A lightweight 1D CNN model is deployed on the vehicle for real-time screening, while an LSTM model and a Bayesian SOH estimation module are deployed in the cloud for deep diagnostics and parameter iteration. The data obtained in step 3 is input into a 1D CNN model as a real-time classification model for hot states. LSTM models are used as time series analyzers to capture temperature evolution patterns; The attention mechanism module is used to couple the output of the LSTM, assign weights to the hidden states at different time steps, focus on the key time segments most relevant to the fault, and output the fault type, development trend and contribution of key features. The LSTM outputs a hidden state vector, which is then normalized to obtain the attention weight distribution. The calculated attention weights are then used to sum the original hidden state sequence, generating a focus context vector. Finally, a Bayesian inference-based SOH estimation module and dynamic threshold adjustment are applied. The 1D CNN model makes the judgment, and the LSTM model is triggered for deep analysis when an anomaly occurs. The LSTM model integrates attention focus information and SOH context, and finally outputs the fault type, trend prediction, root cause analysis and SOH confidence interval. Step 5: Model training and optimization, and the evolution of virtual-real fusion. Data acquisition involves collecting data from normal operating conditions, faults, and the entire lifecycle of aging and real-vehicle historical data. Divide the dataset into training set, validation set, and test set in a 7:2:1 ratio; Model training involves inputting the divided dataset into two models, training the model with the training set data, and testing the model with the validation set data. Training stops when the model's accuracy and recall reach the set values, thus completing the model training process. Step Six: Thermal State Analysis and Fault Handling. Input the preprocessed data into the trained model. First, the 1D CNN model completes the thermal state analysis of the power battery. If it is normal, it continues to monitor the next round of sampling data. If it is abnormal, the LSTM model starts time series analysis and combines it with recent historical data to match the fault cause of the abnormal features according to the temperature change trend curve. A solution is generated and sent to the vehicle terminal.

2. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration as described in claim 1, characterized in that, The method for installing the sparse sensor in step one is as follows: Sparse contact temperature measurement: NTC temperature sensors are arranged at key thermal characteristic points of the battery module, including the module center, the tab connection and the cooling inlet. The ratio of the number of sensors to the number of cells is not less than 1:

10. Global thermal field reconstruction: Based on sparse NTC sampling data, a graph neural network (GNN) is used to reconstruct the virtual temperature field of the area inside the battery pack where no sensors are placed, generating virtual temperature data of the entire cell. Cooling system soft sensing: Eliminates physical intrusive sensors in the fluid loop, and estimates the coolant flow rate and pipeline pressure in real time based on the speed feedback, duty cycle signal and inlet / outlet temperature difference of the electronic water pump, using the Bernoulli equation of the fluid dynamics model.

3. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration as described in claim 1, characterized in that: The data acquisition conditions in step one are a dual-mode sampling strategy combining routine monitoring and event triggering, as detailed below: Normal safety monitoring mode: When the battery is charging, discharging or resting, the system maintains continuous real-time monitoring with a sampling frequency set from 1Hz to 10Hz to capture sudden abnormal fluctuations in temperature and voltage. High-frequency diagnostic trigger mode: High-frequency synchronous sampling is triggered when any of the following conditions are met, the sampling frequency is switched to 100Hz, and the data is transmitted to the dual model for in-depth diagnostics: Condition A (Abnormal Trigger): In normal monitoring data, if the temperature change rate of any cell exceeds the set threshold, the voltage difference abnormally increases, or an insulation fault alarm occurs, it will be triggered immediately. Condition B (Periodic Aging Monitoring): Triggered once every 10% change in battery SOC; Condition C (Timed Inspection): During continuous battery operation, it is forcibly triggered once at set intervals.

4. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration as described in claim 2, characterized in that: The collected data includes the cell temperature collected by the NTC temperature sensor, the airflow temperature collected by the infrared sensor, the pipeline pressure collected by the pressure sensor, and the coolant flow rate collected by the flow sensor. All physical data are synchronously converted into 16-bit digital electrical signals. The acquired digital signals are transmitted to the preprocessing module of the battery management system via shielded cables. During the transmission process, CRC cyclic redundancy check is used to ensure data integrity.

5. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration as described in claim 1, characterized in that, The data preprocessing method in step three is as follows: Time synchronization: The IEEE 1588-2008 protocol is used to achieve accurate time alignment of multi-source data; Noise reduction processing: Kalman filtering algorithm is used for coolant pressure and flow data; Feature extraction targets time-domain and frequency-domain features: Time-domain characteristics: Calculate the standard deviation of the temperature of all individual cells, the integral value of the temperature difference between the inlet and outlet of the battery pack, with the integration interval being the current 10% SOC cycle, in °C·min, and the difference between the maximum and minimum values ​​of the coolant pressure fluctuation. Missing value fault handling: When individual sensor data loss or drift is detected, the spatiotemporal correlation filling algorithm is activated. Using the valid value of the sensor at the previous moment and the current reading of the adjacent sensors in the physical location, a replacement value is generated by weighted averaging to ensure the integrity of the feature vector of the input model. Internal resistance feature extraction: Recursive least squares with forgetting factor (FF-RLS) is used; Based on driving segment data uploaded from the vehicle, including voltage V and current I, the equivalent circuit equation for the battery is constructed as U = OCV - I × R. meas Real-time iterative identification of the observed ohmic internal resistance R of the battery meas ; Calculate the standard deviation of the temperature of all cells: ， Where Ti is the temperature of the i-th cell. The average temperature of the battery cell; Calculate the integral value of the temperature difference between the inlet and outlet of the battery pack: ， Where t0 and t1 are the start and end times of the current 10% SOC cycle, and ΔT(t) is the real-time temperature difference; Calculate the coolant pressure fluctuation range: ， P max P min These are the maximum and minimum pressure values ​​within the sampling period; Frequency domain characteristics: The discrete Fourier transform of the cell temperature fluctuation signal for each window is given by the following formula: ， Where x(n) is a discrete temperature fluctuation signal sequence, n=0,1,...,N−1 is the sampling point index, N is the number of sampling points, k=0,1,...,N−1 is the frequency point index, and j is the imaginary unit; Bandwidth energy calculation: In a certain frequency band [f a f b The corresponding energy is ,k belongs to [f a f b ], |X(k)| is the magnitude of the Fourier transform result; Normalization: The mean-variance normalization method is used. ， in, X represents the normalized feature value, μ represents the original feature data in the preprocessing stage, σ represents the feature mean of the bench test statistics, and σ represents the standard deviation. Outlier data points exceeding μ±5σ are deleted, and the data scale is unified to the interval [-1,1].

6. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration as described in claim 1, characterized in that, In step four, the method for building a sensing and diagnostic model based on dual-model collaboration and Bayesian SOH estimation for thermal state sensing and diagnosis is as follows: S4.1 Real-time initial screening at the vehicle end: The pruned and quantized 1D CNN model is deployed in the vehicle BMS chip; the model only takes into account the real-time features at the current moment and completes the binary classification of "normal / suspected abnormal" within 50ms; If the result is normal, discard the high-frequency data. If a suspected anomaly is detected, the data upload mechanism is triggered; It consists of three convolutional layers, employing 3×1, 5×1, and 7×1 convolutional kernels, max pooling, two pooling layers with a stride of 2, and one fully connected layer. This is used to capture spatial and transient thermal anomaly features at multiple scales. The activation function is ReLU, enabling rapid determination of cell temperature anomaly or normality within 50ms. ， in, This is the output feature map of the current 1D CNN model. W is the activation function for the 1D CNN model. (l) For convolution kernel weights, For 1D CNN The output feature map, b (l) The bias is *, which represents the convolution operation; the final fully connected layer outputs a binary classification probability P. normal With P abnormal ; S4.2, Cloud-based Deep Diagnostics: The LSTM model and attention mechanism module are deployed on a cloud server to receive abnormal data packets uploaded from the vehicle and capture long-term temperature evolution patterns. The LSTM model, used as a time series analyzer, has three hidden layers with 128 neurons per layer. It captures the time series changes in battery temperature within one hour through forget gates, input gates, output gates, and cell state structures. It adopts a "sliding window + overlapping sampling" approach, dividing one hour into 11 overlapping 5-minute windows, with each window overlapping for 2.5 minutes and each window corresponding to 500 sampling points. This multi-window fusion method ensures continuous coverage of the entire one-hour time series and captures the temperature evolution pattern. The attention mechanism module is used to couple the output of the LSTM, assign weights to the hidden states at different time steps, focus on the key time segments most relevant to the fault, and output the fault type, development trend and contribution of key features. The LSTM outputs hidden state vectors, h1, h2, ..., h3. n Each h n Both encode the sequence information up to time t; a training vector u and a weight matrix are introduced into the attention mechanism to compute the hidden state h at each step t. t Relevance score e to the current query t ， in, It is the transpose of the learnable parameter (u) in the attention mechanism; Then, the attention weight distribution is obtained by normalization using the softmax function: ， weight α t It intuitively reflects the relative importance of the information at time t to the current fault diagnosis. Its value is between 0 and 1, and the sum of the weights of all time steps is 1. The calculated attention weights are used to sum the original hidden state sequence to generate a focusing context vector c: ， Where, α t It is the attention weight, h t These are the hidden states of the LSTM at each time step; S4.3 Cloud-based SOH update: The Bayesian SOH estimation module runs in the cloud, uses vehicle nighttime charging data to update the SOH, and sends the calculated "dynamic threshold adjustment parameters" back to the vehicle via OTA to update the judgment criteria of the 1D CNN; An Arrhenius temperature compensation mechanism is introduced, utilizing the synchronously acquired cell temperature T to adjust the identified real-time internal resistance R. meas Normalized internal resistance converted to standard reference temperature (25℃), ， Among them, R meas The measured internal resistance is given by Ea, the activation energy of the battery material is given by k, and T is given by T. meas T represents the current temperature. ref The reference temperature is 25℃. Prior distribution: ~ in, The probability distribution representing the battery health status is: μ0 describes the average level of SOH as the initial health baseline; σ0² describes the dispersion of SOH as the initial health difference. The prior experience distribution is the probability distribution estimate of SOH based on historical data or industry experience when acquiring data; Bayesian update, when a new R is obtained ohm Update the empirical distribution of SOH: ， Among them, E new E represents new observational data. old Historically collected observational data; Constructing a unified aging model for the entire life cycle: Establishing a globally continuous and differentiable nonlinear model describing the change of internal resistance with health status; this model adopts the form of a "linear-exponential composite function", and its specific expression is as follows: Among them, R norm The normalized internal resistance is after temperature compensation; SOH represents the battery health state (value range 0 to 1). Characterizes the initial internal resistance intercept of the battery at the time of manufacture; The linear decay coefficient characterizes the linear increase component of internal resistance caused by the steady-state growth of the SEI film. and The exponential acceleration factor characterizes the component of accelerated increase in internal resistance at the end of battery life due to nonlinear degradation mechanism. parameter[ The vectors are obtained through offline training throughout their entire lifecycle and are fine-tuned and updated online using recursive least squares (RLS) running in the cloud. Constructing a Bayesian network: The health state SOH is defined as a hidden variable, and the normalized internal resistance R is set. norm With cumulative ampere-hour throughput Ah total For observational evidence; Posterior probability update: When the cloud receives a new valid R norm When processing data, the likelihood function P(R) is calculated based on a unified aging model. norm |SOH), combined with the prior distribution P(SOH) from the previous time step. t-1 The posterior probability distribution P(SOH|R) of SOH at the current time is updated using Bayes' theorem. norm Ah total Output the expected value of SOH and its confidence interval. ; Calculation of aging factor: Calculation of dynamic aging factor based on expected SOH value. ； in, This indicates the degree of battery aging. This represents the mean of the updated SOH; Generate dynamic thresholds: via formula Calculate the threshold adjustment factor, where k adj This is the preset tolerance coefficient; coefficients Over-the-air (OTA) updates are sent to the vehicle's Battery Management System (BMS) via over-the-air (OTA) technology to adjust the thermal fault detection threshold of the 1D CNN model in real time. Achieve adaptive and precise early warning for batteries at different stages of aging; 1D CNN quickly determines the fault, and LSTM deep analysis is triggered when an anomaly occurs. LSTM integrates attention focus information and SOH context, and finally outputs fault type, trend prediction, root cause analysis and SOH confidence interval.

7. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration as described in claim 1, characterized in that, The data collection in step five includes the collection of three main categories of data: The first normal operating condition data collection covers the discharge rate of 1C, 2C, 3C, and 4C, the ambient temperature from -20℃ to 60℃, and the combination of battery charge from 0% to 100% SOC, collecting no less than 100,000 sets of data. The second type of fault data covers thermal runaway, cooling failure, cell short circuit, and sensor failure, with no less than 5,000 sets of data for each category; The third part involves collecting historical vehicle failure data and extracting valid historical data, based on the entire lifecycle aging and real-vehicle historical data.

8. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration as described in claim 7, characterized in that, The types of historical fault data, judgment criteria, and handling solutions for the actual vehicles are as follows: A. Early risks of thermal runaway: The criteria for judgment must meet the following characteristics: The core cell temperature change rate continued to exceed the standard, and this state lasted for more than 30 seconds. The temperature difference between battery cells continues to widen, reaching the highest cell temperature, and this difference shows a monotonically increasing trend. An abnormal voltage plateau occurs when the voltage of the corresponding abnormal cell drops by more than 5% from its nominal voltage during the stable phase. The energy of the abnormal cell temperature fluctuation signal at mid-to-high frequencies was more than 100% higher than the historical baseline value; Handling plan: Immediately trigger the highest level alarm, execute emergency commands through the battery management system: forcibly disconnect the main circuit high voltage contactor, request the vehicle controller to enter emergency power-down mode, start all cooling units to maximum power, and send a remote alarm signal containing vehicle location information through the vehicle remote communication module; B. Decline in cooling system performance: The criteria for judgment must meet the following characteristics: battery pack Integral value of temperature difference between inlet and outlet I ΔT Exceeding 150% of the dynamic safety threshold calculated based on the current SOH value. The coolant flow sensor reading remains below 70% of the rated design value. System pressure fluctuation amplitude A P It exceeds twice the standard deviation of the historical average under normal operating conditions. The coolant flow or pressure signals exhibit significant energy distribution distortion or disappearance of characteristic peaks in the low-frequency range characterizing the system's circulation characteristics. Handling plan: Trigger an advanced warning. The control strategy is as follows: immediately limit the charging rate to below 0.3C and force the cooling pump speed to run at the highest speed. The user will be prompted through the vehicle's human-machine interface that "the cooling system efficiency has decreased. Please check it in time." If the flow rate fails to return to the normal range within the next 5 minutes, the driver is advised to find a safe place to stop and contact maintenance. C. Severe degradation of cell consistency or connection failure: The criteria for judgment must meet the following characteristics: The standard deviation of all cell temperatures showed a continuous upward trend. The voltage difference ΔV between the maximum and minimum cell voltages within the battery pack exceeds a preset threshold. The internal resistance exhibits a step increase while the capacity does not decrease accordingly: the overall ohmic internal resistance R of the battery pack is estimated using the DC pulse method. ohm , A single step increase exceeding 10% occurred, but during the same period, the Bayesian SOH estimation module did not report a corresponding capacity jump signal. Solution: This triggers a low-level warning. It is recommended that the user perform battery equalization maintenance or check the electrical and mechanical connection status between battery modules during the next parking session. In subsequent operation, it is recommended that the vehicle control system reduce the continuous discharge rate and avoid the battery state of charge (SOC) value falling below 20%. D. Natural aging in a healthy state: The criteria for judgment must meet the following characteristics: The SOH estimate shows a continuous, slow downward trend: the expected health status value μ output by the Bayesian SOH estimation module. SOH <85%, of which And its output 95% confidence region [μ SOH −2σ,μ SOH The convergence of the +2σ width indicates high estimation certainty. The internal resistance curve matches the aging curve and the prediction curve based on the aging model, and no step jump as seen in Class C faults is observed. The proposed solution is as follows: No fault alarm will be triggered, but the system will automatically generate a battery health status report. Simultaneously, the system will adjust the report based on the calculated dynamic aging factor η. aging It automatically and individually adjusts various thermal protection thresholds and provides gentle information prompts on the vehicle's infotainment system interface.

9. The dual-model thermal early warning method for automotive power batteries based on virtual-real fusion and cloud-edge collaboration according to claim 7, characterized in that, The model training method in step five is as follows: The labeled data is input into the dual model, and the loss value is calculated using the cross-entropy loss function. The Adam optimizer is used with an initial learning rate of 0.001, which decays by 10% every 100 rounds. Backpropagation is performed to adjust the convolutional kernel weights and neural network connection weights. The model is tested on a validation set every 50 rounds. Training is stopped when the model accuracy is ≥96% and the recall is ≥92% to ensure the model's generalization ability.

Images