CN-122020298-A - Virtual-real fusion cloud edge cooperative double-model thermal early warning method for automobile power battery

Abstract

The invention discloses a virtual-actual fusion cloud edge collaborative double-model thermal early warning method for an automobile power battery, which comprises six steps of hardware deployment and software integration, multi-mode data synchronous acquisition, data pre-processing, construction of a perception and diagnosis model, model training and optimization, virtual-actual fusion evolution, thermal state analysis and fault processing. The model introduces a notice mechanism, so that the decision process is transparent, the fault occurrence cause can be found, the self-adaptive fusion and the double-model cooperation are realized, and the accuracy of the classification of the complex faults is improved. And introducing Bayesian reasoning into SOH real-time estimation at a vehicle end, and dynamically adjusting a thermal safety threshold according to the probability so as to achieve reference control of safety risk prevention.

Inventors

• PAN MINGZHANG
• LI ZHIMING
• GUAN WEI
• WEI YUANHAI
• QIN JUNCHAO
• YE NIANYE
• MAN XINGJIA
• QIN HAIFENG
• ZHANG SONG

Assignees

• 广西大学

Dates

Publication Date

20260512

Application Date

20260127

Claims (9)

1. 1. The virtual-real fusion cloud edge cooperative double-model thermal early warning method for the automobile power battery is characterized by comprising the following steps of: Step one, integrating hardware deployment and software, installing a sparse sensor, deploying a soft sensing algorithm, and setting data acquisition conditions; step two, synchronously collecting multi-mode data, triggering a sensor array to start data sampling according to preset logic, and transmitting collected data to a preprocessing module; Step three, preprocessing the data, namely preprocessing the obtained original data, wherein the preprocessing comprises time synchronization, noise reduction processing, feature extraction and normalization; step four, constructing a perception and diagnosis model, adopting a mode of cooperative vehicle end edge calculation and cloud end deep learning, performing real-time screening by vehicle end deployment lightweight 1D CNN model, performing deep diagnosis and parameter iteration by cloud end deployment LSTM model and Bayesian SOH estimation module, Inputting the obtained data in the step 3 into a 1D CNN model to serve as a thermal state real-time classification model; The LSTM model is used as a time sequence analyzer to capture a temperature evolution rule; The attention mechanism module is used for coupling the output end of the LSTM, distributing weights for hidden states of different time steps, focusing on a key time sequence segment most relevant to the fault, and outputting fault type, development trend and key feature contribution degree; The LSTM outputs hidden state vector, then obtains attention weight distribution through normalization, uses calculated attention weight to weight sum the original hidden state sequence to generate a focusing context vector, then based on SOH estimation module of Bayesian reasoning and dynamic threshold adjustment, 1D CNN model judgment, triggering LSTM model depth analysis when abnormality occurs, fusing attention focusing information and SOH context by LSTM model, and finally outputting fault type, trend prediction, root cause analysis and SOH confidence interval; Fifthly, model training and optimization and virtual-real fusion evolution, Data acquisition, namely collecting normal working condition data, fault data, full life cycle aging and real vehicle historical data; Dividing a data set into a training set, a verification set and a test set according to the proportion of 7:2:1; Model training, namely inputting the divided data sets into a double model, training the model by using training set data, testing the model by using verification set data, stopping training when the model accuracy and recall reach set values, and completing model training; step six, thermal state analysis and fault processing, the data after pretreatment is input into a model after training is completed, The method comprises the steps of completing thermal state analysis of a power battery by a 1D CNN model, continuously monitoring the next round of sampling data if the thermal state analysis is normal, starting time sequence analysis by an LSTM model if the thermal state analysis is abnormal, matching fault reasons of abnormal features according to a temperature change trend curve by combining recent historical data, and generating a solution and sending the solution to a vehicle-mounted terminal.
2. 2. The method for thermal early warning of the virtual-real fusion cloud edge collaborative double model of the automobile power battery according to claim 1, wherein the method for installing the sparse sensor in the first step is as follows: Sparse contact type temperature measurement, namely arranging NTC temperature sensors at key thermal characteristic points of a battery module, including the center of the module, the connection position of a tab and a cooling inlet, wherein the ratio of the number of the sensors to the number of the electric cores is not lower than 1:10; reconstructing a global thermal field, namely reconstructing a virtual temperature field of a sensor region which is not arranged in a battery pack by utilizing a graph neural network GNN based on sparse NTC sampling data to generate virtual temperature data of a global battery cell; And (3) soft sensing of the cooling system, namely canceling a physical intrusion sensor in the fluid loop, and estimating the flow rate and the pipeline pressure value of the cooling liquid in real time through a Bernoulli equation of a fluid mechanics model based on the rotating speed feedback, the duty ratio signal and the inlet-outlet temperature difference of the electronic water pump.
3. 3. The virtual-real fusion cloud edge collaborative double-model thermal early warning method of the automobile power battery according to claim 1 is characterized in that the data acquisition condition in the first step is a double-mode sampling strategy combining normal monitoring and event triggering, and specifically comprises the following steps: The normal safety monitoring mode is that when the battery is in a charging and discharging state or a standing state, the system keeps continuous and uninterrupted real-time monitoring, the sampling frequency is set to be 1Hz to 10Hz, and the sampling frequency is used for capturing sudden abnormal temperature and voltage fluctuation; Triggering high-frequency synchronous sampling when any one of the following conditions is met, switching the sampling frequency to 100Hz, and transmitting data to a dual-mode for deep diagnosis: In the normal state monitoring data, when the temperature change rate of any battery cell exceeds a set threshold value, the voltage differential is abnormally enlarged or an insulation fault alarm occurs, the normal state monitoring data is immediately triggered; Condition B (periodic aging monitoring) that battery SOC is triggered once every 10% change; and C (timing inspection) is that the battery is forcedly triggered once every set time in the continuous running process of the battery.
4. 4. The method for the thermal early warning of the virtual-real fusion cloud edge collaborative double model of the automobile power battery according to claim 2, wherein the collected data comprises the temperature of a battery cell collected by an NTC temperature sensor, the temperature of air flow collected by an infrared sensor, the pressure of a pipeline collected by a pressure sensor and the flow rate of cooling liquid collected by a flow sensor, and all physical data are synchronously converted into 16-bit digital electric signals; the collected digital signals are transmitted to a preprocessing module of the battery management system through a shielding cable, and CRC cyclic redundancy check is adopted to ensure the integrity of data in the transmission process.
5. 5. The method for thermal early warning of the virtual-real fusion cloud edge collaborative double model of the automobile power battery according to claim 1, wherein the data preprocessing method in the third step is as follows: time synchronization, namely realizing precise time alignment of multi-source data by adopting an IEEE1588-2008 protocol; noise reduction treatment, namely adopting a Kalman filtering algorithm aiming at cooling liquid pressure and flow data; feature extraction is directed to time domain feature and frequency domain feature extraction: Calculating standard deviation of all cell unit temperatures and integral values of inlet and outlet temperature differences of the battery pack, wherein the integral interval is the current 10% SOC period, the unit is seed min, the maximum value and the minimum value of fluctuation amplitude of the cooling fluid pressure, When the loss or drift of the data of an individual sensor is monitored, starting a space-time correlation filling algorithm, and generating a substitution value by using the effective value of the sensor at the last moment and the current reading of a sensor adjacent to a physical position through weighted average to ensure the completeness of the feature vector of an input model; Based on driving fragment data uploaded by a vehicle end, voltage V and current I, constructing a battery equivalent circuit equation U=OCV-I×R meas , and iteratively identifying an ohmic internal resistance observation value R meas of the battery in real time; calculating standard deviation of all cell temperatures: , Wherein Ti is the temperature of the ith power core, The average temperature of the battery cell; Calculating the temperature difference integral value of the inlet and the outlet of the battery pack: , Wherein T 0 、t 1 is the start-stop time of the current 10% SOC period, and DeltaT (T) is the real-time temperature difference; calculating the fluctuation amplitude of the cooling liquid pressure: , p max 、P min is the maximum and minimum values of the pressure in the sampling period; The frequency domain characteristics are that the discrete Fourier transform is carried out on the battery cell temperature fluctuation signals of each window, and the specific formula is as follows: , Wherein x (N) is a discrete temperature fluctuation signal sequence, n=0, 1,..n-1 is a sampling point index, N is a sampling point number, k=0, 1,..n-1 is a frequency point index, j is an imaginary unit; Band energy calculation the energy corresponding to a certain band [ f a ,f b ] is K belongs to [ f a ,f b ], |X (k) | is the amplitude of the Fourier transform result; normalization by means of mean-variance normalization , Wherein, the For normalized characteristic values, X is the original characteristic data of the preprocessing stage, mu is the characteristic mean value of the bench test statistics, sigma is the standard deviation, abnormal data points exceeding mu+/-5 sigma are deleted, and the data scale is unified to the interval of [ -1,1 ].
6. 6. The method for thermal early warning of the virtual-real fusion cloud edge collaborative double model of the automobile power battery according to claim 1, wherein the method for sensing and diagnosing the thermal state based on the dual-model collaboration and the Bayesian SOH estimation is characterized by comprising the following steps: S4.1, real-time preliminary screening at a vehicle end, namely deploying a pruning and quantized 1D CNN model in a vehicle-mounted BMS chip, inputting real-time features of the model at the current moment, completing the classification judgment of normal/suspected abnormality within 50ms, discarding high-frequency data if the judgment is normal, and triggering a data uploading mechanism if the judgment is suspected abnormal; Comprises 3 layers of convolution layers, adopts convolution kernels with the sizes of 3 multiplied by 1, 5 multiplied by 1 and 7 multiplied by 1, maximally pools, 2 layers of pooling layers with the step length of 2 and 1 layer of full-connection layers, is used for capturing the space and instantaneous thermal anomaly characteristics in multiple scales, adopts a ReLU as an activation function, completes the abnormal or normal rapid judgment of the temperature of a battery cell within 50ms, , Wherein, the For the output feature map of the current 1D CNN model, For the activation function of the 1D CNN model, W (l) is the convolution kernel weight, 1D CNN B (l) is offset, ∗ represents convolution operation, and the final full-connection layer outputs two classification probabilities P normal and P abnormal ; S4.2, cloud depth diagnosis, namely, an LSTM model and an attention mechanism module are deployed on a cloud server, abnormal data packets uploaded by a vehicle end are received, a long time sequence temperature evolution rule is captured, The LSTM model is used as a time sequence analyzer, 3 hidden layers are arranged, 128 neurons in each layer are captured through forgetting gates, input gates, output gates and cell state structures, a time sequence change rule of a battery in 1 hour is captured, one hour is divided into 11 overlapped 5-minute windows by adopting a sliding window and overlapped sampling, each window is attached for 2.5 minutes, each window corresponds to 500 sampling points, and the continuous coverage of the time sequence of all 1 hour is ensured by adopting the multi-window fusion mode, and the temperature evolution rule is captured; The attention mechanism module is used for coupling the output end of the LSTM, distributing weights for hidden states of different time steps, focusing on a key time sequence segment most relevant to the fault, and outputting fault type, development trend and key feature contribution degree; LSTM outputs hidden state vectors h 1 、h 2 ......h n , each h n encodes sequence information up to time t, a trained vector u and a weight matrix are introduced into the attention mechanism, in order to calculate the relevance score e of hidden state h t of each step t and the current query t , Wherein, the Is a transpose of the learnable parameter (u) in the attention mechanism; The attention weight distribution was then obtained by normalization with a softmax function: , The weight alpha t intuitively reflects the relative importance of the information at the time t to the current fault diagnosis, the value of the relative importance is between 0 and 1, and the sum of the weights of all time steps is 1; the calculated attention weights are used to weight sum the original hidden state sequence to generate a focus context vector c: , Wherein α t is the attention weight, h t is the hidden state of each time step of the LSTM; S4.3, updating the cloud end SOH, namely, running a Bayesian SOH estimation module at the cloud end, updating the SOH by utilizing vehicle night charging data, sending the calculated dynamic threshold adjustment parameter to a vehicle returning end through OTA, and updating the judgment standard of the 1D CNN; introducing Arrhenius temperature compensation mechanism, converting identified real-time internal resistance R meas into normalized internal resistance at standard reference temperature (25deg.C) by using synchronously collected cell temperature T, , Wherein R meas is the measured internal resistance, ea is the activation energy of the battery material, k is the Boltzmann constant, T meas is the current temperature, T ref is the reference temperature of 25 ℃, A priori distribution: ~ Wherein, the The probability distribution representing the state of health of the battery, mu 0 describing the average level of SOH as an initial health benchmark, sigma 0 2 describing the degree of dispersion of SOH as an initial health discrepancy; Firstly, estimating the probability distribution of SOH according to historical data or industry experience when acquiring data; Bayesian update, updating the empirical distribution of SOH as new R ohm is obtained: , Wherein E new represents new observation data, and E old is the historically collected observation data; The method comprises the steps of constructing a full life cycle unified aging model, namely, establishing a global continuous derivable nonlinear model describing the change of internal resistance along with the health state, wherein the model adopts a form of a linear-exponential composite function, and the specific expression of the model is as follows: wherein R norm is normalized internal resistance after temperature compensation, SOH is battery health state (value range 0 to 1); Characterizing an initial internal resistance intercept of the battery when the battery leaves the factory; Representing an internal resistance linear increase component caused by SEI film steady-state growth as a linear attenuation coefficient; And (3) with The method is characterized by an exponential acceleration factor, namely an internal resistance acceleration rising component caused by a nonlinear degradation mechanism at the end of the service life of the battery; Parameters [ The ] vector is obtained through full life cycle offline training, and online fine adjustment updating is carried out by using a Recursive Least Squares (RLS) method operated by a cloud; Establishing a Bayesian network, namely setting a health state SOH as a hidden variable, and setting a normalized internal resistance R norm and a cumulative ampere-hour throughput Ah total as observation evidences; Posterior probability updating, namely when the cloud receives new effective R norm data, calculating a likelihood function P (R norm |SOH) based on a unified aging model, updating posterior probability distribution P (SOH|R norm ,Ah total ) of SOH at the current moment by using a Bayesian formula in combination with prior distribution P (SOH) t-1 at the last moment, and outputting expected value of SOH and confidence interval thereof ; Calculating an aging factor by calculating a dynamic aging factor based on the SOH expected value, ; Wherein, the Representing the degree of degradation of the battery, Representing the average value of the updated SOH; generating dynamic threshold values by the formula Calculating a threshold adjustment coefficient, wherein k adj is a preset tolerance coefficient; Coefficient of the OTA is transmitted to a vehicle-end battery management system BMS through an over-the-air technology, and a thermal fault judgment threshold value of a 1D CNN model is adjusted in real time The self-adaptive accurate early warning of batteries with different aging degrees is realized; 1D CNN fast judging, triggering LSTM depth analysis when abnormal, and fusing the attention focusing information and SOH context by LSTM, and finally outputting fault type, trend prediction, root cause analysis and SOH confidence interval.
7. 7. The method for thermal early warning of the virtual-real fusion cloud edge collaborative double model of the automobile power battery according to claim 1, wherein the data acquisition in the fifth step comprises three types of data acquisition: Collecting first normal working condition data, namely covering a combined scene of discharge multiplying power of 1C, 2C, 3C and 4C, ambient temperature of-20 ℃ to 60 ℃ and battery power of 0% -100% SOC, and collecting not less than 10 ten thousand groups; The second fault data comprise thermal runaway, cooling failure, cell short circuit and sensor fault, and each type of data is not less than 5000 groups of data; And the third full life cycle aging and the actual vehicle history data are used for collecting the actual vehicle history fault data and extracting effective history data.
8. 8. The automobile power battery virtual-real fusion cloud edge cooperative double-model thermal early warning method according to claim 7, wherein the type, the judgment basis and the treatment scheme of the actual automobile historical fault data are as follows: A. Early risk of thermal runaway: The judgment basis needs to meet the following characteristics: the temperature change rate of the core cell continuously exceeds the standard, and the duration of the state exceeds 30 seconds, The temperature difference between the electric cores is continuously enlarged, the highest electric core temperature is the highest, the difference value is in a monotonic increasing trend, Abnormal drop of the voltage platform, drop of 5% exceeding the nominal voltage occurs in the voltage of the corresponding abnormal battery cell in the stable section, The energy ratio of the abnormal cell temperature fluctuation signal at medium and high frequency is 100% or more than the historical baseline value; The method comprises the steps of immediately triggering the highest-level alarm, executing an emergency instruction through a battery management system, namely forcibly disconnecting a main loop high-voltage contactor, requesting a whole vehicle controller to enter an emergency power-down mode, starting all cooling units to the maximum power, and sending a remote alarm signal containing vehicle positioning information through a vehicle-mounted remote communication module; B. Cooling system performance degradation: The judgment basis needs to meet the following characteristics: The battery pack inlet and outlet temperature difference integral value I ΔT exceeds 150% of the dynamic safety threshold calculated from the current SOH value, The coolant flow sensor reading continues to be below 70% of the nominal design value, The system pressure fluctuation amplitude A P exceeds 2 times of standard deviation of the average value under the history normal working condition, The coolant flow or pressure signal has significant distortion or characteristic peak disappearance in the low-frequency band energy distribution which characterizes the circulation characteristic of the system, Triggering advanced early warning, wherein a control strategy is to limit the charging multiplying power to below 0.3 ℃ immediately, forcibly set the rotation speed of a cooling pump to be the highest speed for running, prompt a user to 'the efficiency of a cooling system is reduced and check timely' through a vehicle-mounted human-computer interface, and suggest a driver to find a safe place to stop and connect for maintenance if the flow rate fails to return to a normal range within the following 5 minutes; C. Cell-to-cell consistency is severely degraded or connectivity fails: The judgment basis needs to meet the following characteristics: The standard deviation of all cell temperatures shows a continuous upward trend, The maximum and minimum cell voltage differences av in the battery pack exceed a preset threshold, The internal resistance shows step-like increase and the capacity does not correspondingly decline, namely the integral ohmic internal resistance R ohm of the battery pack estimated by a direct current pulse method, The single step increase is more than 10%, but in the same period, the bayesian SOH estimation module does not feed back the corresponding capacity hopping signal, Triggering low-level early warning, suggesting that a user performs battery balance maintenance or checks the electric and mechanical connection state between battery modules when parking next time, and in the follow-up operation, suggesting that a vehicle control system reduces the continuous discharge multiplying power and avoids that the SOC value of the battery is lower than 20%; D. natural aging in healthy state: The judgment basis needs to meet the following characteristics: the SOH estimation value shows a continuous and slow descending trend that the expected health state value mu SOH <85% output by the Bayesian SOH estimation module, wherein And the width of the 95% confidence region [ mu SOH −2σ,μ SOH +2σ ] outputted by the device is converged, which shows that the estimation certainty degree is high, The internal resistance is consistent with the aging and the prediction curve based on the aging model, and no step jump as in the c-type fault occurs, Marking a disposal scheme, namely automatically generating a battery health state report by the system without triggering fault alarm, and automatically and individually adjusting each thermal protection threshold value by the system according to the calculated dynamic aging factor eta aging to carry out mild information prompt on a vehicle-computer system interface.
9. 9. The automobile power battery virtual-real fusion cloud edge cooperative double-model thermal early warning method according to claim 7, wherein the method for training the fifth model is as follows: The method comprises the steps of inputting labeling data into a double model, calculating a loss value by adopting a cross entropy loss function, carrying out back propagation by adopting an Adam optimizer, enabling the learning rate to be 0.001 and attenuate 10% every 100 rounds, adjusting convolution kernel weight and neural network connection weight parameters, testing by using a verification set every 50 rounds of training, and stopping training when the model accuracy is more than or equal to 96% and the recall rate is more than or equal to 92%, so that the generalization capability of the model is ensured.

Description

Virtual-real fusion cloud edge cooperative double-model thermal early warning method for automobile power battery Technical Field The invention relates to a fault early warning method, in particular to a virtual-real fusion cloud edge cooperative double-model thermal early warning method for an automobile power battery, and belongs to the technical field of power batteries. Background With the development of new energy automobiles, a power battery is used as a core power source of the new energy automobiles, and the battery is aged for a long time in charge and discharge cycles, so that the problems of uneven temperature distribution, local heat accumulation and the like of the battery are easy to occur. If the battery problems in the new energy automobile are not found in time, the problems of thermal runaway, cooling liquid failure, battery core short circuit and the like can be caused, and riding safety is seriously endangered. The traditional method cannot adapt to the attenuation of the health of the battery, can possibly cause false alarm of the fault of the battery, lacks uncertainty of the estimation of the health of the battery, causes inaccurate decision, and is difficult to timely realize early warning and trend prediction of the fault. Chinese patent CN120481800A discloses a method and a device for monitoring a vehicle fuel cell, which are characterized in that the scheme is used for analyzing and predicting the operation parameters of the fuel cell, extracting the characteristics of alarm data of the fuel cell by using a 1D-CNN model, establishing and training a diagnosis model, extracting the characteristics of the vehicle data by using an LSTM model, establishing and training an early warning model according to the extracted characteristics, inputting the operation data of the fuel cell collected in real time into the model for analysis, and controlling fuel cell vehicle equipment in real time according to the analysis result. The proposal analyzes the overall operation parameters of the fuel cell, can not find out the problem of uneven local heat accumulation of the internal temperature distribution of the power cell in time, and takes corresponding measures. Disclosure of Invention The application aims to provide a virtual-real fusion cloud edge cooperative double-model thermal early warning method for an automobile power battery, aiming at the problems in the prior art, and the application is based on a 1D CNN model and LSTM dual-model cooperative framework, combines multi-element data acquisition and fusion, constructs a complete closed loop system from data acquisition to intelligent decision on the basis of a notice mechanism and a battery health state module, can realize self-adaptive battery aging, and remarkably improves the safety and reliability of a battery system. The technical scheme is that the virtual-real fusion cloud edge cooperative double-model thermal early warning method for the automobile power battery comprises the following steps of: Step one, integrating hardware deployment and software, installing a sparse sensor, deploying a soft sensing algorithm, and setting data acquisition conditions; step two, synchronously collecting multi-mode data, triggering a sensor array to start data sampling according to preset logic, and transmitting collected data to a preprocessing module; Step three, preprocessing the data, namely preprocessing the obtained original data, wherein the preprocessing comprises time synchronization, noise reduction processing, feature extraction and normalization; step four, constructing a perception and diagnosis model, adopting a mode of cooperative vehicle end edge calculation and cloud end deep learning, performing real-time screening by vehicle end deployment lightweight 1D CNN model, performing deep diagnosis and parameter iteration by cloud end deployment LSTM model and Bayesian SOH estimation module, Inputting the obtained data in the step 3 into a 1D CNN model to serve as a thermal state real-time classification model; The LSTM model is used as a time sequence analyzer to capture a temperature evolution rule; The attention mechanism module is used for coupling the output end of the LSTM, distributing weights for hidden states of different time steps, focusing on a key time sequence segment most relevant to the fault, and outputting fault type, development trend and key feature contribution degree; The LSTM outputs hidden state vector, then obtains attention weight distribution through normalization, uses calculated attention weight to weight sum the original hidden state sequence to generate a focusing context vector, then based on SOH estimation module of Bayesian reasoning and dynamic threshold adjustment, 1D CNN model judgment, triggering LSTM model depth analysis when abnormality occurs, fusing attention focusing information and SOH context by LSTM model, and finally outputting fault type, trend prediction, root cause analysis and SOH conf