Lithium ion battery temperature anomaly detection and fault diagnosis method based on modal decomposition and infrared imaging

Abstract

The invention discloses a lithium ion battery temperature anomaly detection and fault diagnosis method based on modal decomposition and infrared imaging, and belongs to the technical field of battery management. According to the method, basic operation data such as voltage, current and temperature of the lithium ion battery are collected, and an infrared imaging device is utilized to obtain a temperature distribution diagram of the surface of the battery, so that the temperature information of the battery is captured. VMD (variational mode decomposition) and BEMD (two-dimensional empirical mode decomposition) are adopted to perform mode decomposition on acquired temperature signals, and mode signals highly related to battery temperature abnormity or fault modes are extracted. Then, a CNN-LSTM model constructed by using a convolutional neural network (CNN) and a long short-term memory (LSTM) network is used for carrying out time sequence modeling on the extracted features, and hyper-parameters of the model are optimized through a particle swarm optimization (PSO) method; and finally, deploying the trained CNN-LSTM model into embedded equipment to realize real-time monitoring and fault diagnosis of the temperature state of the lithium ion battery.

Description

Technical Field

[0001] The present invention relates to the field of power battery management, and specifically to a method for detecting temperature anomalies and diagnosing faults of lithium-ion batteries based on modal decomposition and infrared imaging. Background Art

[0002] Due to their high energy density, long lifespan, and environmental friendliness, lithium-ion batteries are widely used in electric vehicles and energy storage systems. However, during operation, the performance and safety of batteries are significantly affected by environmental temperature and internal thermal runaway. In particular, abnormal temperatures can lead to battery capacity degradation, shortened lifespan, and even serious safety accidents. Therefore, monitoring and diagnosing the temperature state (SOT) of batteries during operation can not only prevent the occurrence of battery thermal runaway but also ensure the stability and safety of the system.

[0003] Existing fault diagnosis techniques mainly include methods based on physical models, empirical rules, and machine learning. Methods based on physical models infer the battery state by establishing an equivalent circuit model or thermal model of the battery and using measured parameters such as voltage and temperature. Such methods have a strong theoretical basis but require high precision for the model, and the modeling process is complex. Methods based on empirical rules judge whether the battery has a fault according to pre-set rules or thresholds. Such methods are simple to implement but have poor robustness and are difficult to cover complex fault modes, especially unable to handle non-linear and multi-fault coupling situations. Among methods based on machine learning, neural network models have been widely used in the fault diagnosis technology of lithium-ion batteries due to their excellent self-learning, non-linear fitting, and other capabilities. Although high-performance battery fault diagnosis can be achieved using neural network models, there are still some problems. For example, there is significant noise interference in complex signals, and it is impossible to accurately extract fault modes related to temperature, which poses higher requirements for the intelligence of the battery management system. Existing models usually adopt a cascaded structure of a convolutional neural network (CNN) and a long short-term memory network (LSTM), using CNN to extract local features and LSTM to capture the dynamic relationships in time series. However, the features extracted by CNN may cause information leakage, affecting LSTM's learning of time dependence and reducing the diagnostic accuracy. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the problem to be solved by the present invention is to provide a method for detecting temperature anomalies and diagnosing faults of lithium-ion batteries based on modal decomposition and infrared imaging.

[0005] The present invention adopts the following technical solution to solve the above-mentioned technical problems:

[0006] A method for detecting temperature anomalies and fault diagnosis of lithium-ion batteries based on modal decomposition and infrared imaging, the method comprising the following steps:

[0007] Step 1, real-time data acquisition and infrared imaging acquisition;

[0008] Step 2, modal decomposition and feature selection;

[0009] Use variational mode decomposition to decompose signals such as voltage, current, and temperature collected by traditional sensors; use two-dimensional empirical mode decomposition to decompose the image signals collected by infrared imaging;

[0010] Step 3, correlation analysis and feature optimization;

[0011] Calculate the correlation coefficients between each modal signal using the Pearson correlation coefficient, and select the modal signals that have a significant linear relationship with the battery temperature anomaly or fault mode as the input features for subsequent model training, that is, select the modal signals with an absolute value of the Pearson correlation coefficient with the battery core temperature above 0.05;

[0012] Step 4, CNN-LSTM model construction and training;

[0013] Step 5, SOT estimation and fault diagnosis;

[0014] Based on the temperature state and change trend of the battery, different types of faults are identified in a timely manner; the temperature state of the battery is evaluated by combining the estimation results obtained by using the CNN-LSTM model in real time with the infrared imaging data;

[0015] Step 6, deploy the trained CNN-LSTM model to an embedded device. The system monitors the battery temperature state in real time. If an anomaly or fault is detected, the fault is classified and the corresponding level alarm is triggered in a timely manner.

[0016] Further, in the above step 2, variational mode decomposition decomposes the signal x(t) into several modal signals , each modal signal having a different center frequency , expressed as:

[0017]

[0018] where the balance factor controls the balance between the signal and the mode;

[0019] By using the cross - validation method, the dataset is divided multiple times to evaluate the impact of different parameter combinations on the model performance, and the parameter combination with the minimum mean squared error is selected as the final modal decomposition parameter;

[0020] In two - dimensional empirical mode decomposition, the image signal is decomposed into multiple modal signals with different spatial frequency characteristics , and the optimization objective is to minimize the error after decomposition:

[0021]

[0022] The cross - validation method divides the image data into multiple subsets, trains and validates under different parameter settings, evaluates the effect of each group of parameters, and selects the best parameter combination by calculating the mean squared error. 5. Further, in step 3, the calculation formula of the Pearson correlation coefficient is as follows:

[0023]

[0024] where is the Pearson correlation coefficient between modal signals X and Y; and are the i - th data points in signals X and Y respectively; and are the means of signal X and signal Y respectively; n is the total number of data points;

[0025] The value range of the Pearson correlation coefficient is [-1, 1]; when , it indicates that the two signals are completely positively correlated; when , it indicates that the two signals are completely negatively correlated; when , it indicates that there is no linear relationship between the two signals.

[0026] Further, the construction and training of the CNN - LSTM model include: using a convolutional neural network to extract features from the modal signals after modal decomposition; using a long short - term memory network to perform time - series modeling on the extracted features; and improving the accuracy and training efficiency of the model through hyperparameter optimization.

[0027] Further, the construction and training of the CNN - LSTM model include the following sub - steps:

[0028] S401: CNN feature extraction

[0029] The structure of the CNN includes an input layer, multiple convolutional layers, pooling layers, and fully - connected layers. The entire model is trained using the Adam optimizer and optimized through the MSE loss function;

[0030] S402: LSTM time - series modeling

[0031] The spatial features extracted by the CNN will be passed as input to the LSTM layer; during the training process, the LSTM model will adjust its weights through the backpropagation algorithm to optimize by minimizing the loss function MSE;

[0032] S403: Hyperparameter Optimization

[0033] The search range of CNN-related hyperparameters includes the number of convolutional layers from 1 to 3, the number of convolutional kernels in each convolutional layer from 10 to 60, the convolutional kernel size from 1 to 10, and the convolutional stride from 1 to 10; LSTM-related hyperparameters include the number of LSTM layers from 1 to 3, the number of units in each layer from 10 to 150, and the dropout rate from 0.01 to 0.2. The number of units in the fully connected layer is from 10 to 150, the learning rate of the Adam optimizer is from 0.00001 to 0.002; the number of training epochs is from 100 to 300, and the batch size is from 10 to 20. PSO optimizes the hyperparameters through iteration, and the loss function in the optimization process is MSE.

[0034] Furthermore, the system deployment and real-time monitoring include: real-time collection of multi-dimensional data of the battery, including the dynamic changes of temperature, current, and voltage, as well as the temperature distribution map captured by infrared thermal imaging; data preprocessing, including normalization, denoising processing, and extraction of hot spot features from infrared data; batch feeding the preprocessed data into the model for calculation, and outputting the current SOT of the battery and the prediction of future temperature changes.

[0035] The method specifically includes the following steps:

[0036] Step 1, Real-time data collection and infrared imaging acquisition. Use traditional temperature sensors to obtain basic operating data such as voltage, current, and temperature of lithium-ion batteries, and equip infrared imaging devices to obtain the temperature distribution map on the battery surface to help detect local overheating or hot spot areas on the battery surface;

[0037] Step 2, Modal decomposition and feature selection. Perform modal decomposition on the collected temperature signals (including traditional sensor and infrared imaging data), and decompose the complex time series signals into multiple modal signals (IMFs). Among them, signals such as voltage, current, and temperature collected by traditional sensors are decomposed by Variational Mode Decomposition (VMD), and image signals collected by infrared imaging are decomposed by Bidimensional Empirical Mode Decomposition (BEMD).

[0038] Step 3: Correlation analysis and feature optimization. Perform correlation analysis on the decomposed IMFs signals, calculate the correlation coefficients between each modal signal using the Pearson correlation coefficient, and select the modal signals that are highly correlated with the abnormal battery temperature or fault mode as the input features for subsequent model training.

[0039] Step 4: CNN-LSTM model construction and training. Use CNN to extract features from the modal signals after modal decomposition, and use LSTM to perform time series modeling on the extracted features. Train the model and improve the accuracy and training efficiency of the model through hyperparameter optimization;

[0040] Step 5: SOT estimation and fault diagnosis. Use the trained CNN-LSTM model to estimate the battery SOT in real time and identify possible abnormalities or faults. The SOT estimation results will be used as input, combined with infrared imaging data, to judge the health status of the battery in real time. Based on the SOT and temperature status, design a fault diagnosis module to classify the fault levels of the battery;

[0041] Step 6: Deploy the trained CNN-LSTM model to the embedded device. The system monitors the battery temperature status in real time. If an abnormality or fault is detected, classify the fault and trigger the corresponding level alarm in a timely manner.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. The present invention combines traditional temperature sensors and infrared imaging technology, and can simultaneously obtain the internal and external temperature information of the battery. By fusing these two types of data, the system can more comprehensively capture battery temperature abnormalities and fault modes, especially with higher sensitivity to surface hot spots and local overheating, improving the accuracy of fault diagnosis.

[0044] 2. The modal decomposition technology is used to process complex battery temperature signals, which can decompose the original signal into modal functions of multiple frequency bands, effectively extracting frequency features related to faults. The CNN-LSTM model combines the local feature extraction ability of the convolutional neural network and the time series modeling ability of the long short-term memory network, further enhancing the accuracy and time sensitivity of temperature anomaly prediction.

[0045] 3. The present invention designs a multi-level fault diagnosis module based on SOT estimation, which can timely identify different types of faults according to the temperature status and change trend of the battery, and automatically adjust the response strategy according to the severity of the faults. This enables the system to timely detect potential faults and quickly take necessary safety measures when the battery is in danger such as overheating or thermal runaway, ensuring the safety of the battery and users.

[0046] 4. Optimize the hyperparameters of the CNN-LSTM model through the Particle Swarm Optimization (PSO) method, enabling the model to automatically adapt to different battery states and temperature change patterns during training. This optimization not only improves the training efficiency of the model but also enhances the prediction accuracy, especially when dealing with complex datasets. Brief Description of the Drawings

[0047] Figure 1 It is a flowchart of the offline training of the SOT estimation and fault classification diagnosis model of the present invention.

[0048] Figure 2 It is the result of Pearson correlation analysis.

[0049] Figure 3 It is a flowchart of the construction and verification of the SOT estimation model of the present invention.

[0050] Figure 4 It is a flowchart of the integrated system of the SOT estimation and fault classification diagnosis model of the present invention. Specific Implementation Method

[0051] The following will describe in detail the specific implementation manner of a lithium-ion battery temperature anomaly detection and fault diagnosis method based on modal decomposition and infrared imaging. The embodiments are only for further illustration of the present invention, and the specific implementation steps should not be construed as a limitation on the technical scope of the present invention.

[0052] Step 1: Real-time data acquisition and infrared imaging acquisition

[0053] Place the lithium-ion battery under specific ambient temperature conditions to ensure that the operating temperature of the battery is consistent with the external environment. Let the battery stand still in the environment until the battery temperature stabilizes to the ambient temperature, ensuring that the initial temperature of the battery is uniform and there is no influence of external temperature fluctuations. Charge the battery completely according to the battery charge and discharge parameters, and the charging process is carried out according to the charging curve recommended by the manufacturer to ensure the charging behavior of the battery under standard operating conditions. After charging, use a load with a certain power to discharge the battery until the battery voltage drops to the predetermined discharge cut-off voltage to complete a charge / discharge cycle. During the entire charge and discharge cycle, real-time collect the basic operating data of the battery, including voltage, current, power and other data.

[0054] Evenly distribute thermocouples on the battery surface to monitor the temperature of the battery surface, and at the same time equip with an infrared thermal imaging device to obtain the temperature distribution map of the battery surface in real time. Infrared imaging can accurately detect local overheating areas or hot spots that may exist on the battery surface, and assist in judging the thermal behavior during the operation of the battery.

[0055] Sensor data and infrared imaging data are synchronized through time alignment technology to ensure the consistency of data from different sources in the time dimension. Preprocess different data, and use the min-max normalization method for unified size conversion to normalize the data between 0 and 1:

[0056]

[0057] Where X is the original data, and are the minimum and maximum values in the dataset respectively, is the normalized data.

[0058] In this embodiment, the sampling frequency is 1 Hz; the ambient temperatures are -10°C, 0°C, 10°C, 20°C, and 30°C to cover the ambient temperatures during the actual use of lithium-ion batteries; the power configurations include BJDST driving cycle, DST driving cycle, US06 driving cycle, FUDS driving cycle, and CLTC driving cycle; the training set is used to train the SOC estimation model, accounting for 60% of the entire dataset; the validation set is used to optimize the trained model, accounting for 10% of the entire dataset; the test set is used to verify the reliability of the model, accounting for 30% of the entire dataset.

[0059] Step 2, Modal decomposition and feature selection

[0060] After data acquisition, perform modal decomposition on the obtained temperature signals to extract effective features therefrom. For the voltage, current, and temperature signals collected by traditional sensors, use VMD to decompose the complex time series signals into multiple IMFs, and these modal signals can capture the change patterns in different frequency ranges. For the image signals collected by infrared imaging, use BEMD to decompose the image signals into multiple IMFs with different spatio-temporal characteristics. BEMD can effectively process spatial information and capture the changes in the hot spot areas in the image.

[0061] The cross-validation method is used to select the optimal parameters in VMD and BEMD to ensure that the decomposed modal signals can effectively reflect the temperature change characteristics of the battery. For VMD, the main adjustment parameters include the number of modes K and the balance factor α, and these parameters have an important impact on the effect of modal decomposition. In VMD, the goal is to decompose the signal x(t) into several modal signals Each modal signal has a different center frequency . Its optimization problem is usually expressed as:

[0062]

[0063] Among them, the balance factor α controls the balance between the signal and the mode. Through the cross-validation method, the dataset can be divided multiple times to evaluate the impact of different parameter combinations on the model performance, so as to select the optimal parameter configuration.

[0064] In the example, the dataset is divided into K subsets. Each time, K - 1 subsets are selected for training, and the remaining subset is used for validation. For each round of training and validation, different parameter combinations are used to decompose the signal, and the effect of each parameter combination is evaluated by calculating the performance of the model on the validation set. The evaluation metric often uses the mean square error (MSE), and its calculation formula is:

[0065]

[0066] Among them, is the actual value, is the predicted value, and N is the number of samples. By comparing the MSE under different parameter configurations, the parameter combination that minimizes the MSE is selected as the final mode decomposition parameter.

[0067] For BEMD, the cross-validation method is also applied to select the optimal parameters, such as the number of modes K and the number of decomposition times. BEMD aims to decompose the image signal into multiple mode signals with different spatial frequency characteristics , and the optimization goal is to minimize the error after decomposition:

[0068]

[0069] Similarly, the cross-validation method divides the image data into multiple subsets, conducts training and validation under different parameter settings, evaluates the effect of each group of parameters, and selects the best parameter combination by calculating the MSE.

[0070] Step 3, Correlation Analysis and Feature Optimization

[0071] After mode decomposition, the multiple obtained IMFs contain different frequency and spatio-temporal characteristics of the signal. In order to extract the key feature signals related to battery temperature anomalies, fault modes or performance degradation from them, correlation analysis is required to screen out the signals highly correlated with the battery temperature parameters as the input features of the subsequent model.

[0072] In this example, the Pearson correlation coefficient is used to measure the linear relationship between different mode signals. The calculation formula of the Pearson correlation coefficient is as follows:

[0073]

[0074] Among them, is the Pearson correlation coefficient between mode signals X and Y; and are the i-th data points in signals X and Y respectively; and are the means of signal X and signal Y respectively; n is the total number of data points. The value range of the Pearson correlation coefficient is [-1, 1]. When

[0075] it indicates that the two signals are completely positively correlated; when it indicates that the two signals are completely negatively correlated; when it indicates that there is no linear relationship between the two signals.

[0076] In this example, the moving window averages of key parameters including voltage, current, cumulative charge, power, energy, surface temperature, and core temperature are calculated. Subsequently, a large amount of battery performance data is read and integrated from multiple datasets to form a comprehensive dataset. During the data integration process, the consistency of data types is ensured, and the missing values are processed using the forward filling method. Finally, the Pearson correlation coefficient matrix between the parameters in the dataset is calculated, and the specific correlation coefficient results are as Figure 2 shown. It can be seen from the figure that voltage, current, surface temperature, and their averages, average voltage, average current, and average temperature have a higher correlation with the battery core temperature among the battery parameters. Therefore, voltage, current, average voltage, average current, surface temperature, and average surface temperature are selected as the input parameters of the battery estimation model.

[0077] Step 4: Construction and training of the CNN-LSTM model

[0078] S401: CNN feature extraction

[0079] The structure of the CNN includes an input layer, multiple convolutional layers, pooling layers, and fully connected layers. The input layer receives the signal or image data after modal decomposition. The convolutional layers use multiple filters for local feature extraction, and the convolutional kernel size, the number of filters, and the stride are adjusted in the subsequent hyperparameter optimization process. Each convolutional layer is followed by a pooling layer, usually using the max pooling operation, which is used to reduce the data dimension and extract the most significant features. The number of neurons and the number of layers in the fully connected layer will also be determined according to the hyperparameter optimization. Finally, the output layer performs regression or classification according to the specific task. The entire model is trained using the Adam optimizer and optimized through the MSE loss function.

[0080] S402: LSTM time series modeling

[0081] The spatial features extracted by the CNN will be passed as input to the LSTM layer. The number of LSTM layers and the number of units in each LSTM layer will be adjusted during the hyperparameter optimization process. Each LSTM unit contains an input gate, a forget gate, and an output gate. Through these gating mechanisms, the LSTM can selectively remember or forget information based on the current input and historical state, thereby capturing temporal dependencies. After the last LSTM layer, the output will be processed through a fully connected layer to generate the final prediction result. During the training process, the LSTM model will continuously adjust its weights through the backpropagation algorithm to minimize the loss function MSE, so as to accurately estimate the temperature dynamics and potential fault modes of the battery.

[0082] S403: Hyperparameter Optimization

[0083] During the hyperparameter optimization process, the Particle Swarm Optimization (PSO) searches a preset hyperparameter space. The search ranges for CNN-related hyperparameters include the number of convolutional layers (1 - 3), the number of convolutional kernels in each convolutional layer (10 - 60), the convolutional kernel size (1 - 10), and the convolutional stride (1 - 10). LSTM-related hyperparameters include the number of LSTM layers (1 - 3), the number of units in each layer (10 - 150), and the dropout rate (0.01 - 0.2). The number of units in the fully connected layer (10 - 150), the learning rate of the Adam optimizer (0.00001 - 0.002). The number of training epochs is 100 - 300, and the batch size is 10 - 20. PSO optimizes the hyperparameters within these ranges through multiple iterations to find the best combination that can improve the model performance. The loss function during the optimization process is MSE, and the hyperparameter configuration for each iteration is evaluated by calculating the loss value on the validation set.

[0084] S404: CNN-LSTM Model Training and Validation

[0085] After the CNN-LSTM model is constructed and the hyperparameters are optimized, training begins. The dataset is divided into a training set, a validation set, and a test set, where the training set accounts for 60%, the validation set accounts for 10%, and the test set accounts for 30%. During the training process, the CNN extracts spatial features, and the LSTM performs temporal modeling to capture the dynamic changes in the battery temperature. Through the Adam optimizer and the MSE loss function, the model learns on the training set and evaluates its performance using the validation set after each epoch. The results of the validation set help optimize the hyperparameters to ensure the accuracy and efficiency of the model. Finally, after training and tuning, the model is validated on the test set to evaluate its generalization ability and the accuracy of SOC estimation.

[0086] Step 5: State of Temperature (SOT) Estimation and Fault Diagnosis

[0087] Using a trained CNN-LSTM model, the battery's SOT is estimated in real time. The SOT estimate is combined with infrared imaging data for a more accurate temperature assessment. By analyzing the battery surface temperature distribution, potential hot spots and localized overheating can be identified.

[0088] Based on the SOT estimation results and infrared imaging data, a fault diagnosis module was designed to classify the battery's health status. This module categorizes battery faults into different levels. In a Level 1 fault (minor abnormality, SOT between 50 and 80 °C), the battery begins to exhibit minor abnormalities. The SOT estimation shows a slight temperature increase above the average level, but the overall fluctuation is small (within ±5 °C). Infrared thermal imaging may detect localized minor hot spots, but these are small and discontinuous. Gas sensors detect no abnormalities. In a Level 2 fault (moderate abnormality, SOT between 80 and 110 °C), internal battery reactions intensify, leading to self-heating. The SOT estimation shows a significant temperature increase, with local hot spots 10%-15% above the overall average temperature. Infrared thermal imaging clearly reveals several hot spots, with a distinct temperature gradient between the hot spots and the surrounding area. Gas detection begins to detect small amounts of decomposed CO2 and H2 gases. In a Level 3 fault (serious abnormality, SOT between 110 and 140 °C), internal battery reactions further deteriorate, and the separator begins to soften or partially melt. SOT estimates indicate rapid temperature fluctuations, with local hot spots approximately 30% above the overall average. Infrared thermal imaging shows an increase in the number of hot spots, a widening distribution, and a trend toward contiguous areas. Gas detection indicates a significant increase in emissions of harmful gases such as CO and H2, with some gas leaks. A Level 4 fault (thermal runaway, SOT > 140°C) involves a violent reaction within the battery, completely melting the separator. The temperature rises exponentially, becoming difficult to control. Infrared thermal imaging shows the entire surface covered in hot spots, which are continuous and exceed the device's range. Gas sensors detect significant CO, HF, and CH4 emissions, and the battery casing is swollen or ruptured.

[0089] Step 6: System deployment and real-time monitoring

[0090] After the model is deployed, the system collects multidimensional battery data in real time, including dynamic changes in temperature, current, and voltage, as well as temperature distribution maps captured by infrared thermal imaging. This data is collected regularly at a preset frequency and immediately preprocessed. This process includes normalization, denoising, and extracting hotspot features from the infrared data, such as temperature gradients and regional distribution. This preprocessed data is then fed into the model in batches for calculation.

[0091] During operation, the model extracts the spatial features of the infrared thermal imaging data through CNN to identify the hot spot distribution and temperature peak. Subsequently, the LSTM module receives these spatial features and time series data such as temperature, current, and voltage to capture the dynamic changes of the battery. LSTM uses a gating mechanism to process historical information and finally outputs the current SOT of the battery and the prediction of future temperature changes. The model results are fed back to the monitoring system for comprehensive analysis with real-time data to determine whether there is a potential failure risk in the battery.

[0092] Matters not described in the present invention are applicable to the prior art. The above embodiments are only used to illustrate the present invention, and any equivalent transformation and improvement made on the basis of the technical solution of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. A method for detecting abnormal temperature and diagnosing faults of a lithium-ion battery, characterized in that, It includes the following steps: Step 1, real-time data acquisition and infrared imaging acquisition; Step 2, modal decomposition and feature selection; Use variational modal decomposition to decompose signals such as voltage, current, and temperature collected by traditional sensors; use two-dimensional empirical modal decomposition to decompose the image signals collected by infrared imaging; Step 3, correlation analysis and feature optimization; Calculate the correlation coefficients between each modal signal using the Pearson correlation coefficient, and select the modal signals with a significant linear relationship with the abnormal battery temperature or fault mode as the input features for subsequent model training, that is, select the modal signals with the absolute value of the Pearson correlation coefficient with the core battery temperature above 0.05; Step 4, CNN-LSTM model construction and training; Step 5, SOT estimation and fault diagnosis; Identify different types of faults in a timely manner based on the temperature state and change trend of the battery; the temperature state of the battery is evaluated by combining the estimation results obtained by real-time estimation using the CNN-LSTM model with the infrared imaging data; Step 6, deploy the trained CNN-LSTM model to the embedded device. The system monitors the battery temperature state in real time. If an anomaly or fault is detected, classify the fault level and trigger the corresponding level alarm in a timely manner.

2. The method according to claim 1, characterized in that: In step 2, variational mode decomposition is performed to decompose the signal x(t) into a number of modal signals , each of which has a different center frequency , which is expressed as: ； Among them, the balance factor controls the balance between the signal and the mode; Through the cross-validation method, divide the dataset multiple times, evaluate the impact of different parameter combinations on the model performance, and select the parameter combination with the smallest mean square error as the final modal decomposition parameter; In two-dimensional empirical mode decomposition, an image signal is decomposed into multiple modal signals with different spatial frequency characteristics , and the optimization objective is to minimize the error after decomposition: ； The cross-validation method divides the image data into multiple subsets, trains and validates under different parameter settings, evaluates the effects of each group of parameters, and selects the best parameter combination by calculating the mean square error.

3. The method according to claim 1, characterized in that: In the said Step 3, the calculation formula of the Pearson correlation coefficient is as follows: ； Among them, is the Pearson correlation coefficient between the modal signals X and Y; and are the i-th data points in signals X and Y respectively; and are the means of signal X and signal Y respectively; n is the total number of data points; The value range of the Pearson correlation coefficient is [-1, 1]; when it indicates that the two signals are completely positively correlated; when it indicates that the two signals are completely negatively correlated; when it indicates that there is no linear relationship between the two signals.

4. The method according to claim 1, wherein: The construction and training of the CNN-LSTM model includes: using a convolutional neural network to extract features from the modal signals after modal decomposition; using a long short-term memory network to perform time series modeling on the extracted features; improving the accuracy and training efficiency of the model through hyperparameter optimization.

5. The method according to claim 4, wherein: The construction and training of the CNN-LSTM model includes the following sub-steps: S401: CNN feature extraction The structure of the CNN includes an input layer, multiple convolutional layers, pooling layers, and fully connected layers. The entire model is trained using the Adam optimizer and optimized through the MSE loss function; S402: LSTM time series modeling The spatial features extracted by the CNN will be passed as input to the LSTM layer; During the training process, the LSTM model will adjust its weights through the backpropagation algorithm to optimize by minimizing the loss function MSE; S403: Hyperparameter optimization The search range of CNN-related hyperparameters includes the number of convolutional layers from 1 to 3, the number of convolutional kernels in each convolutional layer from 10 to 60, the convolutional kernel size from 1 to 10, and the convolutional stride from 1 to 10; the LSTM-related hyperparameters include the number of LSTM layers from 1 to 3, the number of units in each layer from 10 to 150, and the dropout rate from 0.01 to 0.

2. The number of units in the fully connected layer is from 10 to 150, and the learning rate of the Adam optimizer is from 0.00001 to 0.002; the number of training epochs is from 100 to 300, and the batch size is from 10 to 20. PSO optimizes the hyperparameters through iteration, and the loss function in the optimization process is MSE.

6. The method according to claim 1, wherein The system deployment and real-time monitoring include: real-time collection of multi-dimensional data of the battery, including the dynamic changes of temperature, current, and voltage, as well as the temperature distribution map captured by infrared thermal imaging; data preprocessing, including normalization, denoising, and extraction of hot spot features from infrared data; feeding the preprocessed data into the model in batches for calculation, and outputting the current SOT of the battery and the prediction of future temperature changes.

Images