An AI-based multi-modal feature fusion-based power battery thermal runaway precise early warning system

Abstract

The application discloses a kind of AI multi-modal feature fusion-based power battery thermal runaway precision early warning system, and multiple-source data acquisition module obtains battery electric parameter and thermal distribution data, and wavelet packet decomposition and adaptive platform histogram equalization are used to pre-process;AI feature fusion module uses wavelet packet decomposition and improved RepLKNet network cascade to extract multi-channel time-frequency feature, uses bidirectional gate recurrent unit and multi-head global attention mechanism to extract time series feature, realizes adaptive cross-modal fusion by multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction, and intelligent decision early warning module uses reinforcement learning framework to dynamically adjust multi-classifier decision weight, and adaptively adjusts early warning threshold;Cloud iterative optimization module uses momentum federal average algorithm to carry out efficient model iteration and OTA push, and the application realizes the comprehensive extraction and depth fusion of multidimensional fault feature, the intelligent decision of whole life cycle adaptation and efficient cloud collaborative iteration.

Description

Technical Field

[0001] This invention belongs to the field of power battery safety technology, and in particular relates to a precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion. Background Technology

[0002] As the core energy storage component of new energy vehicles, the safety performance of power batteries directly affects the reliability of vehicle operation and the safety of occupants. Battery thermal runaway is the most serious form of safety accident for power batteries. Its occurrence mechanism involves multi-physics field coupling of electrochemistry, thermodynamics, and mechanical stress, and the fault evolution path is hidden and has significant nonlinear characteristics. In the very early stage before thermal runaway occurs, precursor signals such as abnormal lithium-ion migration rate, increased local internal resistance, and heat release caused by micro-short circuits appear inside the battery. These signals are characterized to varying degrees in three dimensions: electrical parameters, thermal distribution, and temporal evolution. However, because the amplitude of the precursor signals of thermal runaway is weak and the duration is short in the very early stage, and there are complex nonlinear coupling relationships between signals in multiple dimensions, how to comprehensively extract fault-related features from multi-source sensor data, achieve effective deep fusion of different modal features, and maintain stable decision-making performance throughout the entire battery life cycle has become the core challenge in this technical field. In existing technologies, early warning systems mostly rely on single-dimensional monitoring of electrical parameters or simple stitching and fusion of electrical parameters and thermal imaging data. This fails to fully exploit the complementarity between time-frequency domain features and temporal evolution patterns, resulting in insufficient ability to identify very early, weak fault signals. Furthermore, as the number of battery charge-discharge cycles accumulates, battery health continuously deteriorates, and parameter baselines drift. If the early warning decision-making mechanism cannot adaptively adjust, the risk of missed detections in the later stages of battery aging will significantly increase.

[0003] Currently, power battery thermal runaway early warning technology is developing along three main technical routes. The first route is a threshold-based early warning method based on traditional battery management system parameters. This method sets upper and lower limits for single or multiple parameters such as voltage, current, and temperature, triggering an alarm when the monitored value exceeds a preset range. This method is simple to implement, provides a direct response, and is widely deployed in battery management systems. The second route is a feature-based early warning method incorporating traditional machine learning algorithms. It uses models such as support vector machines, random forests, or shallow neural networks, employing historical battery operation data as a training set to extract manually designed features such as voltage standard deviation, temperature change rate, and internal resistance increment, training a classifier to distinguish between fault and normal states. This method has stronger pattern recognition capabilities than threshold methods and can handle nonlinear parameter changes to some extent. The third route is an advanced early warning method combining deep learning technology. This method uses convolutional neural networks to extract abnormal thermal distribution features from thermal imaging images, or recurrent neural network structures such as long short-term memory networks and gated recurrent units to extract long-term dependencies in the temporal changes of electrical parameters. Some solutions attempt to concatenate and fuse image features with temporal features, weighting different feature dimensions through an attention mechanism. In addition, to improve the continuous adaptability of the model, some systems have deployed a model update mechanism based on a federated learning framework in the cloud, using multi-terminal data for centralized model fine-tuning. These technologies have played a certain role in improving thermal runaway early warning in specific scenarios and constitute the main technical landscape in the current field of battery risk monitoring.

[0004] The aforementioned existing technologies reveal the following key shortcomings in practical applications. First, the coverage dimension and analytical depth of feature extraction are insufficient. Thresholding methods only focus on the instantaneous amplitude of parameters and cannot capture changes in energy distribution in the time-frequency domain; traditional machine learning relies on manual feature engineering, and feature design depends on expert experience, making it difficult to exhaustively enumerate all potential fault modes; while existing deep learning methods can automatically extract features, most schemes shallowly stitch the time-domain features of electrical signals with the spatial features of thermal imaging, without performing multi-scale decomposition preprocessing of the electrical signals in the frequency domain or time-frequency domain, resulting in weak fault energy hidden in specific frequency bands being submerged by noise. Second, the fusion mechanism for heterogeneous modal features is too crude. Simple feature vector concatenation or fixed-weighted summation cannot reflect the differentiated contributions of different modal features at different stages of the fault, lacking the ability to adaptively adjust the fusion ratio. In the early stages of the fault, when electrical parameter changes are subtle and the thermal distribution is not yet obviously abnormal, the fusion effect is poor. Third, the decision-making mechanism lacks the ability to adapt to changes in the battery's state throughout its entire life cycle. Existing solutions employ classification models with fixed thresholds or static learning, failing to consider the continuous degradation of battery health in the decision-making process. As the battery enters the aging stage, parameter baseline drift blurs the boundary between fault and normal features, and the fixed decision boundary leads to an increase in the false negative rate. Fourth, cloud-based model aggregation and update efficiency is low, and simple averaging algorithms have slow convergence speeds, hindering rapid iterative optimization of the model in multi-terminal scenarios. Therefore, systematic improvements are urgently needed in areas such as multi-scale feature decomposition preprocessing, adaptive cross-modal deep fusion, dynamic decision-making based on health status perception, and efficient cloud-based aggregation iteration.

[0005] To address the shortcomings of the existing technologies, a precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion is needed to solve these problems. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion. It aims to overcome the shortcomings of existing power battery thermal runaway early warning technologies, such as coarse multimodal feature fusion, inability of decision-making mechanisms to adapt to changes in battery state throughout its entire life cycle, and low efficiency of cloud model iteration. It features multi-scale feature decomposition and adaptive deep fusion as the core of feature extraction, health state perception reinforcement learning as the dynamic decision-making mechanism, and momentum federated aggregation as the iterative optimization engine.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion, the system comprising: The multi-source data acquisition module is connected to the AI ​​feature fusion module via a high-speed data bus. The multi-source data acquisition module is used to collect multi-dimensional operating parameters of the battery and preprocess the collected data, and transmit the preprocessed multimodal data to the AI ​​feature fusion module. The AI ​​feature fusion module is connected to the intelligent decision-making and early warning module. The AI ​​feature fusion module is used to extract time-frequency features, mine time-series features, and perform multi-scale adaptive fusion on multimodal data to generate fused feature vectors, and output the fused feature vectors to the intelligent decision-making and early warning module. The time-frequency feature extraction adopts a fusion architecture of wavelet packet decomposition and improved RepLKNet network cascade. The time-series feature mining adopts a BiGRU-MH-GATT model of bidirectional gated recurrent units and multi-head global attention mechanism cascade. The multi-scale adaptive fusion adopts a feature fusion algorithm based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction. The intelligent decision-making and early warning module is connected to the cloud-based iterative optimization module. The intelligent decision-making and early warning module is used to dynamically adjust the decision weights of the main classifier and multiple auxiliary classifiers based on the fused feature vectors and the PER-DuelingDQN reinforcement learning framework that combines priority experience replay and duel network architecture. It generates thermal runaway early warning decision results and uploads decision logs and fault data to the cloud-based iterative optimization module. The cloud-based iterative optimization module is connected to the AI ​​feature fusion module and the intelligent decision-making and early warning module through OTA technology. The cloud-based iterative optimization module is used to aggregate the gradients of multi-terminal model parameters based on the momentum federated averaging algorithm, generate iterative model parameters and push them to the terminal, forming a closed-loop system of collection, fusion, decision-making and iteration.

[0008] Preferably, the multi-source data acquisition module adopts a hardware sampling and software parsing collaborative architecture. The hardware side is equipped with a high-speed ADC chip, a multi-dimensional sensor array, and an image acquisition unit. The multi-dimensional sensor array includes a current sensor, a voltage sensor, a temperature sensor, an impedance sensor, a lithium-ion migration rate sensor, and a SOC detection unit, simultaneously acquiring six core electrical parameters of the battery. The image acquisition unit uses an infrared thermal imaging camera to simultaneously acquire thermal imaging images of the battery pack surface. The software side performs data preprocessing through an edge computing unit. Data preprocessing includes: performing wavelet packet decomposition on the acquired battery electrical signals, extracting the wavelet packet energy of each frequency band as multi-channel input features of the improved RepLKNet network; processing the thermal imaging images using an adaptive platform histogram equalization algorithm, limiting the histogram peak value by setting a platform threshold, cropping parts exceeding the platform threshold, and uniformly distributing them across the entire grayscale range; the preprocessed electrical parameter feature data and thermal imaging feature data are transmitted to the AI ​​feature fusion module via TCP / IP protocol.

[0009] Preferably, the improved RepLKNet network adopts a reparameterized structure design. During the training phase, a multi-branch structure is used, including a large kernel convolution branch, a medium-small kernel convolution auxiliary branch, and an identity mapping branch. The large kernel convolution branch adopts a dynamic kernel adjustment mechanism, which determines the timing of kernel size switching based on the variance threshold of the real-time battery operating parameters. When the parameter variance is less than or equal to the preset threshold, it is determined to be in a stable state and a 31×31 large kernel convolution is used. When the parameter variance is greater than the preset threshold, it is determined to be in an abnormal state and automatically switches to a 15×15 medium kernel convolution. During the inference phase, the multi-branch structure is equivalently converted to a single-branch large kernel convolution structure through reparameterization. The improved RepLKNet network performs 8 layers of convolution operations, 4 layers of pooling operations, and 2 layers of fully connected layers on the multi-channel input features, and outputs a 512-dimensional time-frequency feature vector.

[0010] Preferably, in the BiGRU-MH-GATT model, the bidirectional gated recurrent unit (BiGRU) takes 6-dimensional time-series electrical parameter data as input and outputs a hidden state sequence H. The multi-head global attention mechanism (MH-GATT) performs multi-head parallel attention calculation on the hidden state sequence H, and the attention output of the m-th attention head is: ; in, These are the query projection matrix, key projection matrix, and value projection matrix for the m-th attention head, respectively. Let be the dimension of the key vector, and softmax be the normalization exponential function; The final output of multi-head attention is: ; Where Concat represents the vector concatenation operation, and M is the total number of attention heads. The output projection matrix is ​​obtained by assigning differentiated weights to each time node to enhance the information of key time nodes that are precursors to thermal runaway, and outputting a 256-dimensional time feature vector.

[0011] Preferably, in the feature fusion algorithm based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction, the 512-dimensional time-frequency feature vector is... With 256-dimensional temporal feature vectors Perform L-level discrete wavelet transform decomposition on each. The low-frequency approximation coefficients in the l-th layer are denoted as High-frequency detail coefficients are denoted as , The low-frequency approximation coefficients in the l-th layer are denoted as High-frequency detail coefficients are denoted as l=1,2,...,L; in the l-th layer, calculate the mutual information value of the low-frequency approximation coefficients respectively. Mutual information value of high-frequency detail coefficients : ; ; Where MI represents the mutual information function and Y is the fault label vector; Fusion weights of the low-frequency approximation coefficients in the l-th layer Fusion weights with high-frequency detail coefficients of layer l They are respectively: ; ; Low-frequency coefficients after layer l fusion and high frequency coefficient for: ; ; The fused coefficients of each layer are reconstructed using inverse discrete wavelet transform to generate a 768-dimensional fused feature vector. .

[0012] Preferably, the intelligent decision-making and early warning module includes a main classifier, an auxiliary classifier one, an auxiliary classifier two, a PER-DuelingDQN reinforcement learning decision unit, a counting and diagnosis unit, and a SOH estimation unit. The main classifier is based on an improved SVM algorithm and adopts an adaptive kernel function selection mechanism, dynamically switching between the RBF kernel function and the multinomial kernel function based on feature variance statistics. The auxiliary classifier one is based on a random forest algorithm and makes specific decisions for electrical parameter fault features. The auxiliary classifier two is based on a lightweight CNN and makes specific decisions for thermal imaging features. The SOH estimation unit performs probabilistic modeling of battery health status based on Gaussian process regression and outputs the SOH estimate and its variance. The PER-DuelingDQN reinforcement learning decision unit uses the fused feature vector, the output probabilities of the three classifiers, and the SOH estimate to form a state space, and uses the classifier weight adjustment strategy as the action space to dynamically optimize the decision weights of the three classifiers. The counting and diagnosis unit uses a sliding time window mechanism to verify the continuous matching degree between the fused feature vector and the fault feature template, and triggers the corresponding level of early warning when the preset continuous matching number condition is met.

[0013] Preferably, the PER-DuelingDQN reinforcement learning decision unit adopts a duel network architecture, and the Q-value function is decomposed into a state value function V and an advantage function A: ; Where s represents the state and a represents the action. To share convolutional layer parameters, For value stream branch parameters, For the dominant flow branch parameters, Size of the action space; It is the summation index variable in the action space; In the priority experience replay mechanism, the priority of the i-th experience is... Based on TD error Sure: ; ; in, The immediate reward for the i-th experience. As a discount factor, For the target network parameters, It is a preset positive constant; The reward function R is set as follows: ; Where P represents the accuracy of the early warning system, and F represents the false alarm rate. and These are the weighting coefficients; The SOH estimation unit models the battery capacity decay trajectory based on Gaussian process regression, for new input feature vectors SOH predicted mean and prediction variance for: ; ; Where K is the covariance kernel function, X is the training input matrix, and y is the training SOH label vector. Let I be the noise variance, and I be the identity matrix. This represents the new input feature vector to be predicted. This represents the row vector of covariance between the new sample point and all training points. This represents the prior covariance scalar of the new sample point itself; Warning threshold Dynamically adjusted based on SOH estimates: ; in, The initial warning threshold, To adjust the coefficient, Let be the estimated SOH value at time t.

[0014] Preferably, the cloud-based iterative optimization module includes a data storage unit, a model iteration unit, and an upgrade push unit; the data storage unit adopts a distributed database architecture to store raw monitoring data, fused feature vectors, decision logs, and early warning records; the model iteration unit aggregates the gradients of multi-terminal model parameters based on the momentum federated averaging algorithm, and the iterative formula for momentum federated averaging is: ; ; in, Let be the cumulative momentum gradient of the t-th iteration. Here, K is the momentum coefficient, and K is the total number of terminals participating in the aggregation. Let n be the local data volume of the k-th terminal, and n be the total data volume of all terminals. Let the gradient of the model parameters uploaded by the k-th terminal in the t-th iteration be denoted as . The global model parameters are for the t-th iteration; the upgrade push unit pushes the iterated model parameters to the AI ​​feature fusion module and the intelligent decision-making and early warning module through OTA technology.

[0015] Preferably, the intelligent decision-making and early warning module supports a three-level early warning mechanism: the first-level early warning is a suspected fault, which displays a prompt message through the vehicle terminal; the second-level early warning is a confirmed fault, which activates the audible and visual alarm device; the third-level early warning is an impending thermal runaway, which simultaneously triggers a cloud alarm, a user APP push, and a linkage with the vehicle's emergency system; the early warning response time does not exceed 1 second.

[0016] Preferably, the system hardware adopts a modular design, adapting to lithium iron phosphate batteries and ternary lithium batteries, supporting both pre-installation and retrofitting. The software uses an embedded Linux operating system, and is customized to adapt to the application scenarios of energy storage batteries and electric ship power batteries. The customized development includes interface adaptation of hardware modules and parameter fine-tuning of software algorithms. The interface adaptation of hardware modules includes adjusting the sensor withstand voltage level to 1000V for high-voltage scenarios of energy storage batteries, and improving the hardware waterproof level to IP68 for the humid environment of electric ships. The parameter fine-tuning of software algorithms includes adjusting the switching threshold of dynamic convolution kernels and the weight coefficients of the reward function of reinforcement learning based on the battery operating conditions of different scenarios.

[0017] The beneficial effects of this invention are as follows: 1. This invention employs a time-frequency feature extraction architecture cascaded with wavelet packet decomposition and an improved RepLKNet network, combined with a time-series feature mining model cascaded with bidirectional gated recurrent units and a multi-head global attention mechanism, and an adaptive feature fusion algorithm based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction. This achieves comprehensive three-dimensional capture and deep fusion of the multi-band energy distribution characteristics of battery electrical parameters in the time-frequency domain, the spatial thermal distribution characteristics of thermal imaging, and the temporal evolution of parameters. Compared to the shortcomings of existing technologies, such as single-dimensional feature extraction and coarse heterogeneous modal fusion, this invention can analyze the fault information correlation within multi-modal data at different scales, enabling the full expression of very early weak fault signals in the feature space, significantly improving the completeness and accuracy of fault feature extraction.

[0018] 2. This invention constructs a reinforcement learning decision-making framework that integrates priority experience replay and duel network architecture. It incorporates the battery health state estimated by Gaussian process regression and the fused feature vector as components of the reinforcement learning state space, enabling adaptive dynamic adjustment of the warning threshold as the battery health state deteriorates. Compared to existing technologies that use fixed thresholds or static classification models, this invention's decision-making mechanism can perceive the parameter baseline drift trend throughout the battery's entire lifecycle and dynamically optimize the collaborative decision weights of the main classifier and auxiliary classifier. Even in scenarios where parameter feature boundaries are blurred during battery aging, it maintains stable fault identification performance, effectively addressing the increased risk of missed detections in the later stages of battery aging in existing technologies.

[0019] 3. This invention employs a momentum federated averaging algorithm in the cloud-based iterative optimization module for multi-terminal model parameter gradient aggregation. Compared to the slow convergence speed of simple federated averaging algorithms in existing technologies, the introduction of the momentum mechanism can accumulate historical gradient direction information, accelerating the convergence of the global model and improving iterative stability. Simultaneously, the modular hardware architecture and finely adjustable software algorithm enable the system to be adapted to both lithium iron phosphate and ternary lithium batteries. Furthermore, through hardware interface adaptation and algorithm parameter adjustment, it can be extended to different application scenarios such as energy storage batteries and electric ships, thus addressing the shortcomings of insufficient versatility in existing technologies. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the internal structure and data flow of the AI ​​feature fusion module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the improved RepLKNet network structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of a reinforcement learning-driven multi-classifier collaborative decision-making process in an embodiment of the present invention. Figure 5 This is a schematic diagram of the system early warning response timing in an embodiment of the present invention. Detailed Implementation

[0021] Example 1: like Figure 1 As shown, a precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion is described. The system includes: The multi-source data acquisition module is connected to the AI ​​feature fusion module via a high-speed data bus. The multi-source data acquisition module is used to collect multi-dimensional operating parameters of the battery and preprocess the collected data, and transmit the preprocessed multimodal data to the AI ​​feature fusion module. The AI ​​feature fusion module is connected to the intelligent decision-making and early warning module. The AI ​​feature fusion module is used to extract time-frequency features, mine time-series features, and perform multi-scale adaptive fusion on multimodal data to generate fused feature vectors, and output the fused feature vectors to the intelligent decision-making and early warning module. The time-frequency feature extraction adopts a fusion architecture of wavelet packet decomposition and improved RepLKNet network cascade. The time-series feature mining adopts a BiGRU-MH-GATT model of bidirectional gated recurrent units and multi-head global attention mechanism cascade. The multi-scale adaptive fusion adopts a feature fusion algorithm based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction. The intelligent decision-making and early warning module is connected to the cloud-based iterative optimization module. The intelligent decision-making and early warning module is used to dynamically adjust the decision weights of the main classifier and multiple auxiliary classifiers based on the fused feature vectors and the PER-DuelingDQN reinforcement learning framework that combines priority experience replay and duel network architecture. It generates thermal runaway early warning decision results and uploads decision logs and fault data to the cloud-based iterative optimization module. The cloud-based iterative optimization module is connected to the AI ​​feature fusion module and the intelligent decision-making and early warning module through OTA technology. The cloud-based iterative optimization module is used to aggregate the gradients of multi-terminal model parameters based on the momentum federated averaging algorithm, generate iterative model parameters and push them to the terminal, forming a closed-loop system of collection, fusion, decision-making and iteration.

[0022] like Figure 2As shown, preferably, the multi-source data acquisition module adopts a hardware sampling and software parsing collaborative architecture. The hardware side is equipped with a high-speed ADC chip, a multi-dimensional sensor array, and an image acquisition unit. The multi-dimensional sensor array includes a current sensor, a voltage sensor, a temperature sensor, an impedance sensor, a lithium-ion migration rate sensor, and a SOC detection unit, simultaneously acquiring six core electrical parameters of the battery. The image acquisition unit uses an infrared thermal imaging camera to simultaneously acquire thermal imaging images of the battery pack surface. The software side performs data preprocessing through an edge computing unit. Data preprocessing includes: performing wavelet packet decomposition on the acquired battery electrical signals, extracting the wavelet packet energy of each frequency band as multi-channel input features of the improved RepLKNet network; processing the thermal imaging images using an adaptive platform histogram equalization algorithm, limiting the histogram peak value by setting a platform threshold, cropping parts exceeding the platform threshold, and uniformly distributing them across the entire grayscale range; the preprocessed electrical parameter feature data and thermal imaging feature data are transmitted to the AI ​​feature fusion module via TCP / IP protocol.

[0023] like Figure 3 As shown, preferably, the improved RepLKNet network adopts a structural reparameterization design. During the training phase, a multi-branch structure is used, including a large kernel convolution branch, a medium-small kernel convolution auxiliary branch, and an identity mapping branch. The large kernel convolution branch adopts a dynamic convolution kernel adjustment mechanism, which determines the timing of kernel size switching based on the variance threshold of the battery's real-time operating parameters. When the parameter variance is less than or equal to the preset threshold, it is determined to be in a stable state and a 31×31 large kernel convolution is used. When the parameter variance is greater than the preset threshold, it is determined to be in an abnormal state and automatically switches to a 15×15 medium kernel convolution. During the inference phase, the multi-branch structure is equivalently converted to a single-branch large kernel convolution structure through structural reparameterization. After processing the multi-channel input features with 8 layers of convolution operations, 4 layers of pooling operations, and 2 layers of fully connected layers, the improved RepLKNet network outputs a 512-dimensional time-frequency feature vector.

[0024] Preferably, in the BiGRU-MH-GATT model, the bidirectional gated recurrent unit (BiGRU) takes 6-dimensional time-series electrical parameter data as input and outputs a hidden state sequence H. The multi-head global attention mechanism (MH-GATT) performs multi-head parallel attention calculation on the hidden state sequence H, and the attention output of the m-th attention head is: ; in, These are the query projection matrix, key projection matrix, and value projection matrix for the m-th attention head, respectively. Let be the dimension of the key vector, and softmax be the normalization exponential function; The final output of multi-head attention is: ; Where Concat represents the vector concatenation operation, and M is the total number of attention heads. The output projection matrix is ​​obtained by assigning differentiated weights to each time node to enhance the information of key time nodes that are precursors to thermal runaway, and outputting a 256-dimensional time feature vector.

[0025] Preferably, in the feature fusion algorithm based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction, the 512-dimensional time-frequency feature vector is... With 256-dimensional temporal feature vectors Perform L-level discrete wavelet transform decomposition on each. The low-frequency approximation coefficients in the l-th layer are denoted as High-frequency detail coefficients are denoted as , The low-frequency approximation coefficients in the l-th layer are denoted as High-frequency detail coefficients are denoted as l=1,2,...,L; in the l-th layer, calculate the mutual information value of the low-frequency approximation coefficients respectively. Mutual information value of high-frequency detail coefficients : ; ; Where MI represents the mutual information function and Y is the fault label vector; Fusion weights of the low-frequency approximation coefficients in the l-th layer Fusion weights with high-frequency detail coefficients of layer l They are respectively: ; ; Low-frequency coefficients after layer l fusion and high frequency coefficient for: ; ; The fused coefficients of each layer are reconstructed using inverse discrete wavelet transform to generate a 768-dimensional fused feature vector. .

[0026] like Figure 4As shown, preferably, the intelligent decision-making and early warning module includes a main classifier, an auxiliary classifier one, an auxiliary classifier two, a PER-DuelingDQN reinforcement learning decision unit, a counting and diagnosis unit, and a SOH estimation unit. The main classifier is based on an improved SVM algorithm and adopts an adaptive kernel function selection mechanism, dynamically switching between the RBF kernel function and the multinomial kernel function based on feature variance statistics. The auxiliary classifier one is based on a random forest algorithm and makes specific decisions for electrical parameter fault features. The auxiliary classifier two is based on a lightweight CNN and makes specific decisions for thermal imaging features. The SOH estimation unit performs probabilistic modeling of battery health status based on Gaussian process regression and outputs the SOH estimate and its variance. The PER-DuelingDQN reinforcement learning decision unit uses the fused feature vector, the output probabilities of the three classifiers, and the SOH estimate to form a state space, and uses the classifier weight adjustment strategy as the action space to dynamically optimize the decision weights of the three classifiers. The counting and diagnosis unit uses a sliding time window mechanism to verify the continuous matching degree between the fused feature vector and the fault feature template, and triggers the corresponding level of early warning when the preset continuous matching number condition is met.

[0027] Preferably, the PER-DuelingDQN reinforcement learning decision unit adopts a duel network architecture, and the Q-value function is decomposed into a state value function V and an advantage function A: ; Where s represents the state and a represents the action. To share convolutional layer parameters, For value stream branch parameters, For the dominant flow branch parameters, Size of the action space; It is the summation index variable in the action space; In the priority experience replay mechanism, the priority of the i-th experience is... Based on TD error Sure: ; ; in, The immediate reward for the i-th experience. As a discount factor, For the target network parameters, It is a preset positive constant; The reward function R is set as follows: ; Where P represents the accuracy of the early warning system, and F represents the false alarm rate. and These are the weighting coefficients; The SOH estimation unit models the battery capacity decay trajectory based on Gaussian process regression, for new input feature vectors SOH predicted mean and prediction variance for: ; ; Where K is the covariance kernel function, X is the training input matrix, and y is the training SOH label vector. Let I be the noise variance, and I be the identity matrix. This represents the new input feature vector to be predicted. This represents the row vector of covariance between the new sample point and all training points. This represents the prior covariance scalar of the new sample point itself; Warning threshold Dynamically adjusted based on SOH estimates: ; in, The initial warning threshold, To adjust the coefficient, Let be the estimated SOH value at time t.

[0028] Preferably, the cloud-based iterative optimization module includes a data storage unit, a model iteration unit, and an upgrade push unit; the data storage unit adopts a distributed database architecture to store raw monitoring data, fused feature vectors, decision logs, and early warning records; the model iteration unit aggregates the gradients of multi-terminal model parameters based on the momentum federated averaging algorithm, and the iterative formula for momentum federated averaging is: ; ; in, Let be the cumulative momentum gradient of the t-th iteration. Here, K is the momentum coefficient, and K is the total number of terminals participating in the aggregation. Let n be the local data volume of the k-th terminal, and n be the total data volume of all terminals. Let the gradient of the model parameters uploaded by the k-th terminal in the t-th iteration be denoted as . The global model parameters are for the t-th iteration; the upgrade push unit pushes the iterated model parameters to the AI ​​feature fusion module and the intelligent decision-making and early warning module through OTA technology.

[0029] like Figure 5As shown, preferably, the intelligent decision-making and early warning module supports a three-level early warning mechanism: the first-level early warning is a suspected fault, which displays a prompt message through the vehicle terminal; the second-level early warning is a confirmed fault, which activates the audible and visual alarm device; the third-level early warning is an imminent thermal runaway, which simultaneously triggers a cloud alarm, a user APP push, and a linkage with the vehicle emergency system; the early warning response time does not exceed 1 second.

[0030] Preferably, the system hardware adopts a modular design, adapting to lithium iron phosphate batteries and ternary lithium batteries, supporting both pre-installation and retrofitting. The software uses an embedded Linux operating system, and is customized to adapt to the application scenarios of energy storage batteries and electric ship power batteries. The customized development includes interface adaptation of hardware modules and parameter fine-tuning of software algorithms. The interface adaptation of hardware modules includes adjusting the sensor withstand voltage level to 1000V for high-voltage scenarios of energy storage batteries, and improving the hardware waterproof level to IP68 for the humid environment of electric ships. The parameter fine-tuning of software algorithms includes adjusting the switching threshold of dynamic convolution kernels and the weight coefficients of the reward function of reinforcement learning based on the battery operating conditions of different scenarios.

[0031] Example 2: This embodiment provides a specific scheme for system construction, as follows: 1. System hardware setup: This invention comprises a multi-source data acquisition module, an AI feature fusion module, an intelligent decision-making and early warning module, and a cloud-based iterative optimization module, forming a closed-loop architecture of acquisition, fusion, decision-making, and iteration. The multi-source data acquisition module synchronously acquires battery electrical parameters and thermal distribution data through a sensor array and an infrared thermal imaging camera, and preprocesses them using wavelet packet decomposition and adaptive platform histogram equalization. The AI ​​feature fusion module extracts multi-channel time-frequency features using wavelet packet decomposition and a cascaded improved RepLKNet network, extracts temporal features using bidirectional gated recurrent units and a multi-head global attention mechanism, and achieves adaptive cross-modal fusion through multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction. The intelligent decision-making and early warning module dynamically adjusts the decision weights of the multi-classifier using a reinforcement learning framework that combines priority experience replay and duel network fusion, and adaptively adjusts the early warning threshold based on battery health status perceived by Gaussian process regression. The cloud-based iterative optimization module uses a momentum federated averaging algorithm for efficient model iteration and OTA push. This invention achieves comprehensive extraction and deep fusion of multi-dimensional fault features, adaptive intelligent decision-making throughout the entire lifecycle, and efficient cloud-based collaborative iteration.

[0032] 2. System software and algorithm implementation: The data preprocessing software of the multi-source data acquisition module is written in C language and runs on the FPGA edge computing unit. The preprocessing process is as follows: First, the current, voltage, temperature, impedance, lithium-ion migration rate, and SOC six-dimensional electrical signals are decomposed into three layers of wavelet packets. The db4 wavelet basis function is selected to extract the wavelet packet energy of each terminal node and generate an eight-channel wavelet packet energy feature matrix, which serves as the multi-channel input feature of the improved RepLKNet network. Second, the infrared thermal imaging image is processed using an adaptive platform histogram equalization algorithm. The platform threshold is set to 1.5 times the mean of the histogram. The number of pixels exceeding the platform threshold is cropped, and the cropped pixels are uniformly redistributed to the entire grayscale range. This enhances the local contrast of the battery thermal distribution while avoiding over-enhancement of hot spots. Third, the above wavelet packet energy features and equalized thermal imaging image data are processed by Kalman filtering to remove high-frequency noise. Linear interpolation is used to fill in the occasional missing data. Min-Max normalization is used to standardize all parameters to the [0,1] interval to generate a standardized multimodal dataset, which is then transmitted to the AI ​​feature fusion module via TCP / IP protocol.

[0033] The time-frequency feature extraction algorithm of the AI ​​feature fusion module is implemented based on the PyTorch framework. The improved RepLKNet network adopts a multi-branch structure during training, including a large-kernel convolution branch, a medium-kernel convolution auxiliary branch, and an identity mapping branch. The large-kernel convolution branch has a built-in dynamic kernel adjustment mechanism, which determines the timing of kernel size switching based on the variance threshold of the real-time battery operating parameters. The preset variance threshold is 0.01. When the parameter variance is less than or equal to this threshold, it is considered a stable state, and a 31×31 large-kernel convolution is used to expand the receptive field. When the parameter variance is greater than this threshold, it is considered an abnormal state, and it automatically switches to a 15×15 medium-kernel convolution to improve computational efficiency. The medium-kernel convolution auxiliary branch uses a fixed 5×5 convolution kernel. After training, the kernel parameters of the three branches are equivalently fused into a single large-kernel convolution kernel through structural reparameterization. Only this single-branch structure is retained during the inference phase, maintaining the high fitting ability of the multi-branch structure during training while obtaining the low latency advantage of the single-branch structure during inference. The network input is a 224×224 multi-channel feature map, where eight channels correspond to the eight frequency band energies of wavelet packet decomposition. After processing through 8 convolutional layers, 4 pooling layers, and 2 fully connected layers, the output is a 512-dimensional time-frequency feature vector. The RepLKNet network is trained for 200 iterations with a learning rate of 0.001 and uses the Adam optimizer.

[0034] The temporal feature mining algorithm of the AI ​​feature fusion module adopts the BiGRU-MH-GATT model. The input to the bidirectional gated recurrent unit (BiGRU) is 6-dimensional temporal electrical parameter data, with a sampling frequency of 1kHz. Both the forward and backward GRU hidden layers have a dimension of 128, which are concatenated to form a 256-dimensional hidden state sequence H. A multi-head global attention mechanism performs parallel attention computation on the hidden state sequence H. The query projection matrix, key projection matrix, and value projection matrix of the m-th attention head all have a dimension of 256×64, the key vector has a dimension of 64, the total number of attention heads M is 8, and the output projection matrix has a dimension of 512×256. After the outputs of each attention head are spliced ​​and linearly projected, the attention weights are dynamically adjusted by calculating the cosine similarity between each time-series node and the thermal runaway fault features. This strengthens the information weights of key time-series nodes that are precursors to thermal runaway, such as the 10 seconds before a voltage surge and the slow temperature rise, and outputs a 256-dimensional time-series feature vector. The BiGRU-MH-GATT model is trained using the cross-entropy loss function, with a batch size of 64, a learning rate of 0.0005, and 150 training iterations.

[0035] The multi-scale adaptive fusion algorithm of the AI ​​feature fusion module adopts a feature fusion strategy based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction. The fusion algorithm is written in Python and runs on the Jetson Xavier NX platform, with a feature fusion latency of no more than 50ms. The specific fusion process is as follows: First, the 512-dimensional time-frequency feature vector and the 256-dimensional time-series feature vector are aligned in dimension, and the time-series feature vector is linearly mapped to expand to 512 dimensions; Second, the F... t and F s The process involves three levels of discrete wavelet transform decomposition, with db2 as the wavelet basis. Each level yields a set of low-frequency approximation coefficients and a set of high-frequency detail coefficients. The third step involves calculating the mutual information values ​​of the time-frequency and time-series feature coefficients relative to the fault label vector Y at each decomposition level, using these mutual information values ​​as the basis for allocating the fusion weights for that level. The fourth step involves weighting and fusing the corresponding coefficients of the time-frequency and time-series features at each decomposition level according to the proportion of mutual information values, fusing the low-frequency approximation coefficients and high-frequency detail coefficients separately. The fifth step involves reconstructing the fused coefficients using inverse discrete wavelet transform to generate a 768-dimensional fusion feature vector F. f This algorithm enables the fusion ratio of time-frequency features and time-series features to be adaptively adjusted at different decomposition scales. When a certain mode contains more fault discrimination information at a specific frequency level, its fusion weight is automatically increased.

[0036] In the multi-classifier collaborative architecture of the intelligent decision-making and early warning module, the main classifier adopts an improved SVM algorithm. The kernel function adaptive selection mechanism determines the switching threshold based on the feature variance statistics of 670,000 sets of measured data, with a feature variance threshold of 0.3. When the feature variance is less than or equal to 0.3, the RBF kernel function is used; when the feature variance is greater than 0.3, it automatically switches to the multinomial kernel function. The first auxiliary classifier is based on the random forest algorithm and contains 100 decision trees, specifically for making decisions on fault features related to electrical parameters. The second auxiliary classifier is based on a lightweight CNN and contains 3 convolutional layers and 2 fully connected layers, specifically for making decisions on faults related to thermal imaging features.

[0037] The PER-Dueling DQN reinforcement learning decision-making unit of the intelligent decision-making and early warning module adopts a duel network architecture, where the Q-value function is decomposed into two parts: the state value function V and the advantage function A. Shared convolutional layers are used. Feature encoding of the input state, value stream branch parameters Output state value scalar V, dominant flow branch parameters The output is the advantage value A corresponding to each action. Mean removal is applied to the advantage values ​​to ensure the distinguishability of V and A. The state space consists of a 768-dimensional fused feature vector, the fault probability values ​​output by the main classifier, the fault probability values ​​output by auxiliary classifier one, the fault probability values ​​output by auxiliary classifier two, and the SOH estimate. The action space is the weight adjustment strategy for the three classifiers. The weight of the main classifier can be adjusted from 0.4 to 0.6, and the weights of the two auxiliary classifiers can each be adjusted from 0.2 to 0.3. The sum of the three weights is always 1. In the priority experience replay mechanism, the experience pool capacity is 10,000 experiences, with priority... Based on TD error The absolute value calculation is preset with a positive constant. The initial weight is 0.01, and the exponential decay factor of the importance sampling weight gradually increases to 1 during training. In the reward function R, the weight coefficient of the warning accuracy P is... The weighting coefficient of the false alarm rate F is 0.7. The value is 0.3. The reinforcement learning training iterations are 500 rounds, and the discount factor is... The learning rate is 0.99, the target network performs a hard parameter update every 100 steps.

[0038] The SOH estimation unit of the intelligent decision-making and early warning module uses Gaussian process regression to probabilistically model the battery capacity decay trajectory. The covariance kernel function for the Gaussian process regression is a combination of the radial basis function kernel and the linear kernel, with noise variance... Initialized to 0.01. The training input matrix X consists of features such as the number of battery charge-discharge cycles, internal resistance increment, coulombic efficiency decay rate, and voltage plateau slope change. The training label vector y is the actual SOH value at the corresponding time. For new feature vectors acquired online, the SOH estimation unit simultaneously outputs the predicted mean and predicted variance, with the predicted variance serving as a quantitative indicator of estimation uncertainty. The warning threshold is dynamically adjusted based on the SOH estimate, with an initial warning threshold of 0.85 and an adjustment coefficient of 0.3. When the SOH estimate decreases, the warning threshold is automatically tightened to compensate for the impact of parameter baseline drift on the fault feature boundary.

[0039] The counting and diagnostic mechanism of the intelligent decision-making and early warning module is written in C++ and runs on an ARM processor, with a response latency of no more than 10ms. The mechanism sets a 10-second sliding time window, within which the matching degree between the fused feature vector and the fault feature template is calculated at a sampling frequency of 1Hz, using cosine similarity. When the matching degree is not less than 85% for 7 consecutive times within the sliding time window, the fault risk is confirmed, and the corresponding level of early warning is triggered.

[0040] The specific execution logic of the three-level early warning mechanism of the intelligent decision-making and early warning module is as follows: Level 1 early warning corresponds to a matching degree range of 85% to 90%, which is determined to be a suspected fault. The vehicle terminal displays a yellow warning message and records fault characteristic data simultaneously. Level 2 early warning corresponds to a matching degree range of 90% to 95%, which is determined to be a confirmed fault. A buzzer with a volume of no less than 85dB and a red LED warning light with a flashing frequency of 2Hz are activated, and the vehicle terminal displays the fault type and risk level. Level 3 early warning corresponds to a matching degree of no less than 95%, which is determined to be an imminent thermal runaway. In addition to local alarms, emergency early warning information is pushed to the cloud platform and the user's mobile APP through the 4G or 5G communication module, and the vehicle's emergency system is activated to cut off the battery power.

[0041] The model iteration algorithm of the cloud-based iterative optimization module is implemented based on the momentum federated averaging framework. Terminal devices calculate the gradients of AI feature fusion model parameters and intelligent decision-making model parameters based on local data, encrypt the gradient information using a homomorphic encryption algorithm, and then upload it to the cloud. After receiving the encrypted gradients uploaded by K terminals in the t-th iteration, the cloud first decrypts each gradient, then aggregates and calculates them according to the momentum federated averaging iterative formula, with the momentum coefficient set to 0.9. The momentum cumulative gradient incorporates the direction information of historical gradients in each iteration, and the global model parameters are updated under the guidance of the momentum cumulative gradient. After the global model parameters iteration is complete, they are homomorphically encrypted and sent back to each terminal, which then decrypts and updates its local model parameters. The incremental learning trigger condition is that every 100,000 sets of new data are accumulated, and when the model accuracy improves by at least 1% after iteration, the upgrade push unit generates an upgrade package no larger than 100MB and pushes it to the terminal device via OTA technology. Users can choose automatic or manual upgrade modes.

[0042] 3. System performance testing: A test platform was built for lithium iron phosphate batteries with a rated voltage of 3.2V and a rated capacity of 100Ah, and a test platform for ternary lithium batteries with a rated voltage of 3.7V and a rated capacity of 100Ah. Four typical fault scenarios were simulated: a battery aging scenario involving 1000 charge-discharge cycles; an external mechanical compression scenario involving 80 repeated hammer blows; an internal short circuit scenario simulating a micro-short circuit with an external 1-ohm resistor; and a high-temperature operation scenario in a 60℃ constant temperature chamber. Six-dimensional electrical parameter data and infrared thermal imaging images were simultaneously acquired during the test, and data acquisition continued until a level-three warning was triggered or the battery safety protection system activated.

[0043] Data acquisition accuracy test results: Under all four fault scenarios, the absolute value of the current measurement error does not exceed 1% of the rated current, the absolute value of the voltage measurement error does not exceed 0.5%, the absolute value of the temperature measurement error does not exceed 0.5%, and the absolute value of the thermal imaging temperature measurement error does not exceed 1℃, which meets the system design specifications.

[0044] Feature extraction performance test results: The improved RepLKNet network achieved a time-frequency feature extraction accuracy of 98.5% under dynamic switching of kernel size and convolution; the BiGRU-MH-GATT model achieved a time-series feature extraction accuracy of 98.9%; after multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction fusion, the matching degree between the fused feature vector and the fault label was no less than 98.7%, verifying the effective integration capability of the multimodal feature fusion algorithm for heterogeneous features.

[0045] Performance test results for decision-making and early warning: In the lithium iron phosphate battery scenario, the early warning accuracy is 95.2%, with a false positive and false negative rate of 4.8%; in the ternary lithium battery scenario, the early warning accuracy is 94.8%, with a false positive and false negative rate of 5.0%. The PER-Dueling DQN reinforcement learning decision unit can adaptively adjust the classifier weights in the battery aging scenario. When the SOH estimate decreases, it automatically increases the weights dependent on electrical parameter features, effectively suppressing false negatives caused by parameter baseline drift during the aging stage. The early warning response time is no more than 0.8 seconds, meeting real-time requirements.

[0046] Warning duration test results: The system's average warning duration is 22.3 hours, which is about 95.8% earlier than that of traditional battery management systems. In the internal short circuit simulation scenario, the system completes feature extraction and fusion within 0.6 seconds after the fault characteristics appear, and issues a level-two warning within 3 minutes, leaving sufficient time for emergency response for the occupants.

[0047] System power consumption test results: The overall operating current of the system does not exceed 60mA, and the impact on the battery range does not exceed 3%, meeting the low power consumption requirements of vehicle equipment.

[0048] 4. System application verification: The system was installed on a new energy vehicle test vehicle equipped with a ternary lithium battery pack with a total voltage of 350V and a total capacity of 80Ah, and a 12-month real-vehicle test was conducted.

[0049] Road test scenarios: The test vehicle accumulated 20,000 km of driving, covering three typical operating conditions: urban roads, highways, and mountain roads. During the test, no hardware failures or software crashes occurred, data communication links between modules remained stable, and the cloud-based iterative optimization module completed model gradient aggregation and parameter push according to the preset cycle. In the scenario of frequent start-stop cycles causing current fluctuations in urban road conditions, the system did not trigger false alarms; in the scenario of continuous high-current discharge on highways, the system accurately tracked the slow upward trend of battery temperature and did not generate false alarms due to fluctuations in a single parameter.

[0050] Static Storage Scenario Test: Three static storage scenarios were simulated: static storage in a garage, high-temperature exposure at 65℃, and low-temperature storage at -20℃. In the high-temperature exposure scenario, the system successfully identified five minor fault signs, including voltage drift and abnormal local temperature increases. Level 1 warnings were issued more than 24 hours in advance of the occurrence of these fault characteristics, prompting users to check the battery status promptly. In the low-temperature storage scenario, the system failed to misinterpret the normal increase in internal resistance caused by the low temperature as a fault.

[0051] Fault simulation test: Two highly concealed faults, loose battery tabs and electrolyte leakage, were manually simulated. Loose tabs caused intermittent fluctuations in contact impedance. The SOH estimation unit continuously tracked the internal resistance increment trend through Gaussian process regression. When the SOH estimate showed a significant decay trend, the warning threshold was automatically tightened. The system issued a level-two warning after the fault characteristics accumulated to a identifiable intensity. Electrolyte leakage caused abnormal local thermal distribution. Adaptive platform histogram equalization preprocessing effectively enhanced the thermal distribution details in low-contrast areas. The improved RepLKNet network accurately extracted local hotspot features from thermal imaging, which were then fused with electrical parameter time-series features at multiple scales to trigger a precise warning. No false alarms or missed alarms were observed in either fault scenario. The system performance fully met the requirements for alarm signal output stability and response time in the national standard GB38031-2025 "Safety Requirements for Power Batteries for Electric Vehicles".

Claims

1. A precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion, characterized in that the system... include: The multi-source data acquisition module is connected to the AI ​​feature fusion module via a high-speed data bus. The multi-source data acquisition module is used to collect multi-dimensional operating parameters of the battery and preprocess the collected data, and transmit the preprocessed multimodal data to the AI ​​feature fusion module. The AI ​​feature fusion module is connected to the intelligent decision-making and early warning module. The AI ​​feature fusion module is used to extract time-frequency features, mine time-series features and perform multi-scale adaptive fusion on multimodal data to generate a fused feature vector, and output the fused feature vector to the intelligent decision-making and early warning module. Time-frequency feature extraction adopts a fusion architecture of wavelet packet decomposition and improved RepLKNet network concatenation; time-series feature mining adopts BiGRU-MH-GATT model of bidirectional gated recurrent unit and multi-head global attention mechanism concatenation; multi-scale adaptive fusion adopts feature fusion algorithm based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction. The intelligent decision-making and early warning module is connected to the cloud-based iterative optimization module. The intelligent decision-making and early warning module is used to dynamically adjust the decision weights of the main classifier and multiple auxiliary classifiers based on the fused feature vectors and the PER-DuelingDQN reinforcement learning framework that combines priority experience replay and duel network architecture. It generates thermal runaway early warning decision results and uploads decision logs and fault data to the cloud-based iterative optimization module. The cloud-based iterative optimization module is connected to the AI ​​feature fusion module and the intelligent decision-making and early warning module through OTA technology. The cloud-based iterative optimization module is used to aggregate the gradients of multi-terminal model parameters based on the momentum federated averaging algorithm, generate iterative model parameters and push them to the terminal, forming a closed-loop system of collection, fusion, decision-making and iteration.

2. The precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion as described in claim 1, characterized in that, The multi-source data acquisition module adopts a hardware sampling and software parsing collaborative architecture. The hardware is equipped with a high-speed ADC chip, a multi-dimensional sensor array and an image acquisition unit. The multi-dimensional sensor array includes a current sensor, a voltage sensor, a temperature sensor, an impedance sensor, a lithium-ion migration rate sensor and a SOC detection unit, which simultaneously acquires six core electrical parameters of the battery. The image acquisition unit uses an infrared thermal imaging camera to simultaneously acquire thermal imaging images of the battery pack surface. The software implements data preprocessing through edge computing units. The data preprocessing includes: performing wavelet packet decomposition on the collected battery electrical signals and extracting the wavelet packet energy of each frequency band as multi-channel input features of the improved RepLKNet network. The thermal imaging images are processed using an adaptive platform histogram equalization algorithm. By setting a platform threshold to limit the peak value of the histogram, the portion exceeding the platform threshold is cropped and evenly distributed across the entire grayscale range. The preprocessed electrical parameter feature data and thermal imaging feature data are transmitted to the AI ​​feature fusion module via TCP / IP protocol.

3. The precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion as described in claim 1, characterized in that, The improved RepLKNet network employs a reparameterized architecture design. During training, it uses a multi-branch structure, including a large kernel convolution branch, a medium-small kernel convolution auxiliary branch, and an identity mapping branch. The large kernel convolution branch uses a dynamic kernel adjustment mechanism, determining the timing of kernel size switching based on the variance threshold of the battery's real-time operating parameters. When the parameter variance is less than or equal to a preset threshold, it is considered a stable state, and a 31×31 large kernel convolution is used. When the parameter variance is greater than the preset threshold, it is considered an abnormal state, and it automatically switches to a 15×15 medium kernel convolution. During the inference phase, the multi-branch structure is equivalently converted to a single-branch large kernel convolution structure through reparameterization. The improved RepLKNet network processes the multi-channel input features through 8 layers of convolution operations, 4 layers of pooling operations, and 2 layers of fully connected layers, outputting a 512-dimensional time-frequency feature vector.

4. The precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion as described in claim 1, characterized in that, In the BiGRU-MH-GATT model, the bidirectional gated recurrent unit (BiGRU) takes 6-dimensional time-series electrical parameter data as input and outputs a hidden state sequence H. The multi-head global attention mechanism (MH-GATT) performs multi-head parallel attention calculation on the hidden state sequence H. The attention output of the m-th attention head is: ； in, These are the query projection matrix, key projection matrix, and value projection matrix for the m-th attention head, respectively. Let be the dimension of the key vector, and softmax be the normalization exponential function; The final output of multi-head attention is: ； Where Concat represents the vector concatenation operation, and M is the total number of attention heads. The output projection matrix is ​​obtained by assigning differentiated weights to each time node to enhance the information of key time nodes that are precursors to thermal runaway, and outputting a 256-dimensional time feature vector.

5. The precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion according to claim 1, characterized in that, In the feature fusion algorithm based on multi-scale discrete wavelet transform decomposition and mutual information weighted reconstruction, the 512-dimensional time-frequency feature vector With 256-dimensional temporal feature vectors Perform L-level discrete wavelet transform decomposition on each. The low-frequency approximation coefficients in the l-th layer are denoted as High-frequency detail coefficients are denoted as , The low-frequency approximation coefficients in the l-th layer are denoted as High-frequency detail coefficients are denoted as l=1,2,...,L; in the l-th layer, calculate the mutual information value of the low-frequency approximation coefficients respectively. Mutual information value of high-frequency detail coefficients : ； ； Where MI represents the mutual information function and Y is the fault label vector; Fusion weights of the low-frequency approximation coefficients in the l-th layer Fusion weights with high-frequency detail coefficients of layer l They are respectively: ； ； Low-frequency coefficients after layer l fusion and high frequency coefficient for: ； ； The fused coefficients of each layer are reconstructed using inverse discrete wavelet transform to generate a 768-dimensional fused feature vector. .

6. The precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion according to claim 1, characterized in that, The intelligent decision-making and early warning module includes a main classifier, auxiliary classifier one, auxiliary classifier two, a PER-DuelingDQN reinforcement learning decision unit, a counting and diagnosis unit, and a SOH estimation unit. The main classifier is based on an improved SVM algorithm and adopts an adaptive kernel function selection mechanism, dynamically switching between the RBF kernel function and the multinomial kernel function based on feature variance statistics. Auxiliary classifier one is based on a random forest algorithm and makes specific decisions for electrical parameter fault features. Auxiliary classifier two is based on a lightweight CNN and makes specific decisions for thermal imaging features. The SOH estimation unit performs probabilistic modeling of battery health status based on Gaussian process regression and outputs the SOH estimate and its variance. The PER-DuelingDQN reinforcement learning decision unit uses the fused feature vector, the output probabilities of the three classifiers, and the SOH estimate to form the state space, and uses the classifier weight adjustment strategy as the action space to dynamically optimize the decision weights of the three classifiers. The counting and diagnosis unit uses a sliding time window mechanism to verify the continuous matching degree between the fused feature vector and the fault feature template, and triggers the corresponding level of early warning when the preset continuous matching number condition is met.

7. A precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion as described in claim 6, characterized in that, The PER-DuelingDQN reinforcement learning decision unit adopts a duel network architecture, and the Q-value function is decomposed into a state value function V and an advantage function A: ； Where s represents the state and a represents the action. To share convolutional layer parameters, For value stream branch parameters, For the dominant flow branch parameters, Size of the action space; It is the summation index variable in the action space; In the priority experience replay mechanism, the priority of the i-th experience is... Based on TD error Sure: ； ； in, The immediate reward for the i-th experience. As a discount factor, For the target network parameters, It is a preset positive constant; The reward function R is set as follows: ； Where P represents the accuracy of the early warning system, and F represents the false alarm rate. and These are the weighting coefficients; The SOH estimation unit models the battery capacity decay trajectory based on Gaussian process regression, for new input feature vectors SOH predicted mean and prediction variance for: ； ； Where K is the covariance kernel function, X is the training input matrix, and y is the training SOH label vector. Let I be the noise variance, and I be the identity matrix. This represents the new input feature vector to be predicted. This represents the row vector of covariance between the new sample point and all training points. This represents the prior covariance scalar of the new sample point itself; Warning threshold Dynamically adjusted based on SOH estimates: ； in, The initial warning threshold, To adjust the coefficient, Let be the estimated SOH value at time t.

8. The precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion according to claim 1, characterized in that, The cloud-based iterative optimization module includes a data storage unit, a model iteration unit, and an upgrade push unit. The data storage unit adopts a distributed database architecture to store raw monitoring data, fused feature vectors, decision logs, and early warning records. The model iteration unit aggregates the gradients of multi-terminal model parameters based on a momentum federated averaging algorithm. The iterative formula for momentum federated averaging is: ； ； in, Let be the cumulative momentum gradient of the t-th iteration. Here, K is the momentum coefficient, and K is the total number of terminals participating in the aggregation. Let n be the local data volume of the k-th terminal, and n be the total data volume of all terminals. Let the gradient of the model parameters uploaded by the k-th terminal in the t-th iteration be denoted as . The global model parameters are for the t-th iteration; the upgrade push unit pushes the iterated model parameters to the AI ​​feature fusion module and the intelligent decision-making and early warning module through OTA technology.

9. A precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion as described in claim 1, characterized in that, The intelligent decision-making and early warning module supports a three-level early warning mechanism: the first level is a suspected fault, which displays a prompt message through the vehicle terminal; the second level is a confirmed fault, which activates the audible and visual alarm device. A Level 3 warning indicates impending thermal runaway, simultaneously triggering cloud-based alarms, user app notifications, and vehicle emergency system activation; the warning response time is no more than 1 second.

10. A precise early warning system for thermal runaway of power batteries based on AI multimodal feature fusion as described in claim 1, characterized in that, The system hardware adopts a modular design, adapting to lithium iron phosphate batteries and ternary lithium batteries, and supports pre-installation at the factory and retrofitting later. The software uses an embedded Linux operating system, and is customized to adapt to the application scenarios of energy storage batteries and electric ship power batteries. The customized development includes interface adaptation of hardware modules and parameter fine-tuning of software algorithms. The interface adaptation of hardware modules includes adjusting the sensor withstand voltage level to 1000V for high-voltage scenarios of energy storage batteries, and improving the hardware waterproof level to IP68 for the humid environment of electric ships. The parameter fine-tuning of software algorithms includes adjusting the switching threshold of dynamic convolution kernels and the weight coefficients of the reward function of reinforcement learning based on the battery operating conditions of different scenarios.

Images