Lithium battery safety state online diagnosis method based on acoustic-thermal-electric multi-physical field signals

Abstract

This invention belongs to the field of battery safety monitoring and fault diagnosis technology, specifically involving an online diagnostic method for the safety status of lithium batteries based on acoustic, thermoelectric, and multi-physics field signals. The method includes the following steps: collecting acoustic signature, temperature, and electrical raw data of the lithium battery, and preprocessing the raw data to obtain a sample dataset; extracting features from the sample dataset to obtain corresponding acoustic signature feature vectors, temperature feature vectors, current feature vectors, and voltage feature vectors; and enhancing and optimizing the extracted feature vectors based on a meta-learning model to generate a meta-enhanced feature set. This invention overcomes the limitations of single-signal diagnosis by fusing acoustic, thermal, and electrical multi-physics field signals, enabling comprehensive perception of changes in the internal mechanical structure of the battery, abnormal heat distribution, and electrochemical reaction imbalances, thus improving the sensitivity and accuracy of identifying early safety hazards.

Description

Technical Field

[0001] This invention belongs to the field of battery safety monitoring and fault diagnosis technology, specifically relating to an online diagnostic method for the safety status of lithium batteries based on acoustic, thermoelectric, and multi-physics field signals. Background Technology

[0002] With the rapid development of new energy vehicles and energy storage systems, the safety issues of lithium batteries are becoming increasingly prominent, urgently requiring efficient and reliable online diagnostic technologies to assess their safety status in real time. Traditional lithium battery status diagnostic methods mostly rely on single electrical signals (such as voltage, current, or impedance), which are difficult to comprehensively reflect the complex physicochemical changes within the battery. They suffer from insufficient sensitivity and response lag, especially in identifying early thermal runaway or micro-short circuits. In recent years, multi-physics fusion diagnostics has become a research hotspot, but existing technologies have not yet effectively integrated acoustic, thermal, and electrical signals for collaborative analysis, limiting diagnostic accuracy and applicable scenarios.

[0003] A search revealed a patent, CN113918889B, entitled "Online Aging Diagnosis Method for Lithium Batteries Based on Spatial Distribution Characteristics of Charging Data," with an authorization announcement date of July 19, 2024. This patent collects voltage and current data during the charging and discharging process of lithium batteries, extracts their spatial distribution characteristics, and combines them with a machine learning model to predict battery capacity, thereby achieving online diagnosis of aging status. However, this method relies solely on electrical signals (voltage and current) and does not incorporate multi-physical field information such as thermal or acoustic fields. It cannot detect non-electrical anomalies caused by dendrite growth, separator damage, or localized overheating within the battery, and lacks early warning capabilities for sudden safety risks (such as the initial stage of an internal short circuit). Furthermore, its diagnostic target focuses on capacity decay (aging) rather than encompassing broader safety states such as thermal runaway and mechanical damage, limiting its application scenarios.

[0004] A search revealed a patent, CN113484784B, entitled "An Online Aging Diagnosis Method for Lithium Batteries Based on Two-Point Impedance Aging Characteristics," with an authorization announcement date of July 8, 2022. This patent utilizes the electrochemical impedance characteristics at two frequencies under specific charging states to construct an aging diagnosis model, improving the practicality of the impedance method. However, this technology is still limited to the single electrical dimension of electrochemical impedance, and measurements must be performed during specific charging stages, making continuous monitoring under all operating conditions difficult. Furthermore, the impedance spectrum is sensitive to temperature changes, and this scheme does not simultaneously acquire or compensate for thermal signals, making it susceptible to environmental interference and potential misjudgments. More importantly, acoustic signals (such as changes in ultrasonic wave propagation characteristics that can reflect electrode structure deformation or gas evolution) are completely excluded from the diagnostic system, failing to capture critical safety events such as mechanical abuse or gas generation.

[0005] The aforementioned problems indicate that existing online diagnostic technologies for lithium batteries generally rely on a single electrical signal and lack comprehensive utilization of the coupling mechanisms of acoustic, thermal, and electrical multi-physics fields. This results in significant deficiencies in early safety hazard identification, adaptability to all operating conditions, and coverage of safety status dimensions. Therefore, this invention proposes an online diagnostic method for lithium battery safety status based on acoustic, thermal, and electrical multi-physics field signals. The aim is to construct a multi-source information fusion diagnostic model by simultaneously sensing acoustic response, temperature distribution, and electrical characteristics. This model achieves highly sensitive, real-time online early warning of multiple safety risks such as thermal runaway, internal short circuits, and mechanical damage in lithium batteries, significantly improving the safety and reliability of the battery system. Summary of the Invention

[0006] The purpose of this invention is to provide an online diagnostic method for the safety status of lithium batteries based on acoustic, thermoelectric, and multi-physical field signals. By fusing acoustic, thermal, and electrical signals, it overcomes the limitations of single-signal diagnosis and can comprehensively perceive changes in the internal mechanical structure of the battery, abnormal heat distribution, and imbalance of electrochemical reactions, thereby improving the sensitivity and accuracy of early safety hazard identification.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An online safety status diagnosis method for lithium batteries based on acoustic, thermoelectric, and multi-physics field signals includes the following steps: Acoustic fingerprints, temperature data, and raw electrical data of lithium batteries are collected, and the raw data are preprocessed to construct a sample dataset. Feature extraction is performed on the sample dataset to obtain corresponding voiceprint feature vectors, temperature feature vectors, current feature vectors and voltage feature vectors. The extracted feature vectors are then enhanced and optimized based on a meta-learning model to generate a meta-enhanced feature set. A primary security diagnostic model based on an attention mechanism is constructed, and the meta-enhanced feature set is time-aligned and spliced ​​together to output the primary security probability. Acoustic, thermal, and electrical single-mode models based on machine learning network structures are constructed respectively. The soundprint, temperature, current, and voltage feature vectors are input into the acoustic, thermal, and electrical single-mode models respectively. Through independent forward propagation, the soundprint prediction security probability, temperature prediction security probability, current prediction security probability, and voltage prediction security probability are output respectively. A weighted classification model is constructed to dynamically weight and fuse the primary safety probability, the voiceprint prediction safety probability, the temperature prediction safety probability, the current prediction safety probability, and the voltage prediction safety probability. The fusion result is then input into a pre-constructed and trained safety status diagnostic model for real-time discrimination to obtain the current safety status indicators of the lithium battery. Based on the comparison between the security status indicators and the preset threshold range, a graded early warning mechanism is triggered, and the corresponding level of early warning response strategy is activated.

[0008] The process involves collecting acoustic signatures, temperature data, and raw electrical data from the lithium battery, and preprocessing the data to obtain a sample dataset, specifically including: By using an array of ultrasonic transducers arranged on the surface of a battery cell or inside a module to transmit and receive ultrasonic signals in real time, the raw acoustic data is obtained, and noise reduction is performed using spectral subtraction to obtain the noise-reduced acoustic data. Raw temperature data of the battery surface and key nodes are collected using a distributed temperature sensor network, and median filtering is used to remove outliers and reduce noise, resulting in smooth temperature data. The battery terminal voltage and charging / discharging current are synchronously recorded by the voltage and current acquisition unit, and the filtered electrical data are obtained by using low-pass filtering for smoothing and noise suppression. Based on the timestamp, the denoised voiceprint data, temperature data, and electrical data within the same time segment are correlated and aligned to form a sample dataset.

[0009] The enhancement and transfer optimization of extracted feature information based on the meta-learning model specifically includes: From sample data covering various battery types, different aging stages, and various temperature conditions, multiple meta-training tasks and meta-testing tasks are divided; each meta-task is further divided into a support set and a query set to simulate the training and testing environment under a few-shot learning scenario. A deep neural network is used as the basic encoder to map the original feature vectors to a high-dimensional feature space in order to extract more expressive battery state features. Perform a model-independent meta-learning-based encoder meta-training process; on each meta-task, calculate the loss using the support set and perform gradient-intra-update to obtain task-specific encoder parameters. Then, use the updated encoder to calculate the loss on the query set, summarize the query set losses of all meta-tasks, and perform global optimization of the encoder's initial parameters through backpropagation until convergence. The trained meta-feature encoder is used to enhance the features of the original sample dataset, resulting in enhanced feature vectors. The enhanced feature vectors of all samples are associated with their corresponding battery state labels in chronological order to form a meta-enhanced feature set that can be used for subsequent modeling and analysis tasks.

[0010] The construction of the primary security diagnostic model based on the attention mechanism, which involves temporally aligning the meta-enhanced feature set and concatenating and fusing it into a primary security probability, specifically includes: Each sample in the meta-enhanced feature set is arranged in the original acquisition time order to form a multi-source feature time series; The feature vector of each time step in the multi-source feature time series contains enhanced acoustic features, temperature features, current features, and voltage features, which together constitute the time series input data. The multi-source feature time series is time-axis aligned using a dynamic time warping algorithm to eliminate the time phase difference between each feature source, resulting in a time-aligned feature series. The time-aligned feature sequence is input into a multi-head self-attention layer, and the long-range dependencies of each time step within the sequence are captured through a parallel attention mechanism, and the attention-enhanced features corresponding to each time step are output. The attention enhancement features from each output are concatenated and then subjected to a linear transformation to obtain a fused attention enhancement feature sequence. The attention-enhanced feature sequences are concatenated along the time dimension to form a global feature vector representing the entire time window; The global feature vector is input into a fully connected layer and mapped to a primary safety probability value through a Softmax activation function. The primary safety probability value is used to characterize the probability that the current battery sample is in a safe state.

[0011] The feature sequence is specifically as follows: Using the current characteristic sequence as a benchmark, the optimal matching path between other characteristic sequences and the benchmark sequence is calculated, and all characteristic sequences are mapped onto a unified time grid to eliminate time misalignment and obtain time-aligned characteristic sequences.

[0012] The attention enhancement feature is specifically as follows: The feature sequence is linearly projected to generate a query matrix, a key matrix, and a value matrix; Attention weights are obtained by calculating the dot product of the query and the key, scaling and Softmax normalization are applied, and the value matrix is ​​weighted to output the attention enhancement features at each time step. Multiple attention heads are set up in parallel to perform attention mechanism operations. The outputs of each attention head are concatenated and then linearly transformed to obtain an attention-enhanced feature sequence that incorporates global context information.

[0013] The weighted classification model is specifically as follows: The weights are calculated based on the statistical characteristics of the current time-series input data, including feature variance, information entropy, or signal-to-noise ratio. Based on the real-time quality assessment results of each modality data, the contribution of the acoustic signature prediction security probability, temperature prediction security probability, current prediction security probability, voltage prediction security probability and primary security probability in the fusion process is adaptively adjusted. When the data quality of a certain modality deteriorates due to sensor failure, environmental interference, or transmission anomalies, the corresponding weight is automatically reduced.

[0014] The safety status diagnostic model employs a machine learning classifier, utilizes a meta-enhanced feature set and corresponding historical fault labels for offline supervised training, and calculates and outputs the current safety status index of the lithium battery through forward propagation.

[0015] An online safety status diagnostic system for lithium batteries based on acoustic, thermoelectric, and multi-physics field signals includes: A data acquisition module is used to collect acoustic signature, temperature, and raw electrical data of the lithium battery. The data acquisition module includes an ultrasonic transducer array, a distributed temperature sensor network, and a voltage and current acquisition unit arranged on the surface of the battery cell or inside the module. The ultrasonic transducer array is used to transmit and receive ultrasonic signals in real time to obtain raw acoustic signature data. The distributed temperature sensor network is used to collect raw temperature data of the battery surface and key nodes. The voltage and current acquisition unit is used to synchronously record raw electrical data of the battery terminal voltage and charging and discharging current. The data preprocessing module, whose input is connected to the data acquisition module, is used to perform spectral subtraction on the original voiceprint data to obtain denoised voiceprint data, to perform outlier removal and denoising on the original temperature data to obtain smoothed temperature data, to perform low-pass filtering on the original electrical data to obtain filtered electrical data, and to associate and align the denoised voiceprint data, temperature data and electrical data within the same time segment according to the timestamp to form a sample dataset; The feature extraction and enhancement module, whose input is connected to the data preprocessing module, is used to extract feature information of voiceprint, temperature, current and voltage parameters from the sample dataset, and enhance and optimize the extracted feature information based on the meta-learning model to generate a meta-enhanced feature set; the meta-learning model includes a meta-task set divided from sample data of different battery types, different aging stages and different temperature conditions, a transferable feature encoder, and optimized parameters trained using a model-independent meta-learning algorithm; The primary security diagnosis module, whose input is connected to the feature extraction and enhancement module, is used to construct a primary security diagnosis model based on the attention mechanism, perform time alignment on the meta-enhanced feature set, capture long-range dependencies of the sequence through a multi-head self-attention mechanism, and map them into primary security probabilities through a fully connected layer. A single-modal diagnostic module, whose input is connected to the feature extraction and enhancement module, includes an acoustic single-modal diagnostic model, a thermal single-modal diagnostic model, and an electrical single-modal diagnostic model, respectively constructed based on a machine learning network structure. The acoustic single-modal diagnostic model is used to input a voiceprint feature vector and output a voiceprint prediction safety probability. The thermal single-modal diagnostic model is used to input a temperature feature vector and output a temperature prediction safety probability. The electrical single-modal diagnostic model is used to input a current feature vector and a voltage feature vector and output a current prediction safety probability and a voltage prediction safety probability, respectively. The weighted fusion and safety diagnosis module, whose input terminals are connected to the primary safety diagnosis module and the single-modal diagnosis module respectively, is used to construct a weighted classification model. It dynamically weights and fuses the primary safety probability, voiceprint prediction safety probability, temperature prediction safety probability, current prediction safety probability, and voltage prediction safety probability to obtain a fused safety probability vector. This fused safety probability vector is then input into a pre-constructed and trained safety status diagnosis model, which calculates and outputs the current lithium battery safety status index through forward propagation. The safety status diagnosis model uses a multilayer perceptron or gradient boosting decision tree as a classifier and utilizes a meta-enhanced feature set and corresponding historical fault labels for offline supervised training. The early warning response module, whose input is connected to the weighted fusion and safety diagnosis module, is used to compare the safety status indicators with preset multi-level threshold ranges, trigger a graded early warning mechanism based on the comparison results, and start the corresponding level of early warning response strategy; the early warning response strategy includes one or more combinations of data recording and tracing, operating parameter adjustment, thermal management intervention, power limiting, safety isolation, and operation and maintenance dispatch.

[0016] The technical effects achieved by this invention are as follows: This invention overcomes the limitations of single-signal diagnosis by integrating acoustic, thermal, and electrical multi-physics field signals. It can comprehensively perceive changes in the internal mechanical structure of the battery, abnormal heat distribution, and imbalance of electrochemical reactions, thereby improving the sensitivity and accuracy of identifying early safety hazards.

[0017] This invention introduces a meta-learning model for feature enhancement and transfer optimization, effectively eliminating data distribution bias caused by individual battery differences, aging stages, and changes in operating conditions, enabling the model to maintain high diagnostic performance in different application scenarios. Simultaneously, the dynamic weighted fusion mechanism adaptively adjusts modal weights based on data quality, enhancing the system's robustness in complex environments.

[0018] This invention combines the outputs of a primary fusion model and a single-modal model, enabling earlier detection of abnormal trends. Combined with a tiered early warning mechanism, it can achieve multi-level alarms ranging from "attention" to "danger," and link these alarms with measures such as thermal management, power limiting, and safety isolation, buying valuable time for emergency response and maximizing the protection of personal and property safety. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0021] like Figure 1 As shown, the online diagnostic method for the safety status of lithium batteries based on acoustic, thermoelectric, and multi-physics field signals includes the following steps: By using an array of ultrasonic transducers arranged on the surface of a battery cell or inside a module to transmit and receive ultrasonic signals in real time, the raw acoustic data is obtained, and noise reduction is performed using spectral subtraction to obtain the noise-reduced acoustic data. Raw temperature data of the battery surface and key nodes are collected using a distributed temperature sensor network, and median filtering is used to remove outliers and reduce noise, resulting in smooth temperature data. The battery terminal voltage and charging / discharging current are synchronously recorded by the voltage and current acquisition unit, and the filtered electrical data are obtained by using low-pass filtering for smoothing and noise suppression. Based on the timestamp, the denoised voiceprint data, temperature data, and electrical data within the same time segment are correlated and aligned to form a sample dataset. Feature extraction was performed on the sample dataset to obtain corresponding voiceprint feature vectors, temperature feature vectors, current feature vectors, and voltage feature vectors, specifically: Short-time Fourier transform is performed on the denoised voiceprint time-series signal to obtain the voiceprint time-spectrum. Then, Mel frequency cepstral coefficients, wavelet packet decomposition energy spectrum, and intrinsic mode function energy moments obtained from empirical mode decomposition are extracted to form the primary voiceprint feature vector. At the same time, the voiceprint time-spectrum is input into a pre-trained convolutional neural network for automatic feature learning, and a deep voiceprint feature vector is output. The primary voiceprint feature vector and the deep voiceprint feature vector are concatenated to form the final voiceprint feature vector. For the filtered temperature time series data, a sliding window method is used to extract the statistical features within the window, including the mean, variance, maximum value, minimum value, rate of change, and temperature gradient; thermal inertia features are extracted in combination with the battery thermal property model; the temperature sequence is input into a long short-term memory network to extract the dynamic time series features of temperature; the above statistical features, thermal inertia features and time series features are concatenated to form a temperature feature vector; For the filtered current time series data, the current rate characteristics, current fluctuation amplitude, constant current charging duration, and pulse current response characteristics during the charging and discharging process are extracted; the current sequence is input into a bidirectional gated recurrent unit network to extract the dynamic time series characteristics of the current; the above features are concatenated to form a current feature vector; For the filtered voltage time series data, the average voltage, voltage plateau duration, voltage fluctuation amplitude, differential voltage curve features, and internal resistance estimate calculated by the voltage drop to current ratio during the charging and discharging process are extracted; the voltage sequence is input into a bidirectional gated recurrent unit network to extract the voltage dynamic time series features; the above features are concatenated to form a voltage feature vector; The four feature vectors mentioned above are associated according to the sample dimension to form the original feature vector set corresponding to each sample; Sample data covering different battery types, aging stages, and temperature conditions were selected from the sample dataset and divided into a meta-training task set and a meta-testing task set. Each meta-task contains a support set and a query set. The support set is used to simulate the rapid adaptation of the model under few-shot learning scenarios, and the query set is used to evaluate the task adaptation effect. The meta-task set is constructed as follows: a small number of samples are randomly selected from samples of the same battery type, the same aging stage, and similar temperature conditions to form the support set, and the remaining samples from the same task domain are selected to form the query set, so as to simulate the real scenario of few-shot learning across operating conditions. A transferable feature encoder based on a deep neural network is constructed to map the original feature vector to a high-dimensional feature space. The feature encoder adopts a multilayer perceptron or residual network structure, with the input layer dimension matching the dimension of the original feature vector and the output layer being an enhanced feature vector of a preset dimension. A model-independent meta-learning algorithm is used to meta-train the feature encoder. The original feature vector set of all samples in the sample dataset is input into the trained meta-feature encoder. The enhanced feature vector is obtained through forward propagation. The enhanced feature vector retains the physical meaning of the original features. At the same time, the cross-task optimization of meta-learning eliminates the distribution shift caused by individual battery differences, changes in operating conditions and different collection environments, thereby enhancing the generalization ability and transferability of the features. The enhanced feature vectors of all samples are arranged in the order of the original acquisition time and associated with the corresponding battery status labels to construct a meta-enhanced feature set. The battery status labels include health status, remaining service life or fault type. The meta-enhanced feature set is used as input for subsequent primary safety diagnostic models and single-modal diagnostic models, providing high-quality feature data support for the accurate assessment of lithium battery safety status. A basic security diagnostic model based on an attention mechanism is constructed. The meta-enhanced feature set is time-aligned and concatenated to form a basic security probability, specifically including: Each sample in the meta-enhanced feature set is arranged in the original acquisition time order to form a multi-source feature time series; Each time step feature vector in the multi-source feature time series contains enhanced acoustic features, temperature features, current features, and voltage features, which together constitute the time series input data; A dynamic time warping algorithm is used to align the time axis of the multi-source feature time series, eliminating the time phase difference between each feature source, and obtaining the time-aligned feature series. The feature series is based on the current feature series, and the optimal matching path between other feature series and the reference series is calculated. All feature series are then mapped onto a unified time grid to eliminate time misalignment and obtain the time-aligned feature series. The time-aligned feature sequence is input into a multi-head self-attention layer. The long-range dependencies between time steps within the sequence are captured through a parallel attention mechanism. First, the feature sequence is linearly projected to generate a query matrix, a key matrix, and a value matrix. Then, attention weights are obtained by calculating the dot product of the query and the key and performing scaling and Softmax normalization. The value matrix is ​​then weighted and summed to output the attention-enhanced features for each time step. Multiple parallel attention heads are used to perform the above operations respectively, and the outputs of each head are concatenated and linearly transformed to obtain an attention-enhanced feature sequence that incorporates global context information. Residual connections and layer normalization are applied to the multi-head attention output to improve training stability. The attention-enhanced feature sequences are concatenated along the time dimension to form a global feature vector representing the entire time window; The global feature vector is input into a fully connected layer and mapped to a primary safety probability value between 0 and 1 using the Softmax activation function. This primary safety probability value characterizes the likelihood that the current battery sample is in a safe state. Using the cross-entropy loss function as the optimization objective, the primary safety diagnosis model is trained under supervision by utilizing the meta-enhanced feature set and its corresponding safety status labels, and the model parameters are iteratively updated until convergence. After training, the model can output the primary safety probability in real time for the input multi-source feature time series, which can be used for subsequent battery safety status warning or diagnosis tasks. Single-mode acoustic, thermal, and electrical models are constructed based on machine learning network structures, as follows: An acoustic single-modal diagnostic model is constructed based on a hybrid architecture of convolutional neural networks and recurrent neural networks. The model takes the voiceprint feature vector as input. First, it extracts the local time-frequency features of the voiceprint signal through multiple one-dimensional convolutional layers. Each convolutional layer is followed by a batch normalization layer to accelerate model convergence, and a nonlinearity is introduced through the ReLU activation function. Then, a max pooling layer is used to reduce the dimensionality of the feature map, preserving salient features and reducing computational complexity. The output of the pooling layer is connected to a long short-term memory network layer to capture the dynamic evolution and long-range dependencies of the voiceprint signal over time. Finally, the output of the long short-term memory network layer is connected to a fully connected layer and mapped to the voiceprint prediction safety probability through a Softmax activation function. The voiceprint prediction safety probability characterizes the probability that the current battery is in a safe state based on the acoustic modality. The model is trained using supervised learning, optimizing network parameters with a cross-entropy loss function and iteratively updating weights through backpropagation until convergence. A thermal single-modal diagnostic model is constructed based on a deep feedforward neural network combined with gated recurrent units (ROUs). The model takes temperature feature vectors as input. First, it standardizes the input data through a batch normalization layer to eliminate the influence of dimensions and accelerate the training process. Then, multiple stacked gated recurrent unit layers are connected to extract the temporal dependencies and thermal inertia evolution features of the temperature sequence. Each gated recurrent unit layer is followed by a dropout layer to randomly deactivate some neurons and prevent overfitting. The output of the gated recurrent unit layers is then fed into a global average pooling layer to aggregate temporal information and obtain a fixed-length feature representation. Finally, a fully connected layer and a Softmax activation function output the temperature prediction safety probability, which characterizes the likelihood that the battery is in a safe state based on the thermal modality. The model training employs an adaptive moment estimation optimizer to minimize the cross-entropy loss between the predicted probability and the true label. An electrical single-modal diagnostic model is constructed based on a fusion architecture of a multilayer perceptron and a bidirectional long short-term memory network. The model takes current and voltage feature vectors as joint inputs. First, a feature concatenation layer fuses the two types of features along their feature dimensions to form a joint electrical feature vector. This joint feature vector is then fed into a bidirectional long short-term memory network layer to simultaneously capture the temporal coupling characteristics and long-range dependencies of current and voltage signals from both forward and reverse directions, fully exploring the intrinsic correlations between electrical parameters. The output of the bidirectional long short-term memory network layer is then fed into a global max pooling layer to extract significant temporal features. Finally, the pooling results are input into two parallel... The model employs fully connected branches, each consisting of multiple stacked fully connected layers, corresponding to the current prediction task and the voltage prediction task, respectively. Each branch ultimately outputs the current prediction safety probability and the voltage prediction safety probability through a Softmax activation function. The current prediction safety probability characterizes the probability that the battery is in a safe state based on the current mode, and the voltage prediction safety probability characterizes the probability that the battery is in a safe state based on the voltage mode. Model training utilizes a multi-task learning strategy, jointly optimizing the loss functions for both current and voltage prediction tasks. The total loss is a weighted sum of two cross-entropy losses. Backpropagation is used to synchronously update the network parameters of the shared layers and the task-specific layers. The constructed acoustic single-mode diagnostic model, thermal single-mode diagnostic model, and electrical single-mode diagnostic model are integrated and deployed to form a single-mode diagnostic module. For each input sample, the independent forward propagation calculations of the three models are executed in parallel, and the predicted safety probabilities of acoustic signature, temperature, current, and voltage are output synchronously. The four predicted safety probabilities describe the safety state of the battery from different physical dimensions, serving as the input basis for subsequent multimodal weighted fusion. Constructing a weighted classification model involves using the statistical characteristics of the current time-series input data as the basis for weight calculation. These statistical characteristics include feature variance, information entropy, or signal-to-noise ratio. Based on the real-time quality assessment results of each modality data, the contribution of the acoustic signature prediction security probability, temperature prediction security probability, current prediction security probability, voltage prediction security probability and primary security probability in the fusion process is adaptively adjusted. When the data quality of a certain modality deteriorates due to sensor failure, environmental interference, or transmission anomaly, the corresponding weight is automatically reduced. The primary safety probability, voiceprint, temperature, current and voltage predicted safety probability are dynamically weighted and fused, and then input into a pre-built and trained safety status diagnostic model for real-time discrimination. The safety status diagnostic model uses a machine learning classifier and uses meta-enhanced feature set and corresponding historical fault labels for offline supervised training. The model calculates and outputs the current safety status index of the lithium battery through forward propagation. Based on the comparison between the safety status indicators and the preset threshold range, a tiered early warning mechanism is triggered, and the corresponding level of early warning response strategy is activated, specifically including the following sub-steps: Multi-level threshold ranges corresponding to safety status indicators are pre-defined in the battery management system or cloud monitoring platform. The threshold ranges are calibrated based on the battery's historical operating data, fault case library, and safety standards, and include at least safety thresholds, warning thresholds, and danger thresholds. The safety probability value or status level in the safety status indicators forms a mapping relationship with the threshold ranges, which is used to classify different safety risk levels. The threshold ranges can be dynamically adjusted according to battery type, aging stage, and operating conditions to adapt to the safety monitoring needs of different application scenarios. The system compares the safety probability value in the safety status indicator with a preset multi-level threshold range in real time, or directly matches the status level with a preset risk level to determine the current risk level of the lithium battery. The risk level is divided into at least normal level, attention level, warning level and dangerous level, which correspond to the safe status, potential risk status, intervention-required status and emergency status of battery operation, respectively. When the safety probability value is in different threshold ranges, the system automatically maps it to the corresponding risk level to provide a basis for subsequent warning triggering. Based on the risk level assessment, the corresponding early warning mechanism will be triggered: When the system is determined to be at a normal level, it maintains normal monitoring status, does not trigger warnings, and only records and uploads data as usual. When the system is determined to be at the level of concern, it records the current status and automatically increases the data collection frequency, enters close monitoring mode, and at the same time prompts the abnormal battery status on the human-machine interface or monitoring platform, suggesting that it should pay more attention. When the warning level is determined, the system immediately issues a warning signal and pushes alarm information to the operation and maintenance personnel through sound and light alarms, SMS push, APP notification or monitoring platform pop-up window, prompting them to check or intervene in a timely manner. When the system is determined to be at a dangerous level, it triggers the highest level alarm. In addition to pushing emergency alarm information, it also sends an emergency control command to the battery management system, requiring immediate execution of safety protection actions. Based on different warning levels, corresponding warning response strategies will be activated. The response strategies shall include at least one or more of the following measures executed in conjunction: Data recording and traceability: The multi-source raw data, feature data and model output results within the preset time window before and after the warning time are stored locally or backed up in the cloud to generate warning logs for subsequent fault tracing, analysis and model optimization; Operating parameter adjustment: By dynamically adjusting the upper limit of charging and discharging current, voltage threshold or temperature control strategy through the battery management system, the operating stress of the battery is reduced and the performance degradation is delayed; Thermal management intervention: Activate active cooling or heating devices to regulate the battery temperature to a preset safe range to prevent thermal runaway or performance degradation due to low temperature; Power limiting: Limits the battery output power and restricts the charge and discharge rate to prevent thermal runaway caused by overcharging, over-discharging, or high-power conditions; Safety isolation: In extreme cases, physical isolation is achieved by disconnecting the battery from the load or charging equipment using a contactor or relay, ensuring system safety; Operation and maintenance dispatch: Automatically sends location information, fault codes and suggested handling measures to the terminals of operation and maintenance personnel to guide the rapid location and handling on site; User prompt: Display safety prompts on the human-machine interface of the vehicle or energy storage system to guide users to take safe operating procedures; The system feeds back warning trigger records, response strategy execution effects, and subsequent battery status changes to a cloud platform or expert system in real time for dynamic optimization of threshold range settings and warning response strategies. Through machine learning algorithms, historical warning data is mined and analyzed to identify the optimal combination of warning thresholds and response strategies under different operating conditions, continuously improving the accuracy and adaptability of the tiered warning model and forming a closed-loop optimized intelligent warning mechanism. The closed-loop feedback system can continuously improve itself based on actual operating data, achieving a dynamic balance between warning accuracy and false alarm rate.

[0022] like Figure 2 As shown, the online safety status diagnostic system for lithium batteries based on acoustic, thermoelectric, and multi-physics field signals includes: A data acquisition module is used to collect acoustic signature, temperature, and raw electrical data of the lithium battery. The data acquisition module includes an ultrasonic transducer array, a distributed temperature sensor network, and a voltage and current acquisition unit arranged on the surface of the battery cell or inside the module. The ultrasonic transducer array is used to transmit and receive ultrasonic signals in real time to obtain raw acoustic signature data. The distributed temperature sensor network is used to collect raw temperature data of the battery surface and key nodes. The voltage and current acquisition unit is used to synchronously record raw electrical data of the battery terminal voltage and charging and discharging current. The data preprocessing module, whose input is connected to the data acquisition module, is used to perform spectral subtraction on the original voiceprint data to obtain denoised voiceprint data, to perform outlier removal and denoising on the original temperature data to obtain smoothed temperature data, to perform low-pass filtering on the original electrical data to obtain filtered electrical data, and to associate and align the denoised voiceprint data, temperature data and electrical data within the same time segment according to the timestamp to form a sample dataset; The feature extraction and enhancement module, whose input is connected to the data preprocessing module, is used to extract feature information of voiceprint, temperature, current and voltage parameters from the sample dataset, and enhance and optimize the extracted feature information based on the meta-learning model to generate a meta-enhanced feature set; the meta-learning model includes a meta-task set divided from sample data of different battery types, different aging stages and different temperature conditions, a transferable feature encoder, and optimized parameters trained using a model-independent meta-learning algorithm; The primary security diagnosis module, whose input is connected to the feature extraction and enhancement module, is used to construct a primary security diagnosis model based on the attention mechanism, perform time alignment on the meta-enhanced feature set, capture long-range dependencies of the sequence through a multi-head self-attention mechanism, and map them into primary security probabilities through a fully connected layer. A single-modal diagnostic module, whose input is connected to the feature extraction and enhancement module, includes an acoustic single-modal diagnostic model, a thermal single-modal diagnostic model, and an electrical single-modal diagnostic model, respectively constructed based on a machine learning network structure. The acoustic single-modal diagnostic model is used to input a voiceprint feature vector and output a voiceprint prediction safety probability. The thermal single-modal diagnostic model is used to input a temperature feature vector and output a temperature prediction safety probability. The electrical single-modal diagnostic model is used to input a current feature vector and a voltage feature vector and output a current prediction safety probability and a voltage prediction safety probability, respectively. The weighted fusion and safety diagnosis module, whose input terminals are connected to the primary safety diagnosis module and the single-modal diagnosis module respectively, is used to construct a weighted classification model. It dynamically weights and fuses the primary safety probability, voiceprint predicted safety probability, temperature predicted safety probability, current predicted safety probability, and voltage predicted safety probability to obtain a fused safety probability vector. This fused safety probability vector is then input into a pre-constructed and trained safety status diagnosis model, which calculates and outputs the current lithium battery safety status index through forward propagation. The safety status index can be a continuous probability value between 0 and 1, representing the probability that the battery is in a safe state; or it can be a discrete state level mapped from this probability value, such as 'safe', 'caution', 'warning', or 'danger'. The safety status diagnosis model uses a multilayer perceptron or gradient boosting decision tree as a classifier and utilizes a meta-enhanced feature set and corresponding historical fault labels to complete offline supervised training. The early warning response module, whose input is connected to the weighted fusion and safety diagnosis module, is used to compare the safety status indicators with preset multi-level threshold ranges, trigger a graded early warning mechanism based on the comparison results, and start the corresponding level of early warning response strategy; the early warning response strategy includes one or more combinations of data recording and tracing, operating parameter adjustment, thermal management intervention, power limiting, safety isolation, and operation and maintenance dispatch. Example 1: Early warning of thermal runaway in electric vehicle power battery system

[0023] Scenario Description: This embodiment focuses on an electric vehicle traveling at high speed, whose power battery system consists of multiple lithium-ion battery modules. The battery management system (BMS) integrates the online diagnostic system proposed in this invention, aiming to monitor the battery status in real time and provide early warnings before thermal runaway occurs.

[0024] Implementation process: Data Acquisition and Preprocessing: Acoustic signals: An array of ultrasonic transducers arranged on the surface of each battery cell transmits and receives ultrasonic signals at a frequency of 1Hz; the raw acoustic data is processed by spectral subtraction to remove vehicle driving vibration and motor noise, resulting in noise-reduced sound wave flight time and energy attenuation data; Thermal signal: Distributed temperature sensors (such as NTC thermistors) deployed in key nodes inside the module (such as cell terminals and module busbars) collect temperature data at a frequency of 10Hz; median filtering is used to remove abnormal temperature spikes caused by poor contact or electromagnetic interference to obtain smooth temperature field distribution data. Electrical signals: The voltage and current acquisition unit inside the BMS synchronously records the terminal voltage of each battery cell and the charging and discharging current of the entire module (sampling frequency is 100Hz); a low-pass filter is used to filter out high-frequency noise to obtain electrical data that reflects the dynamic characteristics of the battery; The data preprocessing module aligns the acoustic, thermal, and electrical data segments within the same 100ms time window according to a unified timestamp to form a sample; Feature extraction and enhancement: The feature extraction module extracts feature vectors from each sample: voiceprint features (such as flight time change rate, attenuation coefficient), temperature features (such as average temperature, maximum temperature difference, temperature rise rate), current features (such as current ratio, current change rate), and voltage features (such as individual voltage dispersion, differential voltage). The feature enhancement module calls a meta-feature encoder pre-trained using the MAML algorithm in a laboratory environment with battery data of different aging levels and different temperature conditions. This encoder maps the original feature vector of the current sample to a high-dimensional space to generate an enhanced feature set, effectively eliminating the data distribution offset caused by the inconsistency of individual battery cells and the current operating conditions (such as high-speed driving and high-current discharge). Multimodal fusion diagnostics: Primary Diagnosis: Enhanced feature sequences from 100 consecutive time windows (corresponding to 10 seconds of historical data) are input into a primary safety diagnosis model based on an attention mechanism. The model aligns the feature sequences through dynamic time warping and then uses a multi-head self-attention mechanism to capture the following: Although voltage and temperature have not yet shown significant anomalies, a weak and continuous energy decay trend compared to the baseline has appeared in the acoustic signature feature sequence (which may indicate microstructural changes caused by membrane melting or lithium dendrite growth). The model outputs a "primary safety probability" of 0.85. Single-modal diagnostics: Simultaneously, the feature vector of the current time window is fed into the acoustic, thermal, and electrical single-modal models respectively; the acoustic model outputs a safety probability of 0.6, the thermal model outputs 0.95, and the electrical model outputs 0.98. Weighted fusion: The weighted fusion module evaluates the quality of each modality data; currently, the temperature and voltage signals have high quality and good signal-to-noise ratio; the acoustic signal has a slightly lower signal-to-noise ratio due to vehicle bumps; according to preset rules, the module assigns a medium weight (0.3) to the acoustic probability, a high weight (0.35 each) to the thermal and electrical probabilities, and a high weight (0.4) to the primary probability because it incorporates time-series information; the final fused safety probability vector is 0.86; Safety Status Assessment and Early Warning: The fused safety probability vector is input into the final safety status diagnostic model (an offline trained XGBoost classifier); the model outputs the current safety status index of the lithium battery as 0.7 (this index is a value between 0 and 1, the lower the value, the more dangerous it is). The early warning response module compares this safety indicator with a preset threshold: if the safety threshold is ≥0.9, the early warning threshold is 0.6 ≤ and <0.9, and the danger threshold is <0.6; then the current status is determined to be "early warning level". The system immediately triggers an alert: ① The in-vehicle central control screen displays "Battery system status is abnormal, it is recommended to go to a repair shop for inspection"; ② High-frequency data from 30 seconds before and after the alert is uploaded to the cloud; ③ The BMS automatically limits the maximum charging and discharging power to 80% of the current level to reduce battery thermal stress.

[0025] Technical Effects: This embodiment, by fusing acoustic signals, successfully captured early microstructural changes inside the battery before traditional signals such as voltage and temperature showed obvious abnormalities. This enabled early warning of potential thermal runaway risks, providing valuable time for the driver to react, several minutes earlier than traditional battery management system methods. Example 2: Rapid Detection of Mechanical Damage in Soft-Pack Lithium Batteries for Consumer Electronics

[0026] Scenario Description: A smartphone uses a pouch lithium battery. If the user accidentally drops the phone, it may damage the internal separator or misalign the electrodes. The diagnostic system of this invention is integrated into the phone's power management chip to quickly detect the battery's safety status after a drop.

[0027] Implementation process: Data Acquisition and Preprocessing: Acoustic signal: The miniature ultrasonic transducer integrated inside the mobile phone (which can be used as an additional function of the existing speaker / microphone assembly) immediately emits and receives an ultrasonic pulse of a specific frequency after a drop; due to the compact structure of the pouch battery, even slight deformation or delamination can significantly change the reflection and transmission characteristics of ultrasonic waves. Thermal and electrical signals: The phone's built-in temperature sensor and battery gauge chip simultaneously collect battery temperature, voltage, and current data; due to the possibility of a micro-short circuit during a drop, the electrical data (especially the voltage) may fluctuate momentarily; Feature extraction and enhancement: The feature extraction module quickly extracts features within a very short time window (e.g., 2 seconds) after the fall; voiceprint features may include peak attenuation of the echo signal and the degree of waveform distortion; electrical features include the magnitude of voltage drop and recovery time; temperature features may not change significantly within this short time window. The feature enhancement module calls a meta-learning model optimized for small sample scenarios of consumer electronics batteries. During the research and development phase, this model was meta-trained using a small amount of "drop test" data of different battery models. It can align the features of the current sample with the feature space of "healthy" and "damaged" samples in the laboratory, so that even with only a small amount of prior data, it can still effectively extract the sensitive features that characterize "mechanical damage". Multimodal fusion diagnostics: Primary Diagnosis and Single-Mode Diagnosis: Considering limited computing resources, the system may not use complex time-series primary diagnostic models, but will rely more on single-mode models; the acoustic single-mode model (a lightweight CNN) outputs a "soundprint prediction safety probability" of 0.3 based on echo characteristics (indicating a significant risk of damage); the electrical model outputs a probability of 0.9 based on voltage fluctuations (voltage has recovered, possibly only a momentary fluctuation); the thermal model outputs a probability of 1.0 (no temperature change); Weighted fusion: The system's default rule is that, in the "physical impact" scenario, the acoustic mode has the highest authority; therefore, the weighted fusion module assigns a very high weight (0.8) to the acoustic probability, while the electrical and thermal probabilities have lower weights (0.1 each); the primary probabilities may not be used in this simplified process or their weights may be set to 0; the final fused safety probability vector is 0.42. Safety Status Assessment and Early Warning: The final safety status diagnostic model (a simple decision tree) outputs a safety status index of "danger" based on the fusion probability. The warning response module immediately triggers the response: ① A warning pops up on the phone screen: "Battery may be damaged. Do not continue to use it. Go to the service center for inspection." ② The system forcibly limits the charging current to the minimum safe level to prevent thermal runaway caused by internal damage during charging. ③ The event log and diagnostic data are recorded locally for maintenance personnel to read.

[0028] Technical Results: This embodiment demonstrates the application of this method in consumer electronics products with limited computing resources and in specific physical damage scenarios. Through effective learning of small samples using a meta-learning model and dynamic weighting tailored to the scenario, the system can quickly and accurately make safety diagnoses based on acoustic signals most sensitive to mechanical damage, effectively avoiding safety risks faced by users due to continued use of damaged batteries.

[0029] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for online diagnosis of the safety status of lithium batteries based on acoustic, thermoelectric, and multi-physics field signals, characterized in that, Includes the following steps: Acoustic fingerprints, temperature, and raw electrical data of lithium batteries are collected, and the raw data are preprocessed to obtain a sample dataset. Feature extraction is performed on the sample dataset to obtain corresponding voiceprint feature vectors, temperature feature vectors, current feature vectors and voltage feature vectors. Based on the meta-learning model, all or part of the extracted feature vectors are enhanced and optimized by transfer to generate a meta-enhanced feature set. A primary security diagnostic model based on an attention mechanism is constructed, and the meta-enhanced feature set is time-aligned and spliced ​​together to form a primary security probability. Based on machine learning network structures, acoustic, thermal, and electrical single-mode models are constructed respectively. The sound signature, temperature, current, and voltage feature vectors are input into the corresponding acoustic, thermal, and electrical single-mode models respectively. Through independent forward propagation calculations, the corresponding sound signature, temperature, current, and voltage predicted security probabilities are output respectively. A weighted classification model is constructed to dynamically weight and fuse the primary safety probability, voiceprint, temperature, current and voltage predicted safety probabilities, and input it into a pre-constructed and trained safety status diagnostic model for real-time discrimination to obtain the current safety status indicators of the lithium battery.

2. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 1, characterized in that, The process involves collecting acoustic signatures, temperature data, and raw electrical data from the lithium battery, and preprocessing the data to obtain a sample dataset, specifically including: By using an array of ultrasonic transducers arranged on the surface of a battery cell or inside a module to transmit and receive ultrasonic signals in real time, the raw acoustic data is obtained, and noise reduction is performed using spectral subtraction to obtain the noise-reduced acoustic data. Raw temperature data of the battery surface and key nodes are collected using a distributed temperature sensor network, and median filtering is used to remove outliers and reduce noise, resulting in smooth temperature data. The battery terminal voltage and charging / discharging current are synchronously recorded by the voltage and current acquisition unit, and the filtered electrical data are obtained by using low-pass filtering for smoothing and noise suppression. Based on the timestamp, the denoised voiceprint data, temperature data, and electrical data within the same time segment are correlated and aligned to form a sample dataset.

3. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 1, characterized in that, The enhancement and transfer optimization of extracted feature information based on the meta-learning model specifically includes: From sample data covering various battery types, different aging stages, and various temperature conditions, multiple meta-training tasks and meta-testing tasks are divided; each meta-task is further divided into a support set and a query set to simulate the training and testing environment under a few-shot learning scenario. A deep neural network is used as the basic encoder to map the original feature vectors to a high-dimensional feature space in order to extract more expressive battery state features. Perform a model-independent meta-learning-based encoder meta-training process; on each meta-task, calculate the loss using the support set and perform gradient-intra-update to obtain task-specific encoder parameters. Then, use the updated encoder to calculate the loss on the query set, summarize the query set losses of all meta-tasks, and perform global optimization of the encoder's initial parameters through backpropagation until convergence. The trained meta-feature encoder is used to enhance the features of the original sample dataset, resulting in enhanced feature vectors. The enhanced feature vectors of all samples are associated with their corresponding battery state labels in chronological order to form a meta-enhanced feature set that can be used for subsequent modeling and analysis tasks.

4. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 1, characterized in that, The construction of the primary security diagnostic model based on the attention mechanism, which involves temporally aligning the meta-enhanced feature set and concatenating and fusing it into a primary security probability, specifically includes: Each sample in the meta-enhanced feature set is arranged in the original acquisition time order to form a multi-source feature time series; The feature vector of each time step in the multi-source feature time series contains enhanced acoustic features, temperature features, current features, and voltage features, which together constitute the time series input data. The multi-source feature time series is time-axis aligned using a dynamic time warping algorithm to eliminate the time phase difference between each feature source, resulting in a time-aligned feature series. The time-aligned feature sequence is input into a multi-head self-attention layer, and the long-range dependencies of each time step within the sequence are captured through a parallel attention mechanism, and the attention-enhanced features corresponding to each time step are output. The attention enhancement features output from multiple attention heads are concatenated along the feature dimension and then subjected to a linear transformation to obtain an attention enhancement feature sequence that incorporates global contextual information. The attention-enhanced feature sequences are concatenated along the time dimension to form a global feature vector representing the entire time window; The global feature vector is input into a fully connected layer and mapped to a primary safety probability value through a Softmax activation function. The primary safety probability value is used to characterize the probability that the current battery sample is in a safe state.

5. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 4, characterized in that, The feature sequence is specifically as follows: Using the current characteristic sequence as a benchmark, the optimal matching path between other characteristic sequences and the benchmark sequence is calculated, and all characteristic sequences are mapped onto a unified time grid to eliminate time misalignment and obtain time-aligned characteristic sequences.

6. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 4, characterized in that, The attention enhancement feature is specifically as follows: The feature sequence is linearly projected to generate a query matrix, a key matrix, and a value matrix; Attention weights are obtained by calculating the dot product of the query and the key, scaling and Softmax normalization are applied, and the value matrix is ​​weighted to output the attention enhancement features at each time step. Multiple attention heads are set up in parallel to perform attention mechanism operations. The outputs of each attention head are concatenated and then linearly transformed to obtain an attention-enhanced feature sequence that incorporates global context information.

7. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 1, characterized in that, The weighted classification model is specifically as follows: The weights are calculated based on the statistical characteristics of the current time-series input data, including feature variance, information entropy, or signal-to-noise ratio. Based on the real-time quality assessment results of each modality data, the contribution of the acoustic signature prediction security probability, temperature prediction security probability, current prediction security probability, voltage prediction security probability and primary security probability in the fusion process is adaptively adjusted. When the data quality of a certain modality deteriorates due to sensor failure, environmental interference, or transmission anomalies, the corresponding weight is automatically reduced.

8. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 1, characterized in that, The safety status diagnostic model employs a machine learning classifier, utilizes a meta-enhanced feature set and corresponding historical fault labels for offline supervised training, and calculates and outputs the current safety status index of the lithium battery through forward propagation.

9. The online diagnostic method for lithium battery safety status based on acoustic-thermal-electric multiphysics field signals according to claim 1, characterized in that, It also includes comparing the security status indicators with preset threshold ranges to trigger a graded early warning mechanism and activate the corresponding level of early warning response strategy.

10. An online diagnostic system for the safety status of lithium batteries based on acoustic, thermoelectric, and multi-physics field signals, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect acoustic signatures, temperature data, and raw electrical data of the lithium battery. The data preprocessing module, whose input end is connected to the data acquisition module, is used to perform preprocessing operations on the raw voiceprint data, raw temperature data, and raw electrical data to construct a sample dataset; The feature extraction and enhancement module, whose input is connected to the data preprocessing module, is used to extract feature information of voiceprint, temperature, current and voltage parameters in the sample dataset, and enhance and optimize the extracted feature information based on the meta-learning model to generate a meta-enhanced feature set. The primary security diagnosis module, whose input is connected to the feature extraction and enhancement module, is used to perform time alignment on the meta-enhanced feature set and capture long-range dependencies of the sequence through a multi-head self-attention mechanism, which are then mapped to primary security probabilities through a fully connected layer. A single-modal diagnostic module, whose input is connected to the feature extraction and enhancement module, is used to calculate the safe probability of voiceprint prediction, temperature prediction, current prediction, and voltage prediction, respectively. The weighted fusion and safety diagnosis module has its input terminals connected to the primary safety diagnosis module and the single-mode diagnosis module, respectively. It is used to dynamically weight and fuse the primary safety probability, the voiceprint predicted safety probability, the temperature predicted safety probability, the current predicted safety probability, and the voltage predicted safety probability, and to obtain the current safety status indicators of the lithium battery using the safety status diagnosis model. The early warning response module, whose input is connected to the weighted fusion and security diagnosis module, is used to compare the security status indicators with preset multi-level threshold ranges, trigger a graded early warning mechanism based on the comparison results, and start the corresponding level of early warning response strategy.

Images