Battery module thermal runaway monitoring and positioning early warning method

Abstract

The invention provides a battery module thermal runaway monitoring and positioning early warning method, and belongs to the technical field of battery module thermal runaway monitoring. A plurality of acoustic sensor arrays are arranged in a battery module energy storage bin to collect acoustic signals in real time, and a target acoustic event feature library containing safety valve opening and battery exhaust features is constructed; a sparse expert hybrid acoustic recognition model is adopted to intelligently recognize and classify acoustic signals, an acoustic region division mechanism based on an exhaust sound attenuation propagation equation is established, and after a safety valve opening event is detected, a thermal runaway module is quickly locked through cross-channel sound pressure level peak value comparison. Battery exhaust acoustic features are continuously monitored, sound pressure level attenuation gradients are calculated and matched with a regional acoustic feature library to achieve accurate regional positioning, meanwhile, a diffusion path is predicted based on a thermal runaway propagation kinetic equation, model parameters are dynamically adjusted and recognized, and the technical problem that a thermal runaway source is difficult to accurately position is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of battery module thermal runaway monitoring, and in particular, relates to a battery module thermal runaway monitoring and positioning early warning method. BACKGROUND

[0002] As an important part of new energy technology, battery energy storage systems are widely used in electric vehicles, energy storage power stations, and portable electronic devices. The battery module thermal runaway monitoring technology is a key technology to ensure the safe operation of the system. Traditional battery thermal runaway monitoring methods mainly rely on temperature sensor arrays and smoke detection systems for monitoring. By setting up a multi-point temperature monitoring network or a smoke concentration detection device in the energy storage warehouse, the location of the thermal runaway event can be determined. These methods have formed a relatively mature technical system in industrial applications and are widely deployed in various battery energy storage facilities. However, traditional positioning technology has the defects of insufficient accuracy and response lag. Temperature sensors need to rely on heat conduction to detect temperature anomalies. In the environment of densely arranged battery modules, heat will quickly spread to adjacent areas, causing temperature distribution to be blurred, making it difficult to accurately determine the location of the initial thermal runaway source. Although the smoke detection system can detect the smoke generated by thermal runaway, the smoke will quickly spread and mix in a closed space, and it is also difficult to provide accurate source positioning information. In the current large-scale energy storage warehouse application scenario, due to the large number of battery modules and the close arrangement, once a thermal runaway event occurs, traditional technology can only determine the approximate area range and cannot accurately lock the specific thermal runaway source battery or module, resulting in a lack of targeted emergency response measures. That is, the existing technology has the technical problem of difficulty in accurately positioning the thermal runaway source after the occurrence of battery module thermal runaway. SUMMARY

[0003] Therefore, the application provides a battery module thermal runaway monitoring and positioning early warning method, which can solve the technical problem of difficulty in accurately positioning the thermal runaway source after the occurrence of battery module thermal runaway in the prior art.

[0004] The application is implemented in the following manner: the application provides a battery module thermal runaway monitoring and positioning early warning method, which comprises arranging a plurality of acoustic sensors in a battery module energy storage bin according to an acoustic sensor array arrangement rule, constructing a target acoustic event feature library containing a safety valve opening acoustic feature and a battery exhaust acoustic feature, processing and identifying a target acoustic event from a real-time collected acoustic signal by using a sparse expert hybrid acoustic recognition model, establishing an acoustic region division mechanism in the module, positioning a thermal runaway module by comparing the peak values of the sound pressure levels across the channels, continuously collecting acoustic signals and extracting battery exhaust acoustic features after locking the thermal runaway module, matching and positioning the regions in the module according to the sound pressure level attenuation gradient and the pre-divided region acoustic feature library, predicting the thermal runaway diffusion path based on the thermal runaway propagation dynamics equation and dynamically adjusting the sparsity parameter of the expert activation mechanism in the sparse expert hybrid acoustic recognition model, and realizing accurate monitoring and positioning of the battery module thermal runaway through acoustic signal feature recognition and multidimensional space positioning.

[0005] The acoustic sensor array arrangement rule is specifically determined according to the module dense arrangement characteristics in the energy storage bin, each acoustic sensor has a coverage radius of 0.5-1.0 m, and the distance between adjacent sensors is not more than 2.0 m, so as to ensure the spatial continuity and redundancy of the acoustic signal collection.

[0006] The safety valve opening acoustic feature is specifically the sound produced by the instantaneous opening of the safety valve when the internal pressure of the battery exceeds the safety threshold, which is characterized by a 20-50 ms burst pulse and a sharp increase in energy in the 500-1500 Hz frequency domain, and has the characteristics of short pulse and high frequency.

[0007] The battery exhaust acoustic feature is specifically the broadband noise signal produced during the continuous exhaust of the battery internal gas after the safety valve is opened, which is characterized by a 2000-5000 Hz frequency band broadband noise and a continuous sound pressure level attenuation.

[0008] The sparse expert hybrid acoustic recognition model is specifically a hybrid architecture containing 8 expert networks, each of which processes different frequency domain ranges or acoustic event types, and the expert selection mechanism dynamically activates 2-3 most relevant experts for collaborative processing through a gating network according to the input signal frequency spectrum characteristics and energy distribution.

[0009] The acoustic region division mechanism is specifically determined according to the module length L and the positions of the M acoustic sensors to establish a spatial coordinate system, divide the module into P equidistant initial intervals, and then perform acoustic calibration to form N locatable unit regions according to the exhaust sound attenuation propagation equation.

[0010] The exhaust sound attenuation propagation equation is specifically an equation for describing energy attenuation rules of battery exhaust sound during propagation in the module, and comprehensively considers influences of geometric diffusion, medium absorption, multi-path reflection, and obstacle shielding of sound waves in a closed space on sound pressure levels.

[0011] The cross-channel sound pressure level peak value contrast is specifically that instantaneous sound pressure level data of multiple acoustic sensor channels are synchronously acquired, the sound source direction is determined by real-time comparison of peak values of each channel, and fast coarse positioning is realized by using a physical law that sound energy attenuates with distance.

[0012] The sound pressure level attenuation gradient is specifically a sound pressure level change rate of a battery exhaust acoustic characteristic signal in a time dimension, is obtained by calculating a sound pressure level difference between continuous sampling points and dividing by a time interval, and reflects dynamic characteristics and sound source intensity change trends of an exhaust process.

[0013] The thermal runaway propagation dynamics equation is specifically an equation for predicting a diffusion process of thermal runaway in the battery module, is established based on heat conduction theory and electrochemical reaction dynamics, and comprehensively considers thermal coupling effects between batteries, heat dissipation conditions, and material thermal physical parameters.

[0014] Before the step of constructing the target acoustic event feature library, the method further includes establishing a sparse expert hybrid acoustic recognition model training data set, collecting thermal runaway experimental data of different types of lithium ion batteries in a controlled environment, and acquiring acoustic data at different spatial positions by using a multi-channel synchronous acquisition mode. After the step of establishing the sparse expert hybrid acoustic recognition model training data set, the method further includes sparse expert hybrid acoustic recognition model training, and a multi-stage training strategy is adopted, that is, first, each expert network is independently pre-trained, and then a joint training stage is performed to optimize the entire hybrid architecture in an end-to-end manner.

[0015] The regional acoustic feature library is specifically a database containing acoustic fingerprint features of each locatable unit region, and features of each region include typical sound pressure level attenuation patterns, spectral distribution characteristics, and time domain waveform characteristics.

[0016] The sparsity parameter is specifically a regulation parameter for controlling the number of activated experts in the sparse expert hybrid acoustic recognition model, and is dynamically adjusted according to a prediction result of the thermal runaway propagation dynamics equation, a complexity of a current acoustic signal, and a real-time calculation resource status.

[0017] The acoustic fingerprint feature is specifically a hierarchical feature including a 1-time distance region sound pressure level attenuation amplitude of 0.1 to 0.2, a 2-time distance region sound pressure level attenuation amplitude of 0.2 to 0.4, a 3-time distance region sound pressure level attenuation amplitude of 0.4 to 1.0, and a 4-time distance region sound pressure level attenuation amplitude of 1.0 to 1.6.

[0018] The adjustment range of the sparsity parameter is 0.2 to 0.8, when the prediction result shows that the risk of thermal runaway diffusion is high, the sparsity parameter is adjusted to 0.6 to 0.8 to activate more experts to improve the identification accuracy, and when the acoustic signal is relatively simple, the sparsity parameter is adjusted to 0.2 to 0.4 to save computing resources.

[0019] Optionally, the value range of M in the M acoustic sensors is 4 to 12, and the plurality of acoustic sensors are uniformly distributed along the length direction of the module, and each acoustic sensor corresponds to monitor the acoustic signal of the module area.

[0020] The value of P in the P equidistant initial intervals is determined according to the module length L and the number of acoustic sensors M, and the module is divided into a plurality of initial intervals according to the equidistant principle, which provides a basis for subsequent acoustic correction.

[0021] The value of N in the N locatable unit areas is formed after the P equidistant initial intervals are acoustically corrected by the exhaust sound attenuation propagation equation, and each locatable unit area has unique acoustic propagation characteristics and attenuation modes.

[0022] Further, the current positioning result is input as an initial condition, historical propagation data is used as a boundary condition constraint, and environmental parameters are substituted into the thermal runaway propagation dynamics equation as solving parameters to solve the diffusion trajectory, and at the same time, the sparsity parameter of the expert activation mechanism in the sparse expert hybrid acoustic identification model is dynamically adjusted according to the solving result to realize real-time tracking and early warning of the thermal runaway propagation trend.

[0023] The present application realizes accurate positioning of the thermal runaway source by constructing a multi-acoustic sensor array to collect battery module acoustic signals in real time, and combining a sparse expert hybrid acoustic identification model to identify key acoustic events such as safety valve opening and battery exhaust. The present application overcomes the defects of traditional temperature and smoke monitoring methods in positioning, and can determine the sound source direction at the moment of safety valve opening through the cross-channel sound pressure level peak comparison technology, quickly lock the specific module where thermal runaway occurs, and then through continuous monitoring of the acoustic characteristics of the battery exhaust and analysis of the sound pressure level attenuation gradient, combined with the pre-established regional acoustic feature library for matching identification, realize the fine positioning of the regional level in the module. The present application establishes an acoustic region division mechanism based on the exhaust sound attenuation propagation equation, accurately partitions the internal space of the module according to the acoustic characteristic difference, each region has unique acoustic fingerprint characteristics, and through the accurate mapping and correlation of acoustic signals and physical positions, the thermal runaway source can be positioned to the specific battery unit level. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 Flow chart of the method of the present application.

[0025] Figure 2 Data graph of the attenuation of exhaust sound signals in different areas.

[0026] Figure 3 Neural network structure diagram for sparse expert mixed acoustic recognition model.

[0027] Figure 4 Graph of the sound pressure level attenuation curve of the battery exhaust process in the embodiment. DETAILED DESCRIPTION

[0028] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0029] As Figure 1 shown is a flow chart of a battery module thermal runaway monitoring and positioning early warning method provided by the present application, the method comprises the following steps:

[0030] S01, a plurality of acoustic sensors are arranged in the battery module energy storage warehouse according to the acoustic sensor array arrangement rule, the plurality of acoustic sensors are uniformly distributed along the length direction of the module, and each acoustic sensor corresponds to monitor the acoustic signal of the module area;

[0031] S02, a target acoustic event feature library is constructed, the target acoustic event feature library contains a safety valve opening acoustic feature and a battery exhaust acoustic feature, wherein the safety valve opening acoustic feature is characterized by a 20 to 50 millisecond burst pulse and a sharp increase in energy in the 500 to 1500 hertz frequency domain, and the battery exhaust acoustic feature is characterized by a 2000 to 5000 hertz frequency band wide frequency noise and a sound pressure level sustained attenuation;

[0032] S03, a sparse expert mixed acoustic recognition model is used to process the real-time collected acoustic signal, the sparse expert mixed acoustic recognition model recognizes the target acoustic event through time-frequency domain feature matching and outputs the event type and confidence;

[0033] S04, a module acoustic area division mechanism is established, a spatial coordinate system is established according to the module length L and the positions of the M acoustic sensors, wherein M is in the range of 4 to 12, the module is divided into P equidistant initial intervals, and then the acoustic school is established according to the exhaust sound attenuation propagation equation to form N locatable unit areas;

[0034] S05, when the safety valve opening event is recognized, the thermal runaway module positioning is realized through the cross-channel sound pressure level peak value comparison, the sensor channel corresponding to the maximum sound pressure level is mapped to the associated battery module and determined as the thermal runaway source;

[0035] S06, after locking the thermal runaway module, continuously collecting the acoustic signals of the module and extracting the battery exhaust acoustic features, calculating the sound pressure level decay gradient according to the set sampling period, and when the absolute value of the sound pressure level decay gradient exceeds the critical threshold, matching with the pre-divided regional acoustic feature library to realize regional positioning in the module;

[0036] S07, predicting the thermal runaway diffusion path based on the thermal runaway propagation dynamics equation, inputting the current positioning result as the initial condition, historical propagation data as the boundary condition constraint, and environmental parameters as the solving parameters into the thermal runaway propagation dynamics equation to solve the diffusion trajectory, and simultaneously dynamically adjusting the sparsity parameters of the expert activation mechanism in the sparse expert hybrid acoustic recognition model to realize real-time tracking and early warning of the thermal runaway propagation trend.

[0037] Among them, the attenuation data of the exhaust sound signal in different regions is as shown in Figure 2 .

[0038] Among them, the acoustic sensor array arrangement rule refers to that according to the dense arrangement characteristics of the module in the energy storage warehouse, the coverage radius of each acoustic sensor is 0.5 meters to 1.0 meter, and the distance between adjacent sensors is not more than 2.0 meters, so as to ensure the spatial continuity and redundancy of acoustic signal collection.

[0039] Among them, the safety valve opening acoustic feature refers to the sound produced by the instantaneous opening of the safety valve when the pressure in the battery exceeds the safety threshold, and the sound has the characteristics of short pulse and high frequency, which is an important acoustic sign of early thermal runaway.

[0040] Among them, the battery exhaust acoustic feature refers to the broadband noise signal generated during the continuous discharge of the battery internal gas after the safety valve is opened, and the signal presents a sound pressure level decay law over time, which is used to realize accurate regional positioning.

[0041] As shown in Figure 3 , the specific structure of the sparse expert hybrid acoustic recognition model is a mixed architecture containing 8 expert networks, each expert network processes different frequency domain ranges or acoustic event types, the first to the third experts are responsible for the safety valve opening event identification of low, medium and high frequency bands respectively, the fourth to the sixth experts are responsible for the battery exhaust acoustic feature extraction of high sound pressure level range 90 to 120 decibels, medium sound pressure level range 60 to 90 decibels and low sound pressure level range 30 to 60 decibels respectively, the seventh expert processes background noise suppression, and the eighth expert is responsible for multi-event concurrent identification. The expert selection mechanism dynamically activates 2 to 3 most relevant experts for collaborative processing through the gating network according to the frequency spectrum characteristics and energy distribution of the input signal.

[0042] The step of establishing the training data set of the sparse expert hybrid acoustic recognition model specifically includes collecting thermal runaway experiment data of different types of lithium ion batteries in a controlled environment, recording complete acoustic signal sequences from abnormal temperature rise of the battery to opening of the safety valve to end of exhaust, acquiring acoustic data at different spatial positions using a multi-channel synchronous acquisition method, pre-processing the original acoustic signals including noise reduction, normalization and time-frequency domain transformation, constructing a balanced data set containing positive samples and negative samples, wherein the positive samples are acoustic features of real thermal runaway events, and the negative samples are interference signals such as normal operating state, mechanical vibration and environmental noise, and finally forming a comprehensive training data set covering multiple battery types, multiple environmental conditions and multiple thermal runaway triggering modes.

[0043] The step of training the sparse expert hybrid acoustic recognition model specifically includes using a multi-stage training strategy, first pre-training each expert network independently to achieve basic performance on sub-tasks, then performing a joint training stage to optimize the entire hybrid architecture in an end-to-end manner, using adaptive loss weight to balance the contribution of different experts during training, using data enhancement techniques including time domain stretching, frequency domain shifting and noise superposition to improve the generalization ability of the model, introducing an adversarial training mechanism to enhance the robustness of the model to environmental interference, dynamically adjusting the sparsity parameter of expert activation during training to balance the calculation efficiency and recognition accuracy, and finally evaluating the model performance through cross-validation and independent test set and optimizing the hyperparameters.

[0044] The exhaust sound attenuation propagation equation is used to describe the energy attenuation law of the battery exhaust sound during the propagation process inside the module, the equation comprehensively considers the influence of geometric diffusion, medium absorption, multi-path reflection and obstacle shielding of sound waves in a closed space on the sound pressure level, the input includes the initial sound pressure level of the sound source, the propagation distance, the medium density, the environmental temperature and the internal structure parameters of the module, and the output is the sound pressure level value at any position, which is used to establish the acoustic feature fingerprint of each locatable unit region in the acoustic region division mechanism of the module.

[0045] The thermal runaway propagation dynamics equation is used to predict the diffusion process of thermal runaway in the battery module, the equation is established based on heat conduction theory and electrochemical reaction dynamics, and comprehensively considers the thermal coupling effect between batteries, heat dissipation conditions and material thermal physical parameters, the input includes the temperature of the initial thermal runaway battery from the temperature monitoring system, the distance between adjacent batteries from the module structure parameters, the environmental temperature from the environmental monitoring system, the heat dissipation coefficient from the material database and the battery heat capacity from the battery specification parameters, and the output is the spatiotemporal evolution trajectory and impact range prediction result of thermal runaway propagation, which is used to dynamically adjust the sparsity parameter of the expert activation mechanism in the sparse expert hybrid acoustic recognition model.

[0046] The acoustic region division mechanism refers to a method of dividing the internal space of the battery module according to acoustic characteristic differences, each region having unique acoustic propagation characteristics and attenuation modes, and the acoustic fingerprint includes a hierarchical feature of a 1-time distance region sound pressure level attenuation amplitude of 0.1 to 0.2, a 2-time distance region sound pressure level attenuation amplitude of 0.2 to 0.4, a 3-time distance region sound pressure level attenuation amplitude of 0.4 to 1.0, and a 4-time distance region sound pressure level attenuation amplitude of 1.0 to 1.6.

[0047] The cross-channel sound pressure level peak value contrast refers to a positioning method of determining the direction of a sound source by synchronously acquiring instantaneous sound pressure level data of multiple acoustic sensor channels and comparing the peak values of each channel in real time. The method uses the physical law of sound energy attenuation with distance to achieve rapid coarse positioning. The instantaneous sound pressure level data is derived from real-time acquisition of multiple acoustic sensors.

[0048] The sound pressure level attenuation gradient refers to the sound pressure level change rate of the battery exhaust acoustic characteristic signal in the time dimension, which is obtained by calculating the sound pressure level difference between consecutive sampling points and dividing by the time interval. The parameter reflects the dynamic characteristics and sound source intensity variation trend of the exhaust process. The sound pressure level difference is derived from continuously acquired module acoustic signals.

[0049] The region acoustic feature library refers to a pre-established database containing acoustic fingerprint features of each locatable unit region. The features of each region include typical sound pressure level attenuation patterns, spectral distribution characteristics, and time-domain waveform characteristics, which are used to achieve rapid matching and identification of real-time acoustic signals and physical locations. The acoustic fingerprint features are derived from the calculation results of the exhaust sound attenuation propagation equation.

[0050] The sparsity parameter refers to a regulation parameter that controls the number of activated experts in the sparse expert hybrid acoustic identification model. The sparsity parameter is dynamically adjusted according to the prediction results of the thermal runaway propagation dynamics equation, the complexity of the current acoustic signal, and the real-time calculation resource status. The adjustment range of the sparsity parameter is 0.2 to 0.8. When the prediction results show that the risk of thermal runaway diffusion is high, the sparsity parameter is adjusted to 0.6 to 0.8 to activate more experts to improve identification accuracy. When the acoustic signal is relatively simple, the sparsity parameter is adjusted to 0.2 to 0.4 to save calculation resources.

[0051] The specific implementation of the above steps is described in detail below.

[0052] The specific implementation of step S01 is to determine the optimal arrangement position of the sensor array through an acoustic sensor spatial geometry distribution algorithm. First, according to the dense arrangement characteristics and geometric size parameters of the module in the energy storage bin, the coverage radius of each acoustic sensor is calculated to be 0.5 to 1.0 m using the hexagonal dense layout principle, ensuring that the distance between adjacent sensors is not more than 2.0 m. Then, based on the spherical diffusion model of sound wave propagation, the uniform distribution spacing along the length direction of the module is calculated through the minimum redundancy constraint and maximum coverage optimization algorithm. Next, using the spatial continuity principle of acoustic signals, sensors are set at key node positions of the module, so that any acoustic event at any position can be monitored by at least two sensors simultaneously. Finally, through the sensor network topology optimization algorithm, the entire array is ensured to have sufficient redundancy and fault tolerance, so that when a single sensor fails, it will not cause a monitoring blind area.

[0053] The specific implementation of step S02 is to construct a target acoustic event feature library using an acoustic signal feature extraction and classification algorithm. First, the feature of the safety valve opening acoustic signal is extracted through time-frequency domain analysis method, and the short-time Fourier transform algorithm is used to analyze the burst pulse characteristics in the 20 to 50 ms time window, and the power spectral density estimation method is used to determine the energy abrupt increase characteristics in the 500 to 1500 Hz frequency range. Then, the wideband noise characteristics of the battery exhaust acoustic signal are analyzed, and the continuous wavelet transform algorithm is used to extract the wideband distribution characteristics in the 2000 to 5000 Hz frequency band, and the sound pressure level time domain attenuation analysis method is used to establish the continuous attenuation rule model. Next, the clustering algorithm is used to classify and label different types of acoustic features, and a multi-dimensional feature vector library containing time domain features, frequency domain features and energy features is established. Finally, the feature template matching algorithm is used to verify the integrity and accuracy of the feature library, ensuring that the target event and environmental interference signals can be effectively distinguished.

[0054] The specific implementation of step S03 is to use a sparse expert hybrid neural network architecture for intelligent identification and processing of real-time acoustic signals. First, the real-time collected acoustic signals are input into the preprocessing module, and the digital filtering algorithm is used to remove noise interference, and the normalization algorithm is used to unify the signal amplitude range. Then, the frequency spectrum characteristics and energy distribution of the input signal are analyzed using the gating network mechanism, and the most relevant 2 to 3 expert networks are dynamically selected for collaborative processing according to the signal features. Next, the activated expert networks process different acoustic features in parallel, the first to third experts use low, medium and high frequency filter combined convolutional neural networks to identify safety valve opening events, and the fourth to sixth experts use sound pressure level segmentation processing algorithms to extract battery exhaust features in different intensity ranges. Then, the seventh expert uses an adaptive noise suppression algorithm to process background interference, and the eighth expert uses a multi-task learning algorithm to process concurrent event identification. Finally, the output results of each expert are integrated through a weighted fusion algorithm to calculate the event type probability distribution and confidence score.

[0055] The specific implementation of step S04 is to establish a spatial region division mechanism in the module based on an acoustic propagation physical model. First, a three-dimensional rectangular coordinate system is established according to the module length L and M acoustic sensor positions, where M ranges from 4 to 12, and the module is divided into P equidistant initial intervals through a geometric analysis algorithm. Then, an acoustic propagation modeling algorithm is used to establish an exhaust sound attenuation propagation equation, taking into account factors such as geometric diffusion, medium absorption, multipath reflection, and obstacle shielding of sound waves in a closed space. Next, a finite element numerical solution method is used to calculate the energy attenuation law of sound waves during propagation inside the module, inputting the initial sound pressure level of the sound source, propagation distance, medium density, environmental temperature, and module internal structure parameters, and outputting the sound pressure level value at any position. Then, the sound school is re-divided to form N locatable unit regions, each with unique acoustic propagation characteristics and attenuation patterns. Finally, an acoustic fingerprint establishment algorithm is used to generate sound pressure level attenuation amplitude grading features for each region, including 1x, 2x, 3x, and 4x distances, corresponding to 0.1-0.2, 0.2-0.4, 0.4-1.0, and 1.0-1.6 attenuation amplitude ranges, respectively.

[0056] The specific implementation of step S05 is to use a multi-channel sound pressure level peak comparison algorithm to achieve rapid positioning of the thermal runaway module. First, the instantaneous sound pressure level data of multiple acoustic sensor channels is synchronously acquired, and a high-precision analog-to-digital converter is used to ensure the time synchronization and amplitude consistency of the data in each channel. Then, a peak detection algorithm is used to identify the sound pressure level peaks of each channel in real time, and a sliding window maximum value search method is used to find the maximum sound pressure level value within a preset time window. Next, using the inverse square law of sound energy distance attenuation, the sound source direction is determined through cross-channel sound pressure level peak comparison analysis, with the sensor channel having the maximum sound pressure level corresponding to the position closest to the sound source. Then, a sensor-module mapping algorithm is used to accurately map the sensor channel corresponding to the maximum sound pressure level to the associated battery module, and through a pre-established spatial topology relationship database, the physical location of the thermal runaway source is determined. Finally, a confidence evaluation algorithm is used to verify the reliability of the positioning result, and when the confidence exceeds the preset threshold of 0.8, the thermal runaway module positioning is confirmed to be effective.

[0057] The specific implementation of step S06 is to realize accurate regional positioning in the module through the sound pressure level attenuation gradient analysis algorithm. First, lock the thermal runaway module and continuously collect the acoustic signals of the module, and use a high sampling rate data acquisition system to ensure the time resolution and amplitude accuracy of the signals. Then use the battery exhaust acoustic feature extraction algorithm, extract the broadband noise signal in the frequency band of 2000 to 5000 Hz through a band-pass filter, and use the sound pressure level calculation algorithm to obtain the real-time sound pressure level value. Then calculate the sound pressure level attenuation gradient according to the set sampling period, and obtain the attenuation gradient value by dividing the sound pressure level difference between consecutive sampling points by the time interval. This parameter reflects the dynamic characteristics and sound source intensity variation trend of the exhaust process. Then use the threshold judgment algorithm, when the absolute value of the sound pressure level attenuation gradient exceeds the critical threshold, trigger the regional positioning process, the critical threshold is usually set to 0.5 to 1.0 dB / s. Then use the pattern matching algorithm to compare the current acoustic features with the pre-divided regional acoustic feature library, and determine the regional position of the sound source through similarity calculation and best matching search algorithm. Finally, use the probability fusion algorithm to integrate multiple feature matching results, and output the final regional positioning result and confidence evaluation in the module.

[0058] The specific implementation of step S07 is to use the thermal runaway propagation dynamics modeling algorithm to predict the diffusion path and realize real-time tracking and early warning. First, based on the heat conduction theory and electrochemical reaction dynamics, establish the thermal runaway propagation dynamics equation, considering the thermal coupling effect between batteries, heat dissipation conditions and material thermal physical parameters. Then use the current positioning result as the initial condition, including the temperature, position and thermal power of the initial thermal runaway battery and other parameters, use the historical propagation data as the boundary condition constraint, and put the environmental parameters into the equation as the solving parameters. Then use the finite difference numerical solution method to calculate the evolution trajectory of thermal runaway in the time and space dimensions, and output the time and space evolution trajectory and impact range prediction result of thermal runaway propagation. Then use the adaptive parameter adjustment algorithm according to the prediction result, dynamically adjust the sparsity parameter of the expert activation mechanism in the sparse expert hybrid acoustic recognition model. When the prediction result shows that the thermal runaway diffusion risk is high, adjust the sparsity parameter to 0.6 to 0.8 to activate more experts to improve the recognition accuracy, and when the acoustic signal is relatively simple, adjust the sparsity parameter to 0.2 to 0.4 to save computing resources. Finally, use the real-time tracking algorithm to continuously monitor the thermal runaway propagation state, and combine the acoustic monitoring result and the dynamics prediction result to realize real-time tracking and early warning of the thermal runaway propagation trend.

[0059] The key technical ideas of the present application mainly include a sparse expert hybrid acoustic recognition model, an acoustic region division and positioning mechanism, and a thermal runaway propagation prediction and adaptive optimization system. Compared with the traditional single neural network recognition method, the sparse expert hybrid acoustic recognition model significantly improves the recognition accuracy and processing efficiency by dividing the complex acoustic recognition task into multiple subtasks and assigning them to different expert networks for processing. At the same time, through the sparse activation mechanism, the computational complexity is greatly reduced, so that the system can realize real-time processing under limited computing resources. Compared with the traditional positioning method based on temperature or voltage monitoring, the acoustic region division and positioning mechanism uses the physical characteristics of acoustic signal propagation to achieve faster and more accurate spatial positioning. By establishing an acoustic fingerprint feature library and an attenuation propagation model, accurate positioning can be achieved in the early stage of thermal runaway, which saves valuable time for subsequent emergency disposal. Compared with the traditional passive monitoring method, the thermal runaway propagation prediction and adaptive optimization system realizes active prediction function through dynamic modeling, which can predict the diffusion path and influence range of thermal runaway in advance, and dynamically adjust the monitoring system parameters according to the prediction results, realizing the organic combination of prediction and monitoring. The synergistic effect of the three key technical ideas forms a complete intelligent monitoring and early warning system, which realizes the whole process coverage of thermal runaway monitoring from early recognition to accurate positioning to diffusion prediction. Compared with the single function module of the existing technology, the present application significantly improves the overall performance and practical value of the system through multi-technology integration, and provides more reliable and intelligent technical support for the safe operation of battery energy storage systems.

[0060] It needs to be further elaborated that the detailed structure of the sparse expert hybrid acoustic recognition model comprises a hybrid architecture of 8 expert networks, each of which adopts a deep convolutional neural network structure, comprising multiple one-dimensional convolution layers, batch normalization layers, activation function layers and pooling layers. The first to third experts respectively adopt a combination of low-pass, band-pass and high-pass filter kernels, and are specially used for processing the safety valve opening event recognition in the low frequency band of 100 to 500 Hz, the medium frequency band of 500 to 1500 Hz and the high frequency band of 1500 to 3000 Hz. Each expert comprises 6 convolution layers, and the convolution kernel sizes are 64, 128 and 256 respectively, and adopts ReLU activation function and maximum pooling operation. The fourth to sixth experts adopt a sound pressure level segmentation processing architecture, and are respectively responsible for the battery exhaust acoustic feature extraction in the high sound pressure level range of 90 to 120 dB, the medium sound pressure level range of 60 to 90 dB and the low sound pressure level range of 30 to 60 dB. Each expert comprises 4 fully connected layers, and the neuron numbers are 512, 256, 128 and 64 respectively, and adopts Dropout regularization technology to prevent overfitting. The seventh expert adopts an autoencoder structure to process background noise suppression, comprising an encoder and a decoder. The encoder compresses the input signal into a low-dimensional feature space, and the decoder reconstructs the pure signal, and extracts the noise component through residual learning. The eighth expert adopts a multi-task learning architecture to be responsible for multi-event concurrent recognition, comprising a shared feature extraction layer and multiple task output layers, and can simultaneously process the concurrent recognition task of multiple acoustic events. The gating network adopts an attention mechanism architecture, analyzes the spectral characteristics, energy distribution and time domain features of the input signal, calculates the activation weight of each expert, and dynamically selects 2 to 3 most relevant experts for collaborative processing.

[0061] The detailed steps of training data set establishment first obtain the original acoustic data through controlled thermal runaway experiment, perform thermal runaway triggering experiment on different types of lithium ion batteries in standard laboratory environment, including overcharge, overheating, mechanical abuse and other triggering methods, use multi-channel high-precision acoustic acquisition equipment to record the complete acoustic signal sequence from the abnormal temperature rise of the battery to the opening of the safety valve to the end of exhaust, the sampling frequency is set to 48 kHz, and the collection time covers the 0 to 30 min time range of the whole thermal runaway process. Then, the original acoustic signal is preprocessed, the wavelet denoising algorithm is used to remove high-frequency noise interference, the signal amplitude is normalized to the range of [-1, 1] through the normalization algorithm, the short-time Fourier transform is used to convert the time-domain signal into time-frequency domain representation, and the mel frequency cepstral coefficient is extracted as the frequency domain feature. Then, the balanced data set is constructed, the positive samples contain the acoustic features of the real thermal runaway events, covering the key events such as safety valve opening and battery exhaust, the negative samples contain various interference signals such as normal operation state, mechanical vibration, environmental noise and personnel activity, and the positive and negative sample ratio is kept at 1:1 to ensure the balance of model training. Then, the data enhancement technology is used to expand the training samples, the signal length is changed through time domain stretching, the frequency domain offset is simulated to simulate different environmental conditions, and the noise superposition is used to enhance the anti-interference ability, and finally a comprehensive training data set containing 50000 samples is formed, covering various battery types such as lithium iron phosphate, ternary lithium and lithium titanate, and various environmental conditions such as normal temperature, high temperature and low temperature.

[0062] It should be noted that the sparse expert hybrid acoustic recognition model is suitable for effectively processing the complexity and diversity characteristics of the battery thermal runaway acoustic signal. The acoustic signal generated in the battery thermal runaway process has the characteristics of multi-band distribution, strong time-varying characteristics, low signal-to-noise ratio, and various event types, and the traditional single neural network model is difficult to process all types of acoustic events at the same time, and the sparse expert hybrid model can use the most suitable network structure and processing algorithm for different types of acoustic events through task decomposition and expert division, which significantly improves the recognition accuracy and processing efficiency. Compared with the existing acoustic recognition technology based on support vector machine, the sparse expert hybrid model of the present application has stronger nonlinear feature extraction capability and better generalization performance, and the support vector machine depends on the manually designed feature engineering, and it is difficult to capture the deep feature relationship in the acoustic signal, while the deep learning architecture can automatically learn the hierarchical feature representation, and better adapt to the complex changes of the acoustic signal. Compared with the existing time series signal processing technology based on recurrent neural network, the hybrid expert architecture of the present application significantly improves the computing efficiency through parallel processing mechanism, and the sequence processing characteristics of the recurrent neural network lead to linear growth of the calculation time with the signal length, while the hybrid expert model can greatly reduce the computational complexity while ensuring the recognition accuracy through expert parallel and sparse activation mechanism, and is more suitable for the needs of real-time monitoring application. The present application realizes the optimal balance of computing efficiency and recognition accuracy through the sparse activation mechanism, dynamically adjusts the number of activated experts according to the complexity of the input signal, reduces the consumption of computing resources in simple signal processing, and increases the expert cooperation to improve the accuracy in complex event recognition. This kind of adaptive adjustment ability is a technical advantage that the traditional fixed architecture model does not have.

[0063] It should be noted that the present application also solves the technical problems that the battery thermal runaway monitoring system in the prior art has the problems of long response time and insufficient early warning capability. The traditional temperature monitoring method needs to rely on the heat conduction process, and there is obvious time lag from the abnormal rise of the battery internal temperature to the detection of temperature change by the sensor, usually tens of seconds or even minutes are needed to trigger the alarm, and the smoke detection technology can only detect the abnormality when a large amount of smoke is generated in the serious stage of thermal runaway, and cannot realize real early warning. The present application can capture abnormal signals and perform identification analysis within milliseconds after the occurrence of thermal runaway by monitoring the opening of the safety valve, the earliest acoustic event of thermal runaway, and using the physical advantage that the sound wave propagation speed is much faster than the heat conduction speed, realizing the ultra-early detection of the thermal runaway event. The safety valve opening event identification expert set in the sparse expert hybrid model can ensure the sensitive capture of the early acoustic event by identifying the burst pulse characteristics of 20 to 50 milliseconds, significantly shortening the response time of the monitoring system, and gaining valuable time window for emergency disposal. In addition, the present application also solves the technical problem that the battery module monitoring system in the prior art has poor anti-interference ability in complex environments. There are a large number of electromagnetic interference sources, mechanical vibration and environmental noise in the energy storage warehouse, and the traditional temperature sensor is easily affected by the electromagnetic field, resulting in a decrease in measurement accuracy, and the smoke detector is prone to false alarms or failure in a strong electromagnetic environment. The present application adopts acoustic monitoring technology which has natural anti-electromagnetic interference advantage, and the acoustic signal is not affected by the electromagnetic field, which can maintain stable detection performance in complex electromagnetic environment. The background noise suppression expert set in the sparse expert hybrid model processes the environmental interference signal, effectively distinguishes the thermal runaway acoustic event from the mechanical vibration, equipment running noise and other interference sources through intelligent filtering and feature extraction technology, significantly improves the reliability and accuracy of the system in harsh environmental conditions, and ensures that the monitoring system can work normally under various complex working conditions.

[0064] Specifically, the principle of the present application is that the core principle of the present application for solving the problems of the prior art lies in fully utilizing the unique physical characteristics and spatial propagation law of acoustic signals in the thermal runaway process. In the thermal runaway process of the battery, the opening of the safety valve will produce a sudden pulse sound with specific frequency domain characteristics in an instant. This acoustic event has the characteristics of short time and high energy, and produces obvious energy steep increase in the frequency band of 500 to 1500 Hz, which provides a reliable acoustic mark for accurate positioning of the thermal runaway source. The present application arranges multiple acoustic sensors in the energy storage bin according to certain rules to form a monitoring array, utilizes the physical characteristics that the speed of sound propagation is constant and isotropic, synchronously measures the sound pressure level data received by the sensors at different positions, and accurately calculates the relative distance relationship between the sound source and each sensor based on the inverse square law of sound energy attenuation with distance. The design logic of the sparse expert hybrid acoustic recognition model lies in decomposing the complex acoustic signal processing task into multiple specialized recognition sub-tasks. The first to third experts respectively process the safety valve opening events in the low, medium and high frequency bands, and ensure accurate identification of key acoustic events through frequency domain feature matching. The fourth to sixth experts process the battery exhaust acoustic characteristics according to different sound pressure level ranges to achieve fine monitoring of the exhaust process. The positioning principle of the cross-channel sound pressure level peak value comparison is based on the physical law that the closer the sound source distance, the greater the sound pressure level. By comparing the instantaneous sound pressure level peak values of each sensor channel in real time, the approximate direction of the sound source can be quickly determined to achieve coarse positioning at the module level. The acoustic region division mechanism describes the propagation and attenuation law of sound waves inside the module through the exhaust sound attenuation propagation equation, considering factors such as geometric diffusion, medium absorption and multipath reflection, and establishes a unique acoustic fingerprint feature for each locatable unit region. By analyzing the attenuation gradient of the sound pressure level in the battery exhaust process and matching with the pre-marked region feature library, the positioning accuracy can be improved to the battery unit level, thereby realizing accurate positioning of the thermal runaway source.

[0065] A specific embodiment 1 of the present application is provided below, and the specific implementation of each step in the embodiment 1 is described in detail as follows.

[0066] The specific implementation of step S01 is to determine the optimal arrangement position of the sensor through the acoustic sensor array geometry optimization algorithm. The spatial coordinate calculation formula of the sensor array arrangement is:

[0067]

[0068] In the formula, x i is the position coordinate of the i-th sensor; L is the module length; M is the total number of sensors, and the value range is 4 to 12; i is the sensor serial number, i = 1, 2, …, M; δ i is the position correction offset. The calculation formula of the sensor coverage radius is:

[0069]

[0070] wherein R c is the sensor coverage radius, ranging from 0.5 to 1.0 m; A total is the total coverage area of the module, in m 2 ; η is the coverage efficiency coefficient, ranging from 0.85 to 0.95. Wherein δ i is obtained by experimental means, including the steps of 1: setting a standard sound source in the module for sound field test; 2: determining the optimal sensor position offset by sound pressure level distribution measurement. A total is directly measured according to the geometric size of the module, and the calculation formula is A total = L x W, wherein W is the width of the module, in m.

[0071] The specific implementation of step S02 is to construct a target acoustic event feature library using a time-frequency domain feature extraction algorithm. The frequency domain energy steep increase calculation formula of the acoustic feature of the safety valve opening is:

[0072]

[0073] wherein E valve (f) is the energy density at frequency f, in J·Hz -1 ; S(f, t) is the time-frequency domain representation of the acoustic signal, in Pa·s 1 / 2 ; t0 is the time of safety valve opening, in s; Δt is the pulse duration, ranging from 20 to 50 ms; f is the frequency, ranging from 500 to 1500 Hz, in Hz. The sound pressure level attenuation model of the battery exhaust acoustic feature is:

[0074] SPL(t) = SPL0·e -αt + β·sin(ωt + φ) + γ.

[0075] wherein SPL(t) is the sound pressure level at time t, in dB; SPL0 is the initial sound pressure level, in dB; α is the attenuation coefficient, in s -1 ; β is the oscillation amplitude, in dB; ω is the oscillation angular frequency, in rad·s -1 ; φ is the phase angle, in rad; γ is the background noise level, in dB; t is the time, in s. Wherein SPL0 is obtained by experimental means, including the steps of 1: triggering the battery exhaust event in a controlled environment; 2: measuring the sound pressure level value at the exhaust start time using a standard sound level meter. The range of α is 0.1 to 0.5 s -1 .

[0076] The specific implementation of step S03 is the same as the foregoing, and will not be described in detail here.

[0077] The specific implementation of step S04 is based on an exhaust sound attenuation propagation equation to establish a mechanism for dividing the acoustic region in the module. The exhaust sound attenuation propagation equation is:

[0078]

[0079] In the formula, SPL(r, θ, φ) is the sound pressure level at the spatial position (r, θ, φ); SPL s is the sound pressure level at the sound source; r is the propagation distance; α a is the medium absorption coefficient; R k is the kth reflection loss; N r is the number of reflections; D j is the diffraction loss of the jth obstacle; N d is the number of obstacles. The number of locatable unit regions for acoustic region division is calculated according to the formula:

[0080]

[0081] In the formula, N is the number of locatable unit regions; P is the initial number of equidistant intervals; ΔSPL th is the sound pressure level resolution threshold, which is 1 to 3 dB. Among them, SPL s is obtained by experimental method, including step 1: setting a standard sound source in the module; step 2: measuring the sound pressure level value at the sound source using a precision sound level meter. α a ranges from 0.001 to 0.01 m -1 .

[0082] The specific implementation of step S05 is to realize the positioning of the thermal runaway module by using a cross-channel sound pressure level peak value comparison algorithm. The multi-channel sound pressure level peak value matrix is:

[0083]

[0084] In the formula, SPL i，max is the peak sound pressure level of the ith sensor channel. The confidence level calculation formula for thermal runaway source positioning is:

[0085]

[0086] In the formula, C loc is the positioning confidence level; SPL max is the maximum sound pressure level; SPL avg is the average sound pressure level; σ SPL is the standard deviation of the sound pressure level; t delay is the signal delay time; τ is the time constant, which is 0.1 to 0.5 s. Among them, SPL i，max is obtained by real-time acquisition, σ SPL is obtained by statistical calculation, and τ ranges from 0.1 to 0.5 s.

[0087] The specific implementation of step S06 is to realize accurate area positioning in the module through sound pressure level attenuation gradient analysis. The calculation formula of the sound pressure level attenuation gradient is:

[0088]

[0089] In the formula, G SPL (t) is the sound pressure level attenuation gradient at time t; SPL(t) is the sound pressure level at time t. The attenuation gradient in discrete form is calculated as:

[0090]

[0091] In the formula, G SPL，discrete (n) is the sound pressure level attenuation gradient of the nth sampling point; T s is the sampling period. The similarity calculation formula of area matching is:

[0092]

[0093] In the formula, S match is the matching similarity; G measured is the measured attenuation gradient; G reference is the reference area attenuation gradient; σ G is the gradient variance. Wherein, T s is set to 0.01-0.1s, σ G is obtained by historical data statistics, ranging from 0.1 to 1.0 dB·s -1 .

[0094] The specific implementation of step S07 is to predict the diffusion path by using the thermal runaway propagation dynamics equation. The thermal runaway propagation dynamics equation is:

[0095]

[0096] In the formula, T(x, y, z, t) is the temperature of the spatial position (x, y, z) at time t, unit K; α T is the thermal diffusion coefficient, unit m 2 ·s -1 ; is the Laplace operator, unit m -2 ; Q gen (x, y, z, t) is the heat generation rate, unit W·m -3 ; ρ is the density, unit kg·m -3 ; c p is the specific heat capacity, unit J·kg -1 ·K -1 ; h c is the convective heat transfer coefficient, unit W·m -2·K -1 ; A s is the surface area, unit m 2 ; V is the volume, unit m 3 ; T amb is the ambient temperature, unit K. The sparsity parameter dynamic adjustment formula is:

[0097]

[0098] wherein, λ sparse (t) is the sparsity parameter at time t; λ base is the basic sparsity, taking a value of 0.2; λ max is the maximum sparsity, taking a value of 0.8; T pred (t) is the predicted temperature; T critical is the critical temperature; T range is the temperature adjustment range. Wherein, α T is obtained through the material database, Q gen (x, y, z, t) is obtained through the electrochemical model calculation, T critical is set to 80-120℃, T range is the temperature adjustment range, taking a value of 20-50K.

[0099] The calculation formula of the acoustic fingerprint feature in the acoustic region division mechanism is:

[0100]

[0101] wherein, F acoustic (d) is the acoustic fingerprint attenuation amplitude at a distance d; d ref is the reference distance, taking a value of 1m. Wherein, d ref is obtained through experimental calibration, including steps 1: setting a point sound source under standard environment; step 2: measuring sound pressure level attenuation data at different distances to determine the reference distance.

[0102] The principle and effect explanation of each formula is as follows. The sensor array arrangement coordinate calculation formula is based on geometric optimization theory, wherein the position correction term δ i is determined by the least square optimization algorithm , wherein SPL measured，i is the measured sound pressure level of the i th sensor, unit dB, SPL predicted，i is the predicted sound pressure level of the i th sensor, unit dB, the spatial continuity of acoustic signal acquisition is ensured by combining uniform distribution and position correction, compared with the traditional equidistant arrangement method, this formula considers the acoustic propagation characteristics and obstacle influence in the actual environment, which can significantly improve the signal coverage quality and positioning accuracy. The frequency energy calculation formula of the safety valve opening acoustic feature is based on the Parseval theorem In the formula, x(t) is a time-domain signal with a unit of Pa, and X(f) is a frequency-domain signal with a unit of Pa·s. The energy density in a certain frequency band is calculated by integration. Compared with the traditional peak detection method, the formula can more accurately capture the burst pulse characteristics and effectively distinguish the safety valve opening event from other transient interference signals.

[0103] The sound pressure level attenuation model of the battery exhaust acoustic characteristics is based on the combination of exponential attenuation theory and oscillation theory, in which the exponential attenuation term SPL0·e -αt The main energy attenuation process is described, and the sinusoidal oscillation term β·sin(ωt+φ) describes the pressure pulsation phenomenon in the exhaust process. Compared with the simple linear attenuation model, the model can more accurately describe the complex time-varying characteristics of the exhaust acoustic signal. The exhaust sound attenuation propagation equation is based on the acoustic propagation theory, in which the geometric diffusion term 20log 10 (r) follows the law of spherical wave propagation, and the medium absorption term α a r is based on the Stokes absorption theory, and comprehensively considers various physical factors such as geometric diffusion, medium absorption, multipath reflection, and diffraction loss. Compared with the traditional inverse square law of distance The equation can provide more accurate sound pressure level prediction in complex closed space environment, and provide reliable theoretical basis for acoustic positioning.

[0104] The confidence calculation formula of the peak value contrast of the sound pressure level across the channels is based on signal processing theory and statistical principles, in which the signal-to-noise ratio term evaluates the signal quality, and the time attenuation factor compensates for the influence of propagation delay, and evaluates the reliability of the positioning result by combining the signal-to-noise ratio and the time attenuation factor. Compared with the simple peak comparison method, the formula can effectively suppress the influence of noise interference and time delay on the positioning accuracy. The sound pressure level attenuation gradient calculation formula is based on the differential theory, and the discrete form uses backward difference approximation In the formula, f(x) is the function value, h is the step size, and the time-domain derivative calculation reflects the dynamic change characteristics of the exhaust process. Compared with the static sound pressure level measurement method, the formula can capture the instantaneous change characteristics of the exhaust acoustic signal, and provide dynamic feature support for accurate area positioning.

[0105] The thermal runaway propagation dynamics equation is based on heat transfer theory and electrochemical reaction dynamics, in which the heat conduction term follows the Fourier heat transfer law, and the heat source term describes the heat generation of electrochemical reaction, and the convection heat dissipation term Based on Newton's cooling law, the evolution process of temperature in space-time dimension is described by partial differential equation, which can predict the diffusion trajectory of thermal runaway from the physical mechanism level compared with the traditional empirical prediction method, and provide scientific basis for active early warning and emergency disposal. The sparsity parameter dynamic adjustment formula is based on the nonlinear mapping characteristics of hyperbolic tangent function In the formula, x is an input variable, and the model complexity is adjusted adaptively through the temperature prediction result. Compared with the identification model with fixed parameters, the formula can dynamically balance the calculation efficiency and identification accuracy according to the actual demand, and realize intelligent resource allocation optimization. The acoustic fingerprint feature segmentation function is based on the distance attenuation law of acoustic propagation, and the piecewise linear function f(x) = a i x+b i , x∈[x i , x i+1 ] describes the attenuation characteristics in different distance ranges, in which a i is the slope coefficient of the ith segment, b i is the intercept coefficient of the ith segment, x i is the starting point of the ith segment, and x i+1 is the end point of the ith segment. Compared with the uniform attenuation model, the function can more accurately describe the different acoustic characteristics of the near field and the far field, and provide reliable feature reference for accurate area division and positioning.

[0106] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: a certain energy storage station contains 2400 lithium iron phosphate battery modules, which are arranged in a matrix of 48 rows and 50 columns in 12 energy storage warehouses, each energy storage warehouse has a length of 60m, a width of 25m and a height of 3.5m. According to the acoustic sensor array arrangement rule of the present application, 8 acoustic sensors are arranged in each energy storage warehouse, the sensor coverage radius is set to 0.8m, and the distance between adjacent sensors is 1.8m, so as to ensure the spatial continuity and redundancy of acoustic signal acquisition in the whole energy storage warehouse.

[0107] The acoustic sensor adopts a wide frequency response capacitive sensor, the frequency response range is 20Hz to 20kHz, the dynamic range is 30dB to 140dB, and the sampling frequency is set to 48kHz. The technical team constructs a target acoustic event feature library according to the method of the present application, and obtains the acoustic feature data of the safety valve opening and the battery exhaust through the thermal runaway experiment in a controlled environment. The acoustic feature of the safety valve opening is a burst pulse of 25ms, and the frequency energy increases obviously in the range of 800Hz to 1200Hz, and the maximum energy density reaches 0.85J·Hz -1 . The acoustic feature of the battery exhaust presents a wide frequency noise distribution in the frequency band of 3000Hz to 4500Hz, the initial sound pressure level is 95dB, and the attenuation coefficient is 0.3s-1 , the oscillation angular frequency is 12.5 rad·s -1 .

[0108] The technical team adopts a sparse expert hybrid acoustic identification model to process the real-time collected acoustic signals, which contains a mixed architecture of 8 expert networks. The 1st to 3rd experts are responsible for identifying the safety valve opening event in the low frequency band 100Hz-500Hz, the medium frequency band 500Hz-1500Hz and the high frequency band 1500Hz-3000Hz respectively, and each expert network contains 6 convolutional layers with 64, 128 and 256 convolutional kernels respectively. The 4th to 6th experts process the battery exhaust acoustic features in high sound pressure level 90dB-120dB, medium sound pressure level 60dB-90dB and low sound pressure level 30dB-60dB respectively, using a 4-layer fully connected layer structure. The 7th expert uses an autoencoder structure to process background noise suppression, and the 8th expert is responsible for multi-event concurrent identification. The gating network dynamically activates 2-3 most relevant experts for collaborative processing according to the spectral characteristics and energy distribution of the input signal.

[0109] According to the acoustic region division mechanism of the application, the technical team divides each energy storage bin into 24 equidistant initial intervals, and then performs acoustic calibration according to the exhaust sound attenuation propagation equation. The input parameters of the exhaust sound attenuation propagation equation include the initial sound pressure level of the sound source 95dB, the medium absorption coefficient 0.005m -1 , the ambient temperature 25℃, and the internal structure parameters of the energy storage bin. After acoustic calibration, 36 locatable unit regions are formed, and the acoustic fingerprint characteristics of each region are shown in Table 1.

[0110] Table 1 Characteristics of acoustic region division in energy storage bin

[0111]

[0112] At 14:32 on a certain day, a battery module thermal runaway event occurred in the 3rd energy storage bin of the energy storage power station. The monitoring system first detected an abnormal temperature rise in the module located in the 23rd row and the 18th column of B area, and the temperature rose rapidly from 32℃ to 78℃. At 14:32:18, the 5th acoustic sensor detected a burst pulse signal lasting 26ms, and the frequency domain analysis showed that the energy increased sharply in the frequency band of 850Hz-1150Hz. The system immediately identified it as a safety valve opening event with a confidence of 0.93.

[0113] The system realizes the rapid positioning of the thermal runaway module through the peak sound pressure level contrast across the channels. The peak sound pressure level data of the 8 acoustic sensor channels is shown in Table 2. The 5th sensor records a maximum sound pressure level of 102.5 dB, which is significantly higher than other sensors. The system maps it to the module at the 23rd row and 18th column in area B, determining it as the thermal runaway source. The positioning confidence calculation result is 0.87, which exceeds the preset threshold of 0.8, confirming the positioning result is valid.

[0114] Table 2 Peak sound pressure level of each sensor during thermal runaway event

[0115]

[0116]

[0117] After locking the thermal runaway module, the system continuously acquires the acoustic signals of the module and extracts the battery exhaust acoustic features. As shown in Figure 4 , the sound pressure level presents a clear exponential decay trend during the exhaust process. The initial sound pressure level is 97.2 dB, and the decay coefficient is 0.28 s -1 . The system calculates the sound pressure level decay gradient according to a sampling period of 0.05 s. When the absolute value of the decay gradient reaches 0.75 dB·s -1 , it exceeds the critical threshold of 0.5 dB·s -1 , triggering the precise area positioning process.

[0118] By matching with the pre-divided area acoustic feature library, the system determines that the thermal runaway occurs in A13 area, with a matching similarity of 0.91. The area positioning result shows that the thermal runaway source is located 1.2 m southeast of the 5th sensor, which completely coincides with the actual position of the module at the 23rd row and 18th column in area B. The entire positioning process takes only 8.5 s from the safety valve opening detection to the completion of the precise area positioning, providing timely and accurate location information for subsequent emergency disposal.

[0119] Based on the thermal runaway propagation dynamics equation, the system predicts the diffusion path of the thermal runaway. The prediction result shows that the thermal runaway will propagate along the module arrangement direction to the adjacent battery, with a diffusion speed of about 0.12 m·min -1 , and is expected to affect 9 modules within a 3x3 range around it within 25 min. According to the prediction result, the system dynamically adjusts the sparsity parameter of the sparse expert hybrid acoustic recognition model from the initial value of 0.4 to 0.75, activating more expert networks to improve the recognition accuracy of the acoustic signals in the diffusion area.

[0120] Based on the predicted diffusion path, the technical team immediately initiates the emergency response procedure, and the hot runaway source module and its surrounding 6 modules are powered off and isolated, and the CO2 fire extinguishing system is started to focus on protecting the 22th-24th row and 17th-19th column area in B area. The monitoring system continuously tracks the propagation process of the thermal runaway, and updates the diffusion range prediction in real time until 14:58, confirming that the thermal runaway is effectively controlled and no further diffusion occurs.

[0121] During the whole process of thermal runaway monitoring and response, the technical solution of the application shows significant technical advantages. The traditional temperature monitoring method usually needs 5 to 10 minutes to detect the thermal runaway signs, while the acoustic monitoring method can detect at the moment when the safety valve is opened, with an advance of several minutes. The traditional positioning method relies on the spatial distribution of temperature sensors, and the positioning accuracy is limited by the sensor density, which can usually only be positioned to the energy storage bin level or module row level, while the acoustic positioning method realizes accurate positioning of individual modules through the characteristics of sound wave propagation and multi-sensor cooperation.

[0122] The progress of the application relative to the traditional method mainly lies in the following aspects. First, the acoustic monitoring principle enables the system to detect thermal runaway at an extremely early stage. The acoustic signal generated by the opening of the safety valve is one of the earliest detectable physical phenomena in the thermal runaway process. Compared with temperature rise, voltage anomaly and other lagging indicators, acoustic signal can provide more timely early warning. Second, the sparse expert hybrid recognition model effectively solves the complex and diverse recognition problem of battery thermal runaway acoustic signal through task decomposition and expert division. Compared with traditional single recognition algorithm, the hybrid expert architecture can handle different types of acoustic events at the same time, significantly improving recognition accuracy and robustness. Third, the establishment of the acoustic propagation physical model enables the system to have accurate spatial positioning capability. Through the propagation and attenuation law of sound waves in the module, it can realize meter-level or sub-meter-level positioning accuracy, far exceeding the positioning capability of traditional methods. Finally, the thermal runaway propagation dynamics prediction function realizes the technical leap from passive monitoring to active early warning. Through physical modeling to predict the spatio-temporal evolution process of thermal runaway, it provides a scientific basis for emergency disposal. Compared with the traditional experience judgment method, the dynamics prediction has higher accuracy and reliability.

[0123] It should be noted that the variables involved in the application are explained in detail as shown in Table 3.

[0124] Table 3 Variable Explanation Table

[0125]

[0126]

[0127] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A method for battery module thermal runaway monitoring and positioning pre-warning, characterized in that, The application relates to a battery module thermal runaway monitoring and positioning method based on acoustic signal processing.

2. The battery module thermal runaway monitoring and locating pre-alarming method of claim 1, wherein, The acoustic sensor array arrangement rule is as follows: according to the module dense arrangement characteristics in the energy storage warehouse, the coverage radius of each acoustic sensor is 0.5-1.0 m, the distance between adjacent sensors is not more than 2.0 m, and the spatial continuity and redundancy of acoustic signal acquisition are ensured.

3. The battery module thermal runaway monitoring and locating pre-alarming method of claim 2, wherein, The safety valve opening acoustic feature is as follows: when the pressure in the battery exceeds the safety threshold, the safety valve is instantaneously opened to generate a sound, which is characterized by a 20-50 ms burst pulse and an energy abrupt increase in the 500-1500 Hz frequency domain, and has the characteristics of short pulse and high frequency.

4. The battery module thermal runaway monitoring and locating pre-alarming method of claim 3, wherein, The battery exhaust acoustic feature is as follows: after the safety valve is opened, a wideband noise signal is generated during the continuous exhaust of the battery internal gas, which is characterized by a 2000-5000 Hz frequency band wideband noise and a continuous attenuation of the sound pressure level.

5. The battery module thermal runaway monitoring and locating pre-alarming method of claim 4, wherein, The sparse expert hybrid acoustic recognition model is a hybrid architecture containing 8 expert networks, each of which processes different frequency domain ranges or acoustic event types, and the expert selection mechanism dynamically activates 2-3 most relevant experts for collaborative processing through a gating network according to the input signal spectrum characteristics and energy distribution.

6. The battery module thermal runaway monitoring and locating pre-alarming method of claim 5, wherein, The acoustic region division mechanism is as follows: according to the module length L and the positions of the M acoustic sensors, a spatial coordinate system is established, the module is divided into P equidistant initial intervals, and then N locatable unit regions are formed through acoustic calibration according to the exhaust sound attenuation propagation equation.

7. The battery module thermal runaway monitoring and locating pre-alarming method of claim 6, wherein, The exhaust sound attenuation propagation equation is used to describe the energy attenuation law of the battery exhaust sound during the propagation process in the module, and comprehensively considers the influence of geometric diffusion, medium absorption, multipath reflection and obstacle shielding of sound waves in a closed space on the sound pressure level.

8. The battery module thermal runaway monitoring and locating pre-alarming method of claim 7, wherein, The cross-channel sound pressure level peak comparison is as follows: the instantaneous sound pressure level data of multiple acoustic sensor channels are synchronously acquired, the sound source direction is determined by comparing the peak values of the channels in real time, and the physical law of sound energy attenuation with distance is used to realize rapid coarse positioning.

9. The battery module thermal runaway monitoring and locating pre-alarming method of claim 8, wherein, The sound pressure level attenuation gradient is as follows: the sound pressure level change rate of the battery exhaust acoustic feature signal in the time dimension, which is obtained by calculating the sound pressure level difference between continuous sampling points and dividing by the time interval, and reflects the dynamic characteristics and sound source intensity change trend of the exhaust process.

10. The battery module thermal runaway monitoring and localization pre-alarming method of claim 9, wherein, The thermal runaway propagation dynamics equation, in particular for predicting the diffusion process of thermal runaway in a battery module, is established based on heat conduction theory and electrochemical reaction kinetics, and comprehensively considers the thermal coupling effect between batteries, heat dissipation conditions and material thermal physical parameters.

Images