Lithium battery thermal runaway fire identification method and system

Abstract

The invention discloses a lithium battery thermal runaway fire identification method and system, and relates to the technical field of lithium battery monitoring. The method comprises the following steps: acquiring historical acoustic signals of a plurality of lithium batteries in different states in real time through an acoustic sensor; preprocessing and converting each historical acoustic signal to obtain a normalized logarithmic Mel spectrogram; all the normalized logarithmic Mel spectrograms and the state of the lithium battery form a data set, and a MobileViT model is trained; the method comprises the following steps: acquiring a real-time acoustic signal of a lithium battery to be detected through an acoustic sensor, and preprocessing and converting the real-time acoustic signal to obtain a normalized logarithmic Mel spectrogram of the real-time acoustic signal; inputting the logarithmic Mel spectrogram normalized by the real-time acoustic signal into the trained MobileViT model to obtain a thermal runaway fire identification result of the lithium battery to be detected; the thermal runaway fire identification result is fire or non-fire. According to the invention, the lithium battery thermal runaway fire can be rapidly and accurately identified.

Description

Technical Field

[0001] The present application relates to the technical field of lithium battery monitoring, and in particular to a method and system for identifying thermal runaway fires of lithium batteries. Background Art

[0002] Lithium batteries have been widely used in transportation, energy storage systems, and consumer goods, facilitating the rapid promotion of electrification throughout society. However, when lithium batteries are subjected to thermal, electrical, mechanical, or electrochemical abuse, they may trigger thermal runaway and cause fires, seriously threatening the safety of people and property. Therefore, timely and accurate identification of lithium battery fires is of great significance for quickly adopting prevention and control strategies, preventing accidents from expanding, and ensuring public safety.

[0003] At present, the identification methods for lithium battery fires mainly focus on the following traditional means:

[0004] 1. Sensor-based fire warning method

[0005] Obtaining external battery parameters: Using the battery management system (BMS) to monitor the external parameters of the battery in real time, such as voltage, current, surface temperature, pressure, etc., as early warning conditions for battery thermal runaway; estimating the internal temperature of the battery: estimating the internal temperature of the battery based on the relevant external parameters of the battery for early warning; detecting characteristic gases: the battery will produce characteristic gases (such as CO, CO2, H2, etc.) during thermal runaway, and early warning of thermal runaway is performed by detecting these gases. In high temperature or high-speed rotation environments, or in environments where sensor deployment is challenging, acoustic sensors have the characteristics of non-contact and flexible layout, and acoustic sensing technology only needs to deploy acoustic probes on the outside of the battery to obtain information such as internal structural changes and gas production, which is an ideal non-destructive monitoring method.

[0006] 2. Traditional methods of post-accident fire cause analysis

[0007] Physical and chemical analysis: Sampling the residues at the fire scene, and using chemical analysis methods (such as chromatography, mass spectrometry, etc.) to detect the characteristic substances of lithium battery materials; Fire trace analysis: Through the investigation of the fire scene, analysis of burning traces, residue morphology, etc., to infer the cause of the fire; Witness and surveillance records: Collect witness testimony and surveillance videos to assist in determining whether the fire was caused by a lithium battery.

[0008] The above traditional methods for lithium battery fire identification have the following shortcomings:

[0009] (1) Lack of accuracy

[0010] Complex physical and chemical analysis: Professional laboratory equipment and technicians are required, and the process is complicated and time-consuming.

[0011] Fire traces are vulnerable to destruction: the fire extinguishing process may destroy key evidence and affect the accuracy of analysis results.

[0012] Eyewitness information is unreliable: Eyewitness accounts may be biased, and surveillance equipment may be damaged or have unclear records.

[0013] (2) Low efficiency

[0014] Long analysis cycle: Traditional methods require a long analysis time and cannot quickly draw conclusions, delaying accident handling.

[0015] High resource requirements: A large amount of manpower and material resources are required, which increases the cost of accident handling.

[0016] (3) Strong professionalism and poor universality

[0017] High technical threshold: Professional knowledge is required, and it is difficult for ordinary investigators to complete it independently.

[0018] Cannot be widely used: Due to limitations on equipment and personnel, it is difficult to be widely used in various fire accidents.

[0019] (4) Susceptible to subjective factors

[0020] Human judgment error: The analysis results may be affected by subjective factors, resulting in inaccurate conclusions.

[0021] Therefore, it is necessary to provide an efficient, accurate and objective identification method for fire accident investigation, and to provide a new technical route for the automation of lithium battery fire identification after a fire occurs. Summary of the invention

[0022] The purpose of this application is to provide a method and system for identifying a lithium battery thermal runaway fire, which can quickly and accurately identify a lithium battery thermal runaway fire.

[0023] To achieve the above objectives, this application provides the following solutions:

[0024] In a first aspect, the present application provides a method for identifying a thermal runaway fire of a lithium battery, comprising: collecting historical acoustic signals of multiple lithium batteries in different states in real time through an acoustic sensor; the different states include: fire and non-fire; the fire includes explosion and non-explosion; preprocessing each historical acoustic signal to obtain each preprocessed historical acoustic signal; the preprocessing includes: removing environmental noise in the historical acoustic signal using an adaptive chirp pattern decomposition method; converting each preprocessed historical acoustic signal into a normalized log-Mel spectrum graph; forming a data set with all normalized log-Mel spectrum graphs and the state of the lithium battery, and training a MobileViT model to obtain a trained MobileViT model; collecting real-time acoustic signals of a lithium battery to be tested through an acoustic sensor, and preprocessing and converting the real-time acoustic signals to obtain a normalized log-Mel spectrum graph of the real-time acoustic signals; inputting the normalized log-Mel spectrum graph of the real-time acoustic signals into the trained MobileViT model to obtain a thermal runaway fire identification result of the lithium battery to be tested; the thermal runaway fire identification result is fire or non-fire.

[0025] In a second aspect, the present application provides a lithium battery thermal runaway fire identification system, comprising: an acoustic sensor and a processing device; the acoustic sensor is connected to the processing device; the acoustic sensor is used to collect historical acoustic signals of multiple lithium batteries in different states in real time; the different states include: fire and non-fire; the fire includes explosion and non-explosion; the processing device is used to preprocess each historical acoustic signal to obtain each preprocessed historical acoustic signal; the preprocessing includes: using an adaptive chirp mode decomposition method to remove environmental noise in the historical acoustic signal; converting each preprocessed historical acoustic signal into a normalized logarithmic plum Mel spectrum diagram; all normalized logarithmic Mel spectrum diagrams and the status of the lithium battery are combined into a data set, and the MobileViT model is trained to obtain a trained MobileViT model; the acoustic sensor is also used to collect the real-time acoustic signal of the lithium battery to be tested; the processing device is also used to preprocess and convert the real-time acoustic signal to obtain a normalized logarithmic Mel spectrum diagram of the real-time acoustic signal; the normalized logarithmic Mel spectrum diagram of the real-time acoustic signal is input into the trained MobileViT model to obtain a thermal runaway fire recognition result of the lithium battery to be tested; the thermal runaway fire recognition result is fire or non-fire.

[0026] According to the specific embodiments provided in this application, this application has the following technical effects:

[0027] The present application provides a method and system for identifying thermal runaway fires of lithium batteries. The acoustic sensor has the characteristics of non-contact and flexible arrangement, avoiding direct contact in high temperature and dangerous environments, and can obtain the acoustic signals of lithium batteries in real time during the occurrence of a fire, providing basic data for accurately identifying the fire; an adaptive chirp pattern decomposition method is used to effectively remove environmental noise, enhance the quality of the acoustic characteristic signals of lithium battery fires, and provide a reliable data basis for subsequent feature extraction; by converting each pre-processed historical acoustic signal into a normalized logarithmic Mel spectrum graph, specific voiceprint features in the audio data of the fire scene can be extracted and analyzed, and the powerful feature extraction and classification capabilities of the MobileViT model are used to improve the accuracy of lithium battery fire identification. At the same time, the MobileViT model has a simple structure and a small amount of calculation, and can quickly process specific voiceprint features to achieve rapid identification of fire types. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 A schematic diagram of a flow chart of a method for identifying a thermal runaway fire of a lithium battery provided in one embodiment of the present application;

[0030] Figure 2 A schematic diagram of the principle of a method for identifying a thermal runaway fire of a lithium battery provided in one embodiment of the present application;

[0031] Figure 3 A schematic diagram of the structure of a processing device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0033] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0034] In the process of investigating lithium battery thermal runaway fire accidents, accident investigators usually rely on video footage at the scene to trace the process of the fire. However, the reality is often unsatisfactory. Due to the complexity and uncertainty of the fire scene, in many cases, video footage may be lost or damaged due to high temperature, smoke or damage, resulting in investigators only being able to obtain audio footage. In this case, it is difficult to accurately determine the cause of the fire based on audio information alone, especially when the fire may be related to different triggers that trigger lithium battery thermal runaway. The process of lithium battery triggering thermal runaway has unique voiceprint features, which may be included in the audio footage at the fire scene. However, current accident investigators are unable to effectively extract and analyze these voiceprint features from audio footage, which limits the ability of accident investigators to quickly identify lithium battery fires. Therefore, developing a technology that can accurately analyze the voiceprint features of lithium battery thermal runaway based on audio footage is crucial to improving the efficiency and accuracy of fire accident investigations.

[0035] In view of this, in an exemplary embodiment, Figure 1 As shown, a method for identifying thermal runaway fire of a lithium battery is provided, comprising the following steps 101 to 106. Among them:

[0036] Step 101: using an acoustic sensor to collect historical acoustic signals of multiple lithium batteries in different states in real time; the different states include: fire and non-fire; the fire includes explosion and non-explosion.

[0037] Step 102: preprocess each historical acoustic signal to obtain each preprocessed historical acoustic signal; the preprocessing includes: using an adaptive chirp mode decomposition method to remove environmental noise in the historical acoustic signal.

[0038] Step 103: Convert each preprocessed historical acoustic signal into a normalized logarithmic Mel-spectrogram.

[0039] Step 104: All normalized logarithmic Mel-spectrograms and the state of the lithium battery are combined to form a data set, and the MobileViT model is trained to obtain a trained MobileViT model.

[0040] Step 105: collect the real-time acoustic signal of the lithium battery to be tested through the acoustic sensor, and pre-process and convert the real-time acoustic signal to obtain a normalized logarithmic Mel spectrum diagram of the real-time acoustic signal.

[0041] Step 106: input the normalized logarithmic Mel-spectrogram of the real-time acoustic signal into the trained MobileViT model to obtain a thermal runaway fire identification result of the lithium battery to be tested; the thermal runaway fire identification result is fire or non-fire.

[0042] By implementing the above steps 101 to 106, the specific voiceprint features in the audio data of the fire scene are extracted and analyzed to achieve rapid and accurate identification of lithium battery thermal runaway fires. This not only helps accident investigators to quickly identify the cause after the fire occurs and provide a scientific basis for accident handling, but also provides strong support for fire prevention and safety management in places involving lithium batteries.

[0043] The method of this application aims to provide an efficient, accurate and objective identification method for fire accident investigation, and provides a new technical route for the automation of lithium battery fire identification after a fire occurs. In high temperature or high-speed rotation environments, or in environments where sensor deployment is challenging, acoustic sensors have the characteristics of non-contact and flexible arrangement. They can use the unique time-frequency characteristics of fire sound signals to improve the accuracy and objectivity of identification, and use the powerful feature extraction and classification capabilities of the MobileViT network to improve the accuracy of lithium battery fire identification. The MobileViT network has a simple structure and low computational complexity. It can quickly process acoustic signal data and realize rapid identification of fire types. Automated identification reduces dependence on professional knowledge and equipment, reduces human subjective judgment, and improves the objectivity and fairness of accident investigations.

[0044] In another exemplary embodiment of the present application, acoustic sensors are deployed around lithium batteries to collect real-time acoustic signals of lithium batteries during normal operation and fire. The acoustic sensor uses a high-sensitivity microphone to obtain acoustic signals in a non-contact manner, avoiding direct contact in high temperature and dangerous environments.

[0045] The sound sensor is used to collect the sound signals of historical explosions and unexplosions in lithium battery fire incidents. These signals contain the acoustic characteristic information of lithium battery fires in different states and are the basic data for subsequent fire identification. By collecting a large amount of explosion and unexplosion sample data, a complete data set covering different states can be established.

[0046] In order to accurately collect the acoustic signals of lithium batteries in normal operation and fire conditions, a highly sensitive sound sensor condenser microphone and a high-precision data acquisition card are selected to support high sampling rate and multi-channel synchronous acquisition, and the data is transmitted to a computer or server for storage and processing through a high-speed interface (such as USB 3.0 or Ethernet). The data block is collected every 100ms, and the lossless audio format WAV is used to ensure data quality.

[0047] In another exemplary embodiment of the present application, the collected historical acoustic signal (original acoustic signal) is preprocessed, including filtering, noise reduction, signal enhancement, etc. Then the above step 102 can be replaced by the following steps 201 to 203:

[0048] Step 201: Use a bandpass filter to filter out high-frequency noise and low-frequency noise in each historical acoustic signal to obtain a historical acoustic signal with primary noise reduction.

[0049] Step 202: Adopting an adaptive chirp pattern decomposition method to remove the environmental noise in the historical acoustic signal of primary noise reduction, and obtaining the historical acoustic signal of secondary noise reduction.

[0050] Step 203: performing amplitude normalization processing on the secondary denoised historical acoustic signals to obtain each pre-processed historical acoustic signal.

[0051] Amplitude normalization processing eliminates the amplitude differences under different acquisition conditions, enhances the quality of the acoustic characteristic signals of lithium battery fires, and provides a reliable data basis for subsequent feature extraction.

[0052] Exemplarily, the specific implementation process of the above step 202 may be: using an adaptive chirp pattern decomposition method to decompose the historical acoustic signal of the primary noise reduction into multiple chirp pattern functions; calculating the correlation coefficient between each chirp pattern function and the historical acoustic signal of the primary noise reduction; retaining the chirp pattern functions whose correlation coefficient is greater than a preset coefficient threshold; and superimposing the retained chirp pattern functions to obtain the historical acoustic signal of the secondary noise reduction.

[0053] When the preset coefficient threshold is 0.3, the more detailed implementation process of the above steps 202 and 203 is as follows:

[0054] The original sound signal is decomposed into multiple chirped mode functions (CMF) using the Adaptive Chirp Mode Decomposition (ACMD) method, which represent different components in the signal. By calculating the correlation coefficient of each CMF, only the CMF with a correlation coefficient greater than 0.3 is retained to reconstruct the signal, filter out the noise component, and enhance the features related to lithium battery fire. The reconstructed signal can more accurately reflect the explosion of the fire, which is helpful for subsequent feature extraction.

[0055] By applying the ACMD algorithm to the original acoustic signal, several CMFs are obtained. The correlation coefficient between each CMF and the original signal is calculated using the formula:

[0056]

[0057] in:

[0058] x i : The value of variable X in the i-th observation sample, N: The number of samples, that is, the number of observation pairs, indicating that there are N pairs (x i , c i )data, The sample mean of variable X is defined as c i : The value of variable C in the i-th observation sample, The sample mean of variable C is defined as

[0059] CMFs with correlation coefficients greater than 0.3 are retained, and they are considered to contain the main signal features. The selected CMFs are superimposed to obtain the reconstructed signal. The reconstructed signal effectively filters out the noise and retains the acoustic features related to lithium battery fires. The reconstructed signal is normalized by Z-score, and the formula is:

[0060]

[0061] in:

[0062] x(t): the original value of the reconstructed signal at time point t (the value of the tth sample point in the signal sequence after CMF screening and superposition);

[0063] μ: the mean (average value) of the reconstructed signal, that is, the average value calculated over the entire reconstructed signal (or a specific analysis interval);

[0064] σ: standard deviation of the reconstructed signal, which describes the dispersion of the signal amplitude distribution by measuring the dispersion of the signal from its mean;

[0065] x norm (t): The result of Z-score normalization of the reconstructed signal at time point t. The normalized signal has a mean of 0 and a standard deviation of 1.

[0066] In another exemplary embodiment of the present application, the preprocessed acoustic signal is converted into a Mel spectrum graph to extract the key acoustic features of lithium battery fires in order to better train and identify the model. First, a short-time Fourier Transform (STFT) is performed to convert the preprocessed time domain signal into a time-frequency domain representation to obtain the distribution characteristics of the signal on the spectrum. Next, the Mel filter group is applied to process the spectrum obtained by STFT to make it consistent with the auditory characteristics of the human ear and obtain a Mel spectrum graph. Subsequently, the logarithmic energy of the Mel spectrum is calculated to generate a logarithmic Mel spectrum graph, thereby enhancing the frequency components to which the human ear is more sensitive. Finally, feature normalization is performed to reduce the differences between samples and improve the training stability of the model.

[0067] Mel spectrogram is a spectrum analysis tool based on the human hearing perception mechanism, which can effectively capture the key features in the sound signal. Since the sound signals of lithium battery fire explosions often have unique frequency distribution characteristics, Mel spectrogram provides a detailed description of acoustic features for subsequent fire identification.

[0068] The time domain signal is converted into the time-frequency domain through short-time Fourier transform to analyze the change of the signal frequency over time. The window function is set to Hamming window, the window length is 1024 points, about 23ms (calculated at 44.1kHz sampling rate), the window overlap is 50%, that is, 512 points are moved each time, and the formula is:

[0069]

[0070] in:

[0071] n: time frame index, indicating the starting point of the nth frame signal.

[0072] m: sampling point index within the window, indicating the mth sampling point starting from the starting point in the current frame.

[0073] k: frequency index, indicating the kth frequency component in the frequency domain.

[0074] M: Window length, indicating the length of the Hamming window.

[0075] X(n, k): The short-time Fourier transform result of the signal at the nth frame and the kth frequency, which is a time-frequency representation in complex form.

[0076] x(n+m): The value of the time domain signal at the mth sampling point in frame n.

[0077] w(m): The weight of the window function value at the mth point.

[0078] This formula describes the basic process of short-time Fourier transform, which converts the signal from the time domain to the time-frequency domain by applying a sliding window (with a window function) to the time domain signal and performing Fourier transform, which is used to analyze the change of the signal's frequency over time.

[0079] Then, the Mel scale is used to simulate the human ear's perception of frequency. 40 filters are selected to cover the frequency range of 20Hz to 20kHz. Triangular filters are used, and each filter is evenly distributed on the Mel scale. The spectrum energy obtained by STFT is weighted and summed through the filter bank to obtain the energy of each Mel band. The energy of the Mel spectrum is converted to a logarithmic scale to enhance the frequency components that the human ear is sensitive to. The formula is:

[0080] LogMek(k)=log 10 (E(k)+∈).

[0081] in:

[0082] E(k): The energy of the k-th Mel band, which represents the band energy on the k-th filter after weighted summation of the STFT results through the Mel filter bank.

[0083] ∈: A small positive value, usually used to prevent the log(0) problem in logarithmic operations and ensure numerical stability.

[0084] LogMel(k): Logarithmic energy of the kth Mel band, which represents the result of converting E(k) to a logarithmic scale to better match the human ear's perception of sound intensity.

[0085] The logarithmic energy of each Mel band is normalized by mean and standard deviation, and the formula is:

[0086]

[0087] in:

[0088] LogMel(k): Logarithmic energy of the kth Mel band, which represents the value after logarithmic conversion of the energy of the Mel spectrum.

[0089] μ k : The mean (average) of the logarithmic energy of the kth Mel band, used to indicate the central tendency of the logarithmic energy of the band.

[0090] σ k : The standard deviation of the logarithmic energy of the kth Mel band, which is used to measure the dispersion of the logarithmic energy value of the band.

[0091] LogMel norm (k): The normalized value of the logarithmic energy of the kth Mel band, which represents the result after normalization by the mean and standard deviation. After normalization, the logarithmic energy data of each Mel band will have the characteristics of a mean of 0 and a standard deviation of 1.

[0092] Finally, the Mel spectrum diagram is generated intuitively. Figure 2 The figure shows the Mel spectrum of the acoustic signal of a lithium battery fire. The horizontal axis is time, the vertical axis is Mel frequency, and the color represents energy intensity, which can intuitively reflect the spectrum characteristics of the fire sound signal.

[0093] In summary, the above step 103 can be replaced by the following steps 301 to 304:

[0094] Step 301: Perform short-time Fourier transform on each preprocessed historical acoustic signal to obtain a frequency spectrum of each historical acoustic signal.

[0095] Step 302: Perform weighted summation on the spectrogram of each historical acoustic signal through a Mel filter bank to obtain the energy of each Mel frequency band.

[0096] Step 303: Convert the energy of each Mel frequency band into a logarithmic scale to obtain the logarithmic energy of each Mel frequency band.

[0097] Step 304: normalize the logarithmic energy of each Mel frequency band by the mean and standard deviation to generate a normalized logarithmic Mel frequency spectrogram of each historical acoustic signal.

[0098] In another exemplary embodiment of the present application, the MobileViT model includes: a convolutional layer, a visual Transformer module and an output layer connected in sequence; the convolutional layer includes a convolutional neural network, and the convolutional neural network uses depthwise separable convolution; the convolutional neural network is used to extract local features of the normalized log-Mel spectrum map; the visual Transformer module is used to model the global features of the normalized log-Mel spectrum map through a multi-head self-attention mechanism based on the local features; the output layer is used to predict the probability distribution of fire and non-fire based on the global features of the normalized log-Mel spectrum map, and output the thermal runaway fire recognition result of the lithium battery.

[0099] The MobileViT lightweight neural network model is constructed. The model design includes three parts: First, the convolutional layer is designed through a shallow convolutional neural network (CNN) to extract the local features of the Mel spectrum graph, and the deep separable convolution is used to reduce the model parameters and computational complexity. Then, the visual transformer (ViT) module is added to model the global features of the Mel spectrum graph and capture the long-distance dependencies between features, thereby improving the expressiveness of the model. Finally, the model is lightweighted through pruning and quantization techniques to reduce the number of parameters and computation, making it suitable for deployment on resource-constrained devices.

[0100] The MobileViT model first extracts local features of lithium battery fire acoustic signals through convolutional layers. The convolutional layers use depthwise separable convolutions, including deep convolutions and point-by-point convolutions, to reduce the number of parameters and computations. Subsequently, the Visual Transformer (ViT) module models global features through a multi-head self-attention mechanism to capture long-distance dependencies between features. The model output layer includes a fully connected layer and a Softmax layer to predict the probability distribution of fire and non-fire.

[0101] In another exemplary embodiment of the present application, during the model training phase, an acoustic data set containing a large number of lithium battery fires and normal states is constructed, converted into a Mel spectrogram, and divided into a training set, a validation set, and a test set. Set appropriate loss functions, optimization algorithms, and other hyperparameters, train on the training set, evaluate the model performance on the validation set, and adjust the hyperparameters to prevent overfitting. Finally, use the test set to evaluate the model to ensure that its accuracy, recall, and F1 score meet the application requirements.

[0102] The MobileViT model combines the advantages of convolutional neural networks and visual transformers (ViT) to effectively process image data. Through training, the model learns the acoustic feature distribution of lithium battery explosions, laying the foundation for subsequent lithium battery fire identification. The model sets a probability threshold for fire determination: 0.8. When the probability exceeds this threshold, the system will determine that it is a lithium battery fire. The lightweight design of MobileViT also makes the model suitable for deployment in resource-limited environments.

[0103] The training set, validation set, and test set are divided into training set, validation set, and test set in a ratio of 7:2:1. The Adam optimizer is used with an adaptive learning rate. The loss function is the cross-entropy loss function, which is suitable for binary classification problems. The initial learning rate is set to 0.001, and the learning rate can be decayed during training. The batch size is 32. The number of training epochs is 100. The model training uses the cross-entropy loss function as the optimization target, and the formula is as follows:

[0104]

[0105] in:

[0106] Loss: The value of the cross entropy loss function, which is used to measure the difference between the model prediction and the true label, is the objective function to be optimized.

[0107] N: The number of samples, that is, the number of samples in the batch.

[0108] y i : The true label of the i-th sample, which takes the value of 0 (non-fire) or 1 (fire).

[0109] The probability of the fire category predicted by the model for the i-th sample (values ​​range from [0, 1]).

[0110] Fire category probability The natural logarithm of is used to calculate the loss when a fire is predicted.

[0111] Non-fire category probability The natural logarithm of is used to calculate the loss when the prediction is not a fire.

[0112] The cross entropy loss function is suitable for binary classification problems and optimizes the accuracy of model prediction by penalizing the wrong prediction probability.

[0113] The training process also uses regularization, dropout, and early stopping to prevent overfitting and ensure the generalization ability of the model. The model is finally evaluated on the test set and the accuracy is calculated: Measuring the overall accuracy and precision of the model: Measuring the accuracy and recall of the model for positive class predictions: Measures the model's ability to identify positive samples, F1 score (F1Score): The model's ability to identify positive samples is measured to measure the overall performance of the model in fire identification tasks.

[0114] In another exemplary embodiment of the present application, a trained MobileViT model is used to identify lithium battery fires from acoustic signals collected in real time. The identification process is as follows: first, a real-time preprocessed acoustic signal is obtained, and the signal is converted into a logarithmic Mel spectrum graph, and normalized at the same time. Subsequently, the processed Mel spectrum graph is input into the MobileViT model for forward calculation to generate a fire identification result. For the identification result, a probability threshold for fire determination is set: 0.8. If the fire probability P output by the model is ≥ 0.8, it is determined to be a lithium battery fire. If P < 0.8, it is determined to be a non-lithium battery fire event. Finally, the identification results, timestamps, signal characteristics and other information are stored in the database for subsequent analysis.

[0115] Based on the same inventive concept, the embodiment of the present application also provides a lithium battery thermal runaway fire identification system for implementing the above-mentioned lithium battery thermal runaway fire identification method. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more lithium battery thermal runaway fire identification system embodiments provided below can refer to the limitations of the lithium battery thermal runaway fire identification method above, and will not be repeated here.

[0116] In an exemplary embodiment, a lithium battery thermal runaway fire identification system is provided, including: an acoustic sensor and a processing device. The acoustic sensor is connected to the processing device; the acoustic sensor is used to collect historical acoustic signals of multiple lithium batteries in different states in real time; the different states include: fire and non-fire; the fire includes explosion and non-explosion. The processing device is used to preprocess each historical acoustic signal to obtain each preprocessed historical acoustic signal; the preprocessing includes: using an adaptive chirp mode decomposition method to remove environmental noise in the historical acoustic signal; converting each preprocessed historical acoustic signal into a normalized logarithmic Mel spectrum graph; all normalized logarithmic Mel spectrum graphs together with the state of the lithium battery constitute a data set, and train the MobileViT model to obtain a trained MobileViT model.

[0117] The acoustic sensor is also used to collect the real-time acoustic signal of the lithium battery to be tested. The processing device is also used to preprocess and convert the real-time acoustic signal to obtain a normalized logarithmic Mel spectrum of the real-time acoustic signal; the normalized logarithmic Mel spectrum of the real-time acoustic signal is input into the trained MobileViT model to obtain a thermal runaway fire recognition result of the lithium battery to be tested; the thermal runaway fire recognition result is fire or non-fire.

[0118] As an optional implementation, the lithium battery thermal runaway fire identification system further includes: a data acquisition card and an interface. The data acquisition card is connected to the acoustic sensor and the interface respectively; the interface is connected to the processing device. The data acquisition card is used to collect the acoustic signal measured by the acoustic sensor and transmit it to the processing device through the interface.

[0119] The acoustic sensor is a high-sensitivity sound sensor condenser microphone, and the data acquisition card is a high-precision data acquisition card that supports high sampling rate and multi-channel synchronous acquisition. The interface is a high-speed interface, such as USB 3.0 or Ethernet. The processing device can be a computer or a server.

[0120] like Figure 3 As shown, the processing device mainly includes the following modules: data acquisition and preprocessing module, Mel spectrum feature extraction module, MobileViT model construction and training module and lithium battery fire identification module.

[0121] The data acquisition and preprocessing module includes an acoustic signal acquisition unit and a signal preprocessing unit. The acoustic signal acquisition unit collects the acoustic signals of lithium batteries in normal operation and during fire in real time through acoustic sensors deployed around the lithium batteries. The acoustic sensor uses a high-sensitivity microphone to obtain acoustic signals in a non-contact manner, avoiding direct contact in high temperature and dangerous environments. The signal preprocessing unit preprocesses the collected original acoustic signals, including filtering, noise reduction, signal enhancement, etc. Specifically: a bandpass filter is used to filter out high-frequency and low-frequency noise in the signal to improve the signal-to-noise ratio of the signal; an adaptive chirp mode decomposition method is used to effectively remove environmental noise, and then the amplitude of the noise-reduced signal is normalized to eliminate the amplitude difference under different acquisition conditions, enhance the quality of the acoustic characteristic signal of the lithium battery fire, and provide a reliable data basis for subsequent feature extraction.

[0122] The Mel spectrum feature extraction module converts the preprocessed acoustic signal into a Mel spectrum graph and extracts the key acoustic features of lithium battery fire in order to better train and identify the model. First, a short-time Fourier transform is performed to convert the preprocessed time domain signal into a time-frequency domain representation to obtain the distribution characteristics of the signal on the spectrum. Next, the Mel filter group is applied to process the spectrum obtained by STFT to make it consistent with the auditory characteristics of the human ear and obtain a Mel spectrum graph. Subsequently, the logarithmic energy of the Mel spectrum is calculated to generate a logarithmic Mel spectrum graph, thereby enhancing the frequency components to which the human ear is more sensitive. Finally, feature normalization is performed to reduce the differences between samples and improve the training stability of the model.

[0123] The MobileViT model building and training module is responsible for building the MobileViT lightweight neural network model and training the model using Mel spectrum features. The model design includes three parts: First, the convolution layer is designed through a shallow convolutional neural network to extract local features of the Mel spectrum graph, and deep separable convolution is used to reduce model parameters and computational complexity. Then, the visual transformer (ViT) module is added to model the global features of the Mel spectrum graph and capture the long-distance dependencies between features, thereby improving the expressiveness of the model. Finally, the model is lightweighted through pruning and quantization techniques to reduce the number of parameters and computational complexity, making it suitable for deployment on resource-constrained devices. In the model training phase, an acoustic dataset containing a large number of lithium battery fires and normal states is constructed, converted into Mel spectrum graphs, and divided into training sets, validation sets, and test sets. Appropriate loss functions, optimization algorithms, and other hyperparameters are set, and training is performed on the training set. At the same time, the model performance is evaluated on the validation set, and the hyperparameters are adjusted to prevent overfitting. Finally, the model is evaluated using the test set to ensure that its accuracy, recall, and F1 score meet the application requirements.

[0124] The lithium battery fire recognition module uses the trained MobileViT model to identify lithium battery fires from the real-time collected acoustic signals. The recognition process is as follows: First, the real-time pre-processed acoustic signal is obtained from the data acquisition and pre-processing module, and the signal is converted into a Mel spectrum graph and normalized at the same time. Subsequently, the processed Mel spectrum graph is input into the MobileViT model for forward calculation to generate the fire recognition result. For the recognition result, a probability threshold for fire judgment is set: 0.8. When the probability exceeds this threshold, the system will determine that it is a lithium battery fire.

[0125] Through the collaborative work of modules such as data collection and preprocessing, Mel spectrum feature extraction, MobileViT model construction and training, and lithium battery fire identification, this application achieves high-accuracy and real-time identification of lithium battery fires.

[0126] This application aims to overcome the limitations of the existing technology and proposes a lithium battery fire identification method based on the Mel-MobileViT lightweight network. This method introduces advanced machine learning algorithms, which can analyze the acoustic signals generated by lithium battery fires in real time, and use Mel spectrograms and MobileViT lightweight networks to train data to improve the accuracy and real-time performance of fire identification.

[0127] The technical solution adopted in this application is: a lithium battery fire identification method based on Mel spectrum features and MobileViT lightweight network, which mainly includes a data acquisition and preprocessing module, a Mel spectrum feature extraction module, a MobileViT model construction and training module, and a lithium battery fire identification module.

[0128] The features and beneficial effects of this application are as follows:

[0129] 1. Introduce machine learning algorithms to analyze acoustic data in real time: By collecting acoustic signals generated by lithium battery fires, the adaptive chirp pattern decomposition method is used to reduce noise and extract pure acoustic features. Subsequently, the processed acoustic signals are converted into Mel spectrograms as the acoustic fingerprint features of the fire. The MobileViT lightweight network is used to train and identify these features, achieving real-time and efficient identification of lithium battery fires.

[0130] 2. Improve the accuracy and real-time performance of fire identification: This application significantly improves the accuracy of fire identification and reduces the occurrence of false positives and false negatives through in-depth mining of acoustic features by deep learning models. The use of lightweight models makes the identification process fast in calculation speed, meeting the needs of real-time detection and rapid response.

[0131] 3. Non-contact detection improves safety: Using acoustic sensors for non-contact data collection avoids the difficulty of sensor deployment in high temperature or dangerous environments, improves the safety and adaptability of the system, reduces dependence on traditional temperature, smoke and other sensors, and reduces hardware costs and maintenance difficulties.

[0132] 4. Lightweight model design, easy to deploy: MobileViT has small network parameters and low computational complexity, making it suitable for running on resource-constrained devices without the need for high-performance computing resources. It reduces the requirements for hardware devices, facilitates deployment on ordinary computing devices and even mobile terminals, and reduces the overall cost of the system.

[0133] 5. Overcoming the defects of existing technologies and achieving advantages: This application overcomes the shortcomings of traditional methods in terms of accuracy, real-time and environmental adaptability through advanced machine learning algorithms and acoustic signal processing technology. It achieves high-accuracy and real-time identification of lithium battery fires, meeting the needs of accident investigation and emergency response.

[0134] The advancement of this application over the prior art is as follows:

[0135] Structure: This application adopts a technical route that combines ACMD noise reduction processing, Mel spectrum feature extraction and MobileViT lightweight network, realizing the complete process from data acquisition, preprocessing, feature extraction to model training and recognition, and improving the overall performance of the system.

[0136] Functional aspects: It improves the accuracy and real-time performance of fire identification, can detect lithium battery fires in time, and reduce accident losses. The non-contact acoustic signal collection method improves the safety and applicability of the system.

[0137] In terms of process: the data processing flow has been optimized, and a lightweight model has been used to reduce computational complexity and improve the efficiency of model training and reasoning.

[0138] Cost: It reduces the dependence on hardware equipment, simplifies the deployment and maintenance of the system, and reduces the construction and operation costs of the system.

[0139] The present invention overcomes the defects of the prior art and achieves the following advantages:

[0140] Overcoming the defects of slow response speed and high false alarm rate: By real-time collection and analysis of acoustic signals of lithium battery fires and using advanced machine learning algorithms, fast and accurate fire identification is achieved, significantly improving response speed and identification accuracy.

[0141] Reduce dependence on hardware devices: The use of lightweight MobileViT models and non-contact acoustic sensors reduces dependence on high-performance hardware and complex sensors, simplifies the system's hardware configuration, and reduces costs.

[0142] Improve the reliability and stability of the system: Through ACMD noise reduction processing and Mel spectrum feature extraction, the model's resistance to environmental noise is enhanced, ensuring the reliability of the recognition results. The highly automated recognition process reduces human intervention and reduces the risk of misjudgment.

[0143] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0144] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for identifying thermal runaway fire of a lithium battery, characterized in that: The lithium battery thermal runaway fire identification method comprises: The acoustic sensor is used to collect historical acoustic signals of multiple lithium batteries in different states in real time; the different states include: fire and non-fire; the fire includes explosion and non-explosion; Preprocessing each historical acoustic signal to obtain each preprocessed historical acoustic signal; the preprocessing includes: using an adaptive chirp mode decomposition method to remove environmental noise in the historical acoustic signal; Convert each preprocessed historical acoustic signal into a normalized logarithmic Mel-spectrogram; All normalized logarithmic Mel-spectrograms and the status of the lithium battery are combined into a data set, and the MobileViT model is trained to obtain a trained MobileViT model; The real-time acoustic signal of the lithium battery to be tested is collected by an acoustic sensor, and the real-time acoustic signal is preprocessed and converted to obtain a normalized logarithmic Mel spectrum of the real-time acoustic signal; The normalized logarithmic Mel-spectrogram of the real-time acoustic signal is input into the trained MobileViT model to obtain a thermal runaway fire recognition result of the lithium battery to be tested; the thermal runaway fire recognition result is fire or non-fire.

2. The method for identifying thermal runaway fire of a lithium battery according to claim 1, characterized in that: Preprocess each historical acoustic signal to obtain each preprocessed historical acoustic signal, specifically including: A bandpass filter is used to filter out high-frequency noise and low-frequency noise in each historical acoustic signal to obtain a noise-reduced historical acoustic signal; Adopting the adaptive chirp mode decomposition method to remove the environmental noise in the historical acoustic signal of primary denoising, and obtaining the historical acoustic signal of secondary denoising; The amplitude of the secondary denoised historical acoustic signal is normalized to obtain each preprocessed historical acoustic signal.

3. The method for identifying thermal runaway fire of a lithium battery according to claim 2, characterized in that: Adopting the adaptive chirp mode decomposition method to remove the environmental noise in the historical acoustic signal of primary noise reduction, the historical acoustic signal of secondary noise reduction is obtained, which specifically includes: An adaptive chirp mode decomposition method is used to decompose the once-denoised historical acoustic signal into multiple chirp mode functions; Calculate the correlation coefficient between each chirp pattern function and the historical acoustic signal of primary noise reduction; retaining chirp pattern functions whose correlation coefficients are greater than a preset coefficient threshold; The retained chirp pattern functions are superimposed to obtain the secondary denoised historical acoustic signal.

4. The method for identifying thermal runaway fire of a lithium battery according to claim 1, characterized in that: Each preprocessed historical acoustic signal is converted into a normalized logarithmic Mel-spectrogram, including: Performing short-time Fourier transform on each preprocessed historical acoustic signal to obtain a frequency spectrum of each historical acoustic signal; The spectrogram of each historical acoustic signal is weighted and summed through a Mel filter bank to obtain the energy of each Mel frequency band; The energy of each Mel frequency band is converted to a logarithmic scale to obtain the logarithmic energy of each Mel frequency band; The logarithmic energy of each Mel frequency band is normalized by mean and standard deviation to generate a normalized log Mel spectrogram of each historical acoustic signal.

5. The method for identifying thermal runaway fire of a lithium battery according to claim 1, characterized in that: The MobileViT model includes: a convolutional layer, a visual Transformer module and an output layer connected in sequence; The convolution layer includes a convolutional neural network, wherein the convolutional neural network uses a depthwise separable convolution; the convolutional neural network is used to extract local features of the normalized logarithmic Mel-spectrogram; The visual Transformer module is used to model the global features of the normalized log-mel spectrogram through a multi-head self-attention mechanism according to the local features; The output layer is used to predict the probability distribution of fire and non-fire according to the global characteristics of the normalized logarithmic Mel-frequency spectrum, and output the thermal runaway fire recognition result of the lithium battery.

6. The method for identifying thermal runaway fire of a lithium battery according to claim 1, characterized in that: The MobileViT model is lightweight using pruning and quantization techniques.

7. The method for identifying thermal runaway fire of a lithium battery according to claim 1, characterized in that: During the training of the MobileViT model: Use Adam optimizer to update the parameters of MobileViT model; The loss function uses the cross entropy loss function; Regularization, Dropout, and early stopping are used to prevent overfitting.

8. The method for identifying thermal runaway fire of a lithium battery according to claim 1, characterized in that: The normalized logarithmic Mel-spectrogram of the real-time acoustic signal is input into the trained MobileViT model to obtain the thermal runaway fire recognition results of the lithium battery to be tested, including: Set the probability threshold for fire determination; After the normalized logarithmic Mel-spectrogram of the real-time acoustic signal is input into the trained MobileViT model, if the fire probability obtained by the trained MobileViT model is greater than or equal to the probability threshold, the thermal runaway fire identification result of the lithium battery to be tested is fire; if the fire probability obtained by the trained MobileViT model is less than the probability threshold, the thermal runaway fire identification result of the lithium battery to be tested is non-fire.

9. A lithium battery thermal runaway fire identification system, characterized in that: The lithium battery thermal runaway fire identification system comprises: an acoustic sensor and a processing device; The acoustic sensor is connected to the processing device; The acoustic sensor is used to collect historical acoustic signals of multiple lithium batteries in different states in real time; the different states include: fire and non-fire; the fire includes explosion and non-explosion; The processing device is used to preprocess each historical acoustic signal to obtain each preprocessed historical acoustic signal; the preprocessing includes: using an adaptive chirp mode decomposition method to remove environmental noise in the historical acoustic signal; converting each preprocessed historical acoustic signal into a normalized logarithmic Mel spectrum; all normalized logarithmic Mel spectrums and the state of the lithium battery together form a data set, and training the MobileViT model to obtain a trained MobileViT model; The acoustic sensor is also used to collect real-time acoustic signals of the lithium battery to be tested; The processing device is also used to preprocess and convert the real-time acoustic signal to obtain a normalized logarithmic Mel spectrum of the real-time acoustic signal; input the normalized logarithmic Mel spectrum of the real-time acoustic signal into the trained MobileViT model to obtain a thermal runaway fire identification result of the lithium battery to be tested; the thermal runaway fire identification result is fire or non-fire.

10. The lithium battery thermal runaway fire identification system according to claim 9, characterized in that: The lithium battery thermal runaway fire identification system also includes: a data acquisition card and an interface; The data acquisition card is connected to the acoustic sensor and the interface respectively; the interface is connected to the processing device; The data acquisition card is used to collect acoustic signals measured by the acoustic sensor and transmit the acoustic signals to the processing device through an interface.

Images