Power battery thermal runaway soundprint early warning method and system

Abstract

The present invention relates to the technical field of power battery safety monitoring, and specifically to a power battery thermal runaway soundprint early warning method and system. The method collects soundprint signals in the battery pack cooling duct through a MEMS microphone array, focusing in particular on the acoustic features in the 23-28kHz frequency band; constructs a multi-scale convolutional recurrent neural network voiceprint recognition model, including a cascaded residual feature extraction network, a three-way parallel multi-scale convolutional layer, an adaptive attention mechanism, a feature fusion layer, and a long short-term memory network layer; calculates the entropy of the voiceprint signal and compares it with a preset threshold; and simultaneously monitors the deformation of the battery cell through a strip Bragg grating sensor; when the soundprint entropy exceeds 2.8 times the historical average value and lasts for 20 seconds, and the battery cell deformation reaches 0.15%, a thermal runaway warning is triggered. The present invention sets a multi-level warning mechanism, which can issue a warning 10 to 30 minutes before thermal runaway occurs, providing sufficient time for taking safety measures.

Description

Technical Field

[0001] The present invention relates to the technical field of power battery safety monitoring, and in particular to a power battery thermal runaway soundprint early warning method and system. Background Art

[0002] With the increasing popularity of new energy vehicles and portable electronic devices, lithium-ion batteries are widely used as mainstream power batteries and energy storage batteries. However, under certain conditions, lithium-ion batteries can experience thermal runaway, leading to serious safety accidents. Traditional thermal runaway warning technologies rely primarily on monitoring parameters such as temperature, voltage, and current. These parameters often only show significant changes when thermal runaway is about to occur, resulting in short warning times and ineffective prevention of safety accidents.

[0003] Existing temperature monitoring methods typically only issue warnings when the battery's surface temperature rises sharply, by which time the internal temperature may have already reached dangerous levels. Methods based on voltage and current monitoring are significantly affected by external loads and have limited accuracy. While gas sensor-based methods can detect battery gas leaks, they have long response times and are susceptible to environmental interference.

[0004] Therefore, there is an urgent need for a technology that can provide early warning at an early stage before thermal runaway occurs, so as to provide sufficient time to take safety measures and prevent accidents. Summary of the Invention

[0005] The purpose of the present invention is to provide a power battery thermal runaway soundprint early warning method and system, by analyzing the specific frequency domain soundprint signal generated by the battery electrolyte in the early stage of thermal runaway micro-boiling, combined with battery cell deformation monitoring, to achieve early warning of thermal runaway.

[0006] The present invention proposes an early warning method for power battery thermal runaway soundprint, comprising:

[0007] Collect the soundprint signal in the battery pack cooling duct to obtain the specific frequency domain acoustic characteristics generated during the boiling process of the electrolyte in the battery pack;

[0008] Constructing a multi-scale convolutional recurrent neural network voiceprint recognition model to perform feature extraction and dynamic analysis on the voiceprint signal;

[0009] Calculating the entropy value of the voiceprint signal and comparing it with a preset threshold;

[0010] Obtain battery cell deformation data and monitor the degree of battery cell expansion;

[0011] When the entropy value of the voiceprint signal exceeds a preset threshold and the deformation of the battery cell reaches a predetermined threshold, generating thermal runaway early warning information;

[0012] The multi-scale convolutional recurrent neural network voiceprint recognition model includes:

[0013] A cascaded residual feature extraction network, used to convert the voiceprint signal into a high-dimensional feature vector;

[0014] Three parallel multi-scale convolutional layers, each using convolution kernels of different sizes and different dilation rates, to capture the multi-scale temporal features in the high-dimensional feature vector;

[0015] An adaptive attention mechanism dynamically adjusts the weights of the three parallel multi-scale convolutional layers based on input features;

[0016] a feature fusion layer for fusing the outputs of the three parallel multi-scale convolutional layers; and a long short-term memory network layer for performing time series modeling on the outputs of the feature fusion layer and calculating the entropy value of the voiceprint signal.

[0017] Preferably, collecting the voiceprint signal in the battery pack cooling duct specifically includes:

[0018] A MEMS microphone array is evenly arranged on the inner wall of the battery pack cooling channel, and the MEMS microphone array is set at the gas-liquid separation layer in the cooling channel;

[0019] Collecting voiceprint signals with a sampling rate of 192kHz through the MEMS microphone array;

[0020] Pre-process the collected voiceprint signal to remove environmental noise and background noise;

[0021] The preprocessed voiceprint signal is converted into a frequency domain representation through Fourier transform; and the acoustic features of the 23-28 kHz frequency band are extracted from the frequency domain representation.

[0022] Preferably, the cascaded residual feature extraction network specifically includes:

[0023] Input layer, receiving the pre-processed voiceprint signal;

[0024] Multiple convolutional layers, with residual connections set every fixed number of layers;

[0025] Batch normalization layer, which normalizes the output of each convolutional layer;

[0026] The activation function layer introduces nonlinear transformation capabilities; and the output layer generates a feature vector with a dimension of 512.

[0027] Preferably, the three-way parallel multi-scale convolutional layer specifically includes:

[0028] The first convolution branch uses a 1×1 convolution kernel and a dilation rate of 1 to capture local point features.

[0029] The second convolution branch uses a 3×3 convolution kernel and a dilation rate of 2 to capture medium-range features; and the third convolution branch uses a 5×5 convolution kernel and a dilation rate of 4 to capture large-scale context features.

[0030] Preferably, the adaptive attention mechanism specifically includes:

[0031] Two-layer fully connected neural network, including input layer, hidden layer and output layer;

[0032] The input layer receives the high-dimensional feature vector, the hidden layer has the same dimension as the input layer, and the output layer contains 3 neurons, which correspond to the weights of three parallel multi-scale convolutional layers respectively;

[0033] Normalizing the three weight values ​​of the output layer so that their sum is 1; and applying the normalized weights to the output features of the three parallel multi-scale convolutional layers respectively.

[0034] Preferably, the feature fusion layer specifically includes:

[0035] Divide the feature vector into odd-dimensional features and even-dimensional features;

[0036] For odd-dimensional features, a weighted summation method is used for fusion, where the odd-dimensional features of the three-way convolution branch are multiplied by the corresponding attention weights and then added together;

[0037] For even-dimensional features, a nonlinear interactive method is used for fusion. The first branch feature is multiplied by its weight, and the second and third branch features are multiplied by their respective weights and then multiplied together. The two results are then added together. The processed odd and even dimension features are recombined to form a complete fused feature vector.

[0038] Preferably, the long short-term memory network layer specifically includes:

[0039] A single-layer long short-term memory network, consisting of an input gate, a forget gate, an output gate, and a memory unit;

[0040] A single output neuron, using an activation function;

[0041] Converting the output value of the long short-term memory network into a probability distribution;

[0042] Calculate the information entropy of the probability distribution as the voiceprint entropy value; and when the voiceprint entropy value exceeds 2.8 times the average entropy value of historical samples and lasts for more than 20 seconds, determine that the voiceprint is abnormal.

[0043] Preferably, obtaining the cell deformation data specifically includes:

[0044] A ribbon Bragg grating sensor made of a single-mode optical fiber built into the battery core;

[0045] The band Bragg grating sensor is used to monitor the range change of the expansion deformation of the battery cell in real time;

[0046] Record the cell deformation change ΔS and the corresponding time T;

[0047] A relationship model between ΔS and time is established through nonlinear fitting; and when ΔS reaches or exceeds 0.15%, it is determined that the battery cell deformation is abnormal.

[0048] Preferably, generating thermal runaway early warning information specifically includes:

[0049] Establish a multi-level early warning mechanism, including level one, level two, and level three warnings;

[0050] When the voiceprint entropy value exceeds 2 times the baseline, a level 1 warning is triggered and the monitoring frequency is increased;

[0051] When the voiceprint entropy value exceeds 2.8 times the baseline and lasts for 20 seconds, and the battery cell deformation does not reach 0.15%, a second-level warning is triggered;

[0052] When the deformation of the battery cell reaches 0.15%, a third-level warning is triggered and emergency response is immediately initiated; the warning information is transmitted to the battery management system, and corresponding safety measures are initiated according to the warning level.

[0053] The power battery thermal runaway soundprint early warning system includes:

[0054] The data acquisition subsystem includes a MEMS microphone array arranged on the inner wall of the battery pack cooling duct and a strip Bragg grating sensor placed inside the battery cell, which is used to collect voiceprint signals and battery cell deformation data;

[0055] Signal preprocessing module, used to filter, convert the frequency domain and filter the frequency band of the collected original signal;

[0056] The deep learning analysis subsystem includes a cascaded residual feature extraction network, a three-way parallel multi-scale convolutional layer, an adaptive attention mechanism, a feature fusion layer, and a long short-term memory network layer. It is used to build a multi-scale convolutional recurrent neural network voiceprint recognition model, perform feature extraction and dynamic analysis on the voiceprint signal, and calculate the entropy value of the voiceprint signal;

[0057] The deformation monitoring subsystem is used to process the data of the strip Bragg grating sensor and establish a deformation-time relationship model through nonlinear fitting;

[0058] The early warning decision subsystem is used to generate early warning information of thermal runaway when the entropy value of the voiceprint signal exceeds a preset threshold and the deformation of the battery cell reaches a predetermined threshold; and the edge computing and data management subsystem includes an edge computing node and a battery pack management terminal, which are used to perform real-time analysis and coordinate multiple edge nodes, where the edge computing node and the battery pack management terminal are connected via wireless communication, and the edge computing node feeds back the data processing results to the battery pack management terminal.

[0059] The present invention has the following beneficial effects:

[0060] 1. Early warning time: Voiceprint analysis can issue an early warning in the electrolyte micro-boiling stage (10 to 30 minutes before thermal runaway), which is significantly earlier than traditional temperature monitoring methods;

[0061] 2. Improved accuracy: A multi-scale convolutional recurrent neural network is used to analyze voiceprint features, combined with cell deformation monitoring to form a dual verification mechanism, significantly reducing the false alarm rate;

[0062] 3. Non-invasive monitoring: does not interfere with the internal structure of the battery and can be directly integrated into the existing battery management system;

[0063] 4. Strong adaptability: Through deep learning models, it can continuously adapt to battery aging and environmental changes to maintain long-term effectiveness;

[0064] 5. Complete solution: A complete closed loop is formed from monitoring, analysis, early warning to response, providing comprehensive security protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a flow chart of the power battery thermal runaway soundprint early warning method of the present invention;

[0066] Figure 2 This is a structural diagram of the multi-scale convolutional recurrent neural network voiceprint recognition model of the present invention;

[0067] Figure 3 This is a schematic diagram of the arrangement of the MEMS microphone array in the battery pack cooling duct of the present invention;

[0068] Figure 4 This is a diagram of the cascaded residual feature extraction network structure of the present invention;

[0069] Figure 5 This is a diagram of the structure of the three-way parallel multi-scale convolutional layer of the present invention;

[0070] Figure 6 This is the structural diagram of the adaptive attention mechanism of the present invention;

[0071] Figure 7 This is a workflow diagram of the feature fusion layer of the present invention;

[0072] Figure 8 This is a diagram of the long short-term memory network layer structure of the present invention;

[0073] Figure 9 Schematic diagram of the cell deformation monitoring and nonlinear fitting model of the present invention. DETAILED DESCRIPTION

[0074] Please refer to Figures 1-9 , the embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0075] Reference Figure 1 The present invention provides a power battery thermal runaway soundprint early warning method. The method first collects the soundprint signal in the battery pack cooling pipe to obtain the specific frequency domain acoustic characteristics generated during the boiling process of the electrolyte in the battery pack; then constructs a multi-scale convolutional recurrent neural network soundprint recognition model to perform feature extraction and dynamic analysis on the soundprint signal; then calculates the entropy value of the soundprint signal and compares it with a preset threshold; at the same time, obtains the battery cell deformation data to monitor the degree of battery cell expansion; finally, when the entropy value of the soundprint signal exceeds the preset threshold and the battery cell deformation reaches a predetermined threshold, generates thermal runaway early warning information.

[0076] In one embodiment of the present invention, the multi-scale convolutional recurrent neural network voiceprint recognition model includes a cascaded residual feature extraction network, three parallel multi-scale convolutional layers, an adaptive attention mechanism, a feature fusion layer, and a long short-term memory network layer. Together, these components form an end-to-end voiceprint analysis system that effectively captures the specific voiceprint characteristics produced by boiling electrolytes.

[0077] Reference Figure 3 The present invention collects soundprint signals in a battery pack cooling duct, specifically comprising: evenly arranging a MEMS microphone array on the inner wall of the battery pack cooling duct, wherein the MEMS microphone array is disposed at a gas-liquid separation layer in the cooling duct; collecting soundprint signals with a sampling rate of 192 kHz through the MEMS microphone array; preprocessing the collected soundprint signals to remove ambient and background noise; converting the preprocessed soundprint signals into a frequency domain representation through Fourier transform; and extracting acoustic features in the 23-28 kHz frequency band from the frequency domain representation.

[0078] Preferably, a MEMS microphone is evenly arranged every 5 to 10 cm on the inner wall of the battery pack cooling duct to form a microphone array. This arrangement ensures that sound signals at different locations within the battery pack can be effectively captured. During the micro-boiling stage before thermal runaway, the battery electrolyte will produce sounds in a specific frequency range. Experiments have found that these sounds are mainly concentrated in the 23-28kHz frequency band, so the present invention pays special attention to the acoustic characteristics of this frequency band.

[0079] The high sampling rate of 192kHz is adopted to meet the Nyquist sampling theorem and ensure accurate capture of sound signals up to 28kHz. Each acquisition lasts 10 seconds, generating approximately 1.92M data (10s × 192kHz × 1024). The acquired raw signal undergoes preprocessing, primarily including denoising, normalization, and frequency domain conversion, to improve the accuracy of subsequent analysis.

[0080] Reference Figure 4 The cascaded residual feature extraction network in the present invention specifically includes: an input layer that receives the preprocessed voiceprint signal; multiple convolutional layers, with residual connections set every fixed number of layers; a batch normalization layer that normalizes the output of each convolutional layer; an activation function layer that introduces nonlinear transformation capabilities; and an output layer that generates a feature vector with a dimension of 512.

[0081] In this embodiment of the present invention, ResNet50 is used as the basic network structure. The network contains 50 convolutional layers and introduces a residual connection mechanism. The core idea of ​​residual connection is to set shortcuts every few layers in the network, so that the deep network can directly learn the residual mapping relative to the output of the shallow network, thereby solving the problem of difficult training of deep networks. In specific implementation, a residual connection is set every 2-3 convolutional layers.

[0082] The batch normalization layer normalizes each small batch of data to make the data distribution more stable and accelerate model convergence. The activation function uses ReLU (Rectified Linear Unit), and its mathematical expression is:

[0083] .

[0084] in, is the output value of the ReLU function, is the input value. hour, ;when hour, 0. The ReLU function remains linear in the positive interval and outputs zero in the negative interval. This simple nonlinear transformation can effectively improve the expressiveness of the network while avoiding the gradient vanishing problem of the traditional sigmoid function in deep networks.

[0085] Finally, the cascaded residual feature extraction network maps the voiceprint signal into a 512-dimensional feature vector. This dimension is the optimal value verified by multiple experiments and can control the computational complexity while maintaining sufficient expressive power.

[0086] Reference Figure 5The three-way parallel multi-scale convolutional layer in the present invention specifically includes: a first convolution branch, which uses a 1×1 convolution kernel and a convolution operation with a dilation rate of 1 to capture local point features; a second convolution branch, which uses a 3×3 convolution kernel and a convolution operation with a dilation rate of 2 to capture medium-range features; and a third convolution branch, which uses a 5×5 convolution kernel and a convolution operation with a dilation rate of 4 to capture large-scale context features.

[0087] The core idea of ​​multi-scale convolution is to use convolution kernels of different sizes simultaneously to capture features at different scales. In this paper, three parallel branches use convolution kernels of 1×1, 3×3, and 5×5, respectively, which can simultaneously capture point-like local features, medium-range features, and large-scale contextual features in the voiceprint signal.

[0088] Dilated convolution is a special convolution operation that expands the receptive field without increasing the number of parameters by inserting holes between the convolution kernel elements. The dilation rate represents the spacing between the convolution kernel elements. In this invention, the three branches use dilation rates of 1, 2, and 4 respectively, further expanding the range difference of the receptive field. For a k×k convolution kernel, the effective receptive field size is calculated as:

[0089] .

[0090] in, is the effective receptive field size, which indicates the range of the input area that can be covered by the convolution operation; is the convolution kernel size, which indicates the side length of the convolution kernel; is the dilation rate, which indicates the number of intervals between convolution kernel elements. For example, when and hour, , that is, a 3×3 convolution kernel with an expansion rate of 2 can obtain an effective receptive field of 5×5.

[0091] Therefore, the effective receptive fields of the three branches in this invention are 1×1, 5×5, and 17×17, respectively, covering multiple scales from local to global. Each convolutional branch is followed by a batch normalization layer and a ReLU activation function to improve training stability and introduce nonlinear transformations. The three branches independently process input features and each outputs a feature map of the same dimension for subsequent fusion processing.

[0092] Reference Figure 6The adaptive attention mechanism in the present invention specifically includes: a two-layer fully connected neural network, including an input layer, a hidden layer and an output layer; wherein the input layer receives the high-dimensional feature vector, the hidden layer has the same dimension as the input layer, and the output layer contains 3 neurons, which correspond to the weights of three parallel multi-scale convolutional layers respectively; the three weight values ​​of the output layer are normalized so that their sum is 1; and the normalized weights are applied to the output features of the three parallel multi-scale convolutional layers respectively.

[0093] The core idea of ​​the attention mechanism is to allow the model to focus on the most important parts of the input. In this paper, the adaptive attention mechanism is used to dynamically adjust the importance weights of the three convolutional branches, enabling the model to automatically adjust its attention to features of different scales based on the different input features.

[0094] The implementation uses a two-layer fully connected neural network. The number of neurons in the first layer is the same as the input feature dimension (512), and the second layer contains 3 neurons, corresponding to the three convolution branches. The ReLU activation function is used between the two layers to introduce nonlinear transformation capabilities. The output of the second layer is normalized by softmax to ensure that the sum of the three weights is 1. The calculation formula is:

[0095] .

[0096] in, For the The weight value of each convolution branch ranges from 0 to 1; is the second layer of the fully connected network corresponding to The output value of a neuron; is the base of natural logarithm, approximately equal to 2.71828; and is the branch index, and its value range is , corresponding to three convolution branches respectively; Represents the summation operation of the three branches. The softmax function converts the original output value into a probability distribution, ensuring that the sum of all weight values ​​is 1.

[0097] The normalized weights are automatically adjusted through backpropagation to adapt to different input features. For example, when the input signal contains obvious high-frequency transient features, the model may assign a higher weight to the first branch (1×1 convolution); while when the signal contains a long-term evolution pattern, the model may increase the weight of the third branch (5×5 convolution).

[0098] Reference Figure 7The feature fusion layer in the present invention specifically includes: dividing the feature vector into odd-dimensional features and even-dimensional features; for odd-dimensional features, adopting weighted summation to fuse, multiplying the odd-dimensional features of the three-way convolution branch with the corresponding attention weights and then adding them; for even-dimensional features, adopting nonlinear interaction to fuse, multiplying the first branch feature with its weight, multiplying the second and third branch features with their respective weights and then multiplying them together, and then adding the two parts of the results; and recombining the processed odd and even dimension features to form a complete fused feature vector.

[0099] The innovation of the feature fusion layer is to distinguish between odd and even dimensions and adopt different fusion strategies. This not only preserves the stability of the linear model, but also introduces nonlinear interactions to enhance feature expression capabilities. For odd-dimensional features, a weighted summation linear fusion method is used, and its calculation formula is:

[0100] .

[0101] in, is the odd-dimensional feature after fusion; For the The weight values ​​of the convolution branches are calculated by the adaptive attention mechanism; For the Output features of convolution branches; , represents the branch index. For even-dimensional features, a multiplication operation is introduced to enhance nonlinear interaction, and its calculation formula is:

[0102] .

[0103] in, is the even-dimensional feature after fusion; and The meaning of is the same as above; Represents element-wise multiplication, which multiplies elements at corresponding positions. The multiplication operation can capture high-order correlations between features and produce a stronger response when two features are activated simultaneously.

[0104] Finally, the processed odd and even dimension features are recombined in their original order to form a complete fused feature vector. This fusion strategy has shown better performance than a single fusion method in experiments and can more effectively capture complex patterns in voiceprint signals.

[0105] Reference Figure 8The long short-term memory network layer in the present invention specifically includes: a single-layer long short-term memory network, including an input gate, a forget gate, an output gate and a memory unit; a single output neuron, using an activation function; converting the output value of the long short-term memory network into a probability distribution; calculating the information entropy of the probability distribution as the voiceprint entropy value; and when the voiceprint entropy value exceeds 2.8 times the average entropy value of historical samples and lasts for more than 20 seconds, it is determined that the voiceprint is abnormal.

[0106] Long short-term memory (LSTM) networks are a special type of recurrent neural network that effectively models long-term dependencies in sequential data. In this paper, LSTM networks are used to capture the temporal evolution of voiceprint signals, which is crucial for detecting gradual changes during thermal runaway.

[0107] The core of LSTM is its memory cell and three gating mechanisms: the input gate, the forget gate, and the output gate. The input gate controls the extent to which new information enters the memory cell, the forget gate controls the extent to which historical information is retained, and the output gate controls the extent to which information is output from the memory cell. The calculation formulas for these three gates are:

[0108] .

[0109] .

[0110] .

[0111] in, For the moment The input gate value of , which ranges from (0,1), controls the degree to which new information enters the memory unit; For the moment The forget gate value ranges from (0,1) to control the degree of historical information retention; For the moment The output gate value of is in the range of (0,1), which controls the output degree of the memory unit information; The sigmoid activation function maps the input to the (0,1) interval; 、 、 are the weight matrices of the input gate, forget gate, and output gate respectively; Indicates that the hidden state of the previous moment With the current input The vector formed by connecting; 、 、 They are the bias items of the input gate, forget gate and output gate respectively; Represents a matrix multiplication operation.

[0112] The update formula of the memory unit is:

[0113] .

[0114] .

[0115] in, For the moment The candidate memory cell value represents the new information generated by the current input; For the moment The memory unit value of represents the complete memory at the current moment; is the memory unit value at the previous moment; It is a hyperbolic tangent activation function that maps the input to the (-1,1) interval; is the weight matrix of the candidate memory unit; is the bias term of the candidate memory unit; Represents an element-wise multiplication operation.

[0116] The hidden state is calculated as:

[0117] .

[0118] in, For the moment The hidden state of , which represents the output features of the network; is the output gate value; The result of applying the hyperbolic tangent function to the memory cell value.

[0119] In this invention, the output of the LSTM network passes through a fully connected layer to generate a single output value, which is then converted into a probability distribution and the information entropy is calculated as the voiceprint entropy value. The calculation formula for information entropy is:

[0120] .

[0121] Among them, Entropy is the voiceprint entropy value, which indicates the complexity of the voiceprint signal; Input fusion feature The element value at the moment; is the length of the input fusion feature; Indicates that the length is The sequence is summed; log represents the natural logarithm, that is, Logarithm of base ; It can be regarded as a probability distribution, indicating the The relative size of the moment element value in the entire sequence. The higher the entropy value, the more complex the voiceprint signal is and the closer it is to thermal runaway.

[0122] A voiceprint anomaly is identified when the voiceprint entropy exceeds 2.8 times the average entropy of historical samples and persists for more than 20 seconds. These two thresholds (2.8 times and 20 seconds) were determined through extensive experimentation and effectively reduce false positives while maintaining a high detection rate. Experiments have shown that in the early stages of battery thermal runaway, the entropy of the voiceprint signal generated by electrolyte micro-boiling typically rises sharply, reaching 2.5-3 times the normal value, and this anomaly persists for a long time (typically exceeding 15 seconds). Therefore, choosing 2.8 times and 20 seconds as the thresholds effectively balances sensitivity and specificity.

[0123] Reference Figure 9 The present invention specifically includes: embedding a strip Bragg grating sensor made of single-mode optical fiber in the battery core; monitoring the value range change of the battery core expansion deformation in real time through the strip Bragg grating sensor; recording the battery core deformation change value ΔS and the corresponding time T; establishing a relationship model between ΔS and time through nonlinear fitting; and determining that the battery core deformation is abnormal when ΔS reaches or exceeds 0.15%.

[0124] A ribbon Bragg grating sensor is a highly accurate strain sensor. It operates by exploiting the Bragg reflection principle of light. When an optical fiber is subjected to strain, the wavelength of the reflected light changes. This change can be used to calculate the strain value. In this invention, a ribbon Bragg grating sensor fabricated using single-mode optical fiber exhibits a stable linear response within a strain range of 0% to 1.6%.

[0125] The sensor is placed inside the battery cell to monitor the cell's expansion and deformation in real time. When the battery experiences thermal runaway, the internal temperature rises, causing the cell to expand. The cell deformation change value ΔS and the corresponding time T are recorded, and then a nonlinear fitting model is established to establish the relationship between the two. The fitting formula is:

[0126] .

[0127] in, is the cell deformation change value, which represents the percentage deformation relative to the initial state; is the time change value, which represents the time increment relative to the initial measurement moment, in seconds; is the slope of the fitting curve, which indicates the rate of deformation change over time, in % / s; is the offset of the fitting curve, indicating the initial deformation value, in %.

[0128] Extensive experiments have shown that when cell deformation reaches 0.15%, the battery is already in the early stages of thermal runaway. If no action is taken, it will quickly develop into full thermal runaway. Therefore, the present invention sets 0.15% as the threshold for determining abnormal cell deformation.

[0129] Reference Figure 1 The generation of thermal runaway early warning information in the present invention specifically includes: setting up a multi-level warning mechanism, including level one warning, level two warning and level three warning; when the voiceprint entropy value exceeds 2 times the baseline, triggering the level one warning and increasing the monitoring frequency; when the voiceprint entropy value exceeds 2.8 times the baseline and lasts for 20 seconds, and the battery cell deformation does not reach 0.15%, triggering the level two warning; when the battery cell deformation reaches 0.15%, triggering the level three warning and immediately initiating emergency response; and transmitting the warning information to the battery management system, and initiating corresponding safety measures according to the warning level.

[0130] The multi-level warning mechanism enables different levels of response based on the severity of the anomaly, ensuring timely prevention of risks while avoiding unnecessary overreaction. A Level 1 warning emphasizes heightened vigilance and increased monitoring frequency; a Level 2 warning indicates a significant increase in the risk of thermal runaway, but not yet an emergency, and emergency measures can be prepared; a Level 3 warning indicates that thermal runaway has already begun, requiring immediate emergency action.

[0131] Warning information is transmitted to the battery management system via a communication interface, including the warning level, anomaly type (voiceprint anomaly / deformation anomaly), anomaly location (determined by the spatial positioning of the MEMS microphone array), and the expected time for thermal runaway to develop (calculated using a nonlinear fitting model). Based on the received warning information, the battery management system initiates appropriate safety measures, such as reducing charge and discharge currents, initiating forced cooling, disconnecting the problematic battery module, or, in extreme cases, activating the passive fire protection system (spraying inert gases such as nitrogen into the battery pack to cool it down).

[0132] The present invention provides an early warning system for thermal runaway soundprints of power batteries, which includes: a data acquisition subsystem, a signal preprocessing module, a deep learning analysis subsystem, a deformation monitoring subsystem, a warning decision subsystem, and an edge computing and data management subsystem.

[0133] The data acquisition subsystem includes a MEMS microphone array located on the inner wall of the battery pack's cooling duct and a ribbon Bragg grating sensor placed inside the battery cell. The MEMS microphone array collects the soundprint signal, while the ribbon Bragg grating sensor collects the battery cell deformation data.

[0134] The MEMS microphones are designed to be industrial-grade, dust-proof and waterproof, with an operating temperature range of -40°C to 85°C, making them suitable for the harsh environment inside the battery pack. The microphone array is evenly spaced along the inner wall of the cooling duct, with a distance of 5 to 10 cm between adjacent microphones, enabling effective capture of sound signals from different locations within the battery pack. Each microphone has a sampling rate of 192kHz and a precision of 24 bits, meeting the requirements for capturing high-frequency sound signals.

[0135] The ribbon Bragg grating sensor, fabricated from single-mode optical fiber and placed inside the battery cell, monitors the cell's expansion deformation in real time. The sensor is highly sensitive, exhibiting a stable linear response within a deformation range of 0% to 1.6%, with a measurement accuracy of 0.001%.

[0136] The signal preprocessing module is responsible for filtering, frequency domain conversion, and frequency band selection of the collected raw signals. For voiceprint signals, high-pass filtering is first performed to remove low-frequency ambient noise, followed by adaptive filtering to remove background noise. Next, a fast Fourier transform is used to convert the time domain signal into a frequency domain representation. Finally, acoustic features in the 23-28kHz frequency band are extracted from the frequency domain representation.

[0137] Preferably, the high-pass filter cutoff frequency is set to 20 kHz to effectively remove most ambient noise (typically concentrated below 20 kHz). The fast Fourier transform uses a Hanning window with a window length of 8192 points and a 50% overlap, which can reduce spectral leakage while maintaining frequency resolution.

[0138] The deep learning analysis subsystem is the core of the present invention, which includes a cascaded residual feature extraction network, a three-way parallel multi-scale convolutional layer, an adaptive attention mechanism, a feature fusion layer and a long short-term memory network layer. It is used to construct a multi-scale convolutional recurrent neural network voiceprint recognition model, perform feature extraction and dynamic analysis on the voiceprint signal, and calculate the entropy value of the voiceprint signal.

[0139] The deformation monitoring subsystem processes the data from the strip Bragg grating sensor and establishes a deformation-time relationship model through nonlinear fitting. It specifically includes a signal acquisition unit, a signal demodulation unit, a data processing unit, and a model calculation unit.

[0140] The signal acquisition unit is responsible for receiving the optical signal from the strip Bragg grating sensor, the signal demodulation unit converts the optical signal into an electrical signal and calculates the deformation value, the data processing unit filters and normalizes the deformation data, and the model calculation unit establishes a relationship model between deformation and time through nonlinear fitting.

[0141] Preferably, signal demodulation uses wavelength demodulation, achieving a measurement accuracy of 1 pm (corresponding to a deformation accuracy of approximately 0.001%). Nonlinear fitting uses the least squares method, with the fitting formula being ΔS = m × ΔT + n, where m and n are obtained through iterative calculations to minimize the mean square error between the fitting curve and the actual data.

[0142] The early warning decision subsystem is used to generate early warning information for thermal runaway when the entropy value of the voiceprint signal exceeds a preset threshold and the battery cell deformation reaches a predetermined threshold. It specifically includes a data fusion unit, a threshold judgment unit, an early warning generation unit, and a communication interface unit.

[0143] The data fusion unit is responsible for integrating the voiceprint analysis results and deformation monitoring results. The threshold judgment unit determines whether to trigger an early warning based on preset rules. The early warning generation unit generates early warning information of different levels according to the degree of abnormality. The communication interface unit transmits the early warning information to the battery management system.

[0144] Preferably, warnings are divided into three levels: Level 1 (voiceprint entropy exceeds 2 times the baseline), Level 2 (voiceprint entropy exceeds 2.8 times the baseline for 20 seconds, and the battery cell deformation does not reach 0.15%), and Level 3 (battery cell deformation reaches 0.15%). Different warning levels will trigger different safety measures, ranging from increasing monitoring frequency to initiating emergency response.

[0145] The edge computing and data management subsystem consists of edge computing nodes and a battery pack management terminal, which are used to perform real-time analysis and coordinate multiple edge nodes. The edge computing nodes are deployed near the battery packs and directly process the collected data, reducing communication latency. The battery pack management terminal coordinates multiple edge nodes and manages global policies.

[0146] The edge computing node should preferably use a low-power ARM processor with a main frequency of no less than 1.5GHz, a memory of no less than 2GB, and a storage space of no less than 16GB, capable of meeting the operational requirements of deep learning models. The edge computing node and the battery pack management terminal are connected via Bluetooth 5.0 or Wi-Fi 802.11ac wireless communication, with a communication delay of no more than 50ms, enabling real-time data exchange.

[0147] The edge computing node is responsible for performing voiceprint signal processing and deep learning analysis, transmitting the analysis results (such as voiceprint entropy and anomaly determination) to the battery pack management terminal. The battery pack management terminal adjusts the warning thresholds of each node based on global information to ensure that the system maintains high accuracy under different operating environments. In addition, the battery pack management terminal is responsible for managing the voiceprint fingerprint database, which stores a large number of normal and abnormal voiceprint samples for model training and updates.

[0148] In summary, the power battery thermal runaway acoustic early warning method and system provided by this invention achieves early warning of thermal runaway by analyzing the specific frequency domain acoustic signals generated by the battery electrolyte during the micro-boiling stage, which is the early stage of thermal runaway, combined with cell deformation monitoring. Compared with traditional methods, this invention can issue an early warning 10 to 30 minutes before thermal runaway occurs, providing ample time for safety measures to be implemented, significantly improving the safety of power battery systems.

[0149] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. The early warning method of power battery thermal runaway soundprint is characterized by: include: Collect the soundprint signal in the battery pack cooling duct to obtain the specific frequency domain acoustic characteristics generated during the boiling process of the electrolyte in the battery pack; Constructing a multi-scale convolutional recurrent neural network voiceprint recognition model to perform feature extraction and dynamic analysis on the voiceprint signal; Calculating the entropy value of the voiceprint signal and comparing it with a preset threshold; Obtain cell deformation data and monitor cell expansion; When the entropy value of the voiceprint signal exceeds a preset threshold and the deformation of the battery cell reaches a predetermined threshold, generating thermal runaway early warning information; The multi-scale convolutional recurrent neural network voiceprint recognition model includes: A cascaded residual feature extraction network, used to convert the voiceprint signal into a high-dimensional feature vector; Three parallel multi-scale convolutional layers, each using convolution kernels of different sizes and different dilation rates, to capture the multi-scale temporal features in the high-dimensional feature vector; An adaptive attention mechanism dynamically adjusts the weights of the three parallel multi-scale convolutional layers based on input features; A feature fusion layer, which fuses the outputs of the three parallel multi-scale convolutional layers; and A long short-term memory network layer performs time series modeling on the output of the feature fusion layer and calculates the entropy value of the voiceprint signal; The collecting of the voiceprint signal in the battery pack cooling duct specifically includes: A MEMS microphone array is evenly arranged on the inner wall of the battery pack cooling channel, and the MEMS microphone array is set at the gas-liquid separation layer in the cooling channel; Collecting voiceprint signals with a sampling rate of 192kHz through the MEMS microphone array; Pre-process the collected voiceprint signal to remove environmental noise and background noise; Convert the preprocessed voiceprint signal into frequency domain representation through Fourier transform; and extracting acoustic features in the 23-28 kHz frequency band from the frequency domain representation; The long short-term memory network layer specifically includes: A single-layer long short-term memory network, consisting of an input gate, a forget gate, an output gate, and a memory unit; A single output neuron, using an activation function; Converting the output value of the long short-term memory network into a probability distribution; Calculating the information entropy of the probability distribution as the voiceprint entropy value; and When the voiceprint entropy value exceeds 2.8 times the average entropy value of historical samples and lasts for more than 20 seconds, it is determined to be a voiceprint abnormality; The feature fusion layer specifically includes: Divide the feature vector into odd-dimensional features and even-dimensional features; For odd-dimensional features, a weighted summation method is used for fusion, where the odd-dimensional features of the three-way convolution branch are multiplied by the corresponding attention weights and then added together; For even-dimensional features, a nonlinear interaction method is used to fuse them. The first branch feature is multiplied by its weight, and the second and third branch features are multiplied by their respective weights and then multiplied together. The two results are then added together. The processed odd and even dimension features are recombined to form a complete fused feature vector.

2. The power battery thermal runaway soundprint early warning method according to claim 1, characterized in that: The cascaded residual feature extraction network specifically includes: Input layer, receiving the pre-processed voiceprint signal; Multiple convolutional layers, with residual connections set every fixed number of layers; Batch normalization layer, which normalizes the output of each convolutional layer; Activation function layer, which introduces nonlinear transformation capabilities; and The output layer generates a feature vector of dimension 512.

3. The power battery thermal runaway soundprint early warning method according to claim 1, characterized in that: The three-way parallel multi-scale convolutional layer specifically includes: The first convolution branch uses a 1×1 convolution kernel and a dilation rate of 1 to capture local point features. The second convolutional branch uses a 3×3 convolution kernel and a dilation rate of 2 to capture medium-range features; and The third convolution branch uses a 5×5 convolution kernel and a convolution operation with a dilation rate of 4 to capture a wide range of context features.

4. The power battery thermal runaway soundprint early warning method according to claim 1, characterized in that: The adaptive attention mechanism specifically includes: Two-layer fully connected neural network, including input layer, hidden layer and output layer; The input layer receives the high-dimensional feature vector, the hidden layer has the same dimension as the input layer, and the output layer contains 3 neurons, which correspond to the weights of three parallel multi-scale convolutional layers respectively; Normalizing the three weight values ​​of the output layer so that their sum is 1; and The normalized weights are applied to the output features of the three parallel multi-scale convolutional layers respectively.

5. The power battery thermal runaway soundprint early warning method according to claim 1, characterized in that: The obtaining of cell deformation data specifically includes: A ribbon Bragg grating sensor made of a single-mode optical fiber built into the battery core; The band Bragg grating sensor is used to monitor the range change of the expansion deformation of the battery cell in real time; Record the cell deformation change ΔS and the corresponding time T; Establishing a ΔS-time relationship model through nonlinear fitting; and When ΔS reaches or exceeds 0.15%, it is determined that the battery cell deformation is abnormal.

6. The power battery thermal runaway soundprint early warning method according to claim 1, characterized in that: The generated thermal runaway early warning information specifically includes: Establish a multi-level early warning mechanism, including level one, level two, and level three warnings; When the voiceprint entropy value exceeds 2 times the baseline, a level 1 warning is triggered and the monitoring frequency is increased; When the voiceprint entropy value exceeds 2.8 times the baseline and lasts for 20 seconds, and the battery cell deformation does not reach 0.15%, a second-level warning is triggered; When the cell deformation reaches 0.15%, a level 3 warning is triggered and emergency response is initiated immediately; and The warning information is transmitted to the battery management system, and corresponding safety measures are initiated according to the warning level.

7. A power battery thermal runaway soundprint early warning system, used to implement the power battery thermal runaway soundprint early warning method according to any one of claims 1 to 6, characterized in that: include: The data acquisition subsystem includes a MEMS microphone array arranged on the inner wall of the battery pack cooling duct and a strip Bragg grating sensor placed inside the battery cell, which is used to collect voiceprint signals and battery cell deformation data; The signal preprocessing module is used to filter, convert the frequency domain and filter the frequency band of the collected original voiceprint signal; The deep learning analysis subsystem includes a cascaded residual feature extraction network, a three-way parallel multi-scale convolutional layer, an adaptive attention mechanism, a feature fusion layer, and a long short-term memory network layer. It is used to build a multi-scale convolutional recurrent neural network voiceprint recognition model, perform feature extraction and dynamic analysis on the voiceprint signal, and calculate the entropy value of the voiceprint signal; The deformation monitoring subsystem is used to process the data of the strip Bragg grating sensor and establish a deformation-time relationship model through nonlinear fitting; The early warning decision subsystem is used to generate early warning information of thermal runaway when the entropy value of the voiceprint signal exceeds a preset threshold and the deformation of the battery cell reaches a predetermined threshold; as well as The edge computing and data management subsystem includes edge computing nodes and battery pack management terminals, which are used to perform real-time analysis and coordinate multiple edge nodes. The edge computing nodes and battery pack management terminals are connected via wireless communication, and the edge computing nodes feed back data processing results to the battery pack management terminals.

Images