A lithium battery thermal runaway early warning method based on a thermally induced sound-producing material

Abstract

The application provides a lithium battery thermal runaway early warning method based on a thermal sound-producing material, and belongs to the technical field of battery safety monitoring; first, a thermal acoustic response layer containing thermal expansion microcapsules is covered on the surface of a battery monomer; when local abnormal temperature rise reaches a preset threshold, the microcapsules are instantaneously brittlely broken, and specific high-frequency pulse sound waves are released; second, a monomer gap is used as a parallel plate acoustic waveguide, and a MEMS microphone array is used to collect mixed sound waves in real time; audio is converted into a Log-Mel spectrogram and input into a lightweight convolutional neural network; deep separable convolution is used to extract voiceprint features, and an adaptive loss function is combined to accurately separate weak signals from complex background noise; finally, time density integration logic is used for confirmation, and BMS linkage is triggered to cut off. The application deeply integrates passive physical phase change and AI recognition, and realizes "extremely early, high-precision and full-coverage" active early warning before the critical point of thermal runaway.

Description

Technical Field

[0001] This invention belongs to the field of battery safety monitoring technology, and in particular relates to an early warning method for thermal runaway of lithium batteries based on thermo-acoustic materials. Background Technology

[0002] With the explosive growth of new energy vehicles and electrochemical energy storage power stations, the safety issues of high-energy-density lithium-ion batteries are becoming increasingly prominent. Lithium-ion batteries are highly susceptible to thermal runaway under abusive conditions such as overcharging, internal short circuits, or high external temperatures. Thermal runaway is essentially a chain reaction of exothermic chemical reactions within the lithium battery, typically involving SEI film decomposition, separator melting, and cathode material decomposition, ultimately leading to battery fire or even explosion. Currently, the key to preventing and controlling lithium battery thermal runaway lies in "early detection and early warning." However, the time window from a micro-short circuit or localized overheating within the lithium battery to a full-blown thermal runaway often lasts only a few minutes or even seconds. How to quickly and accurately detect abnormal signals in the early stages of thermal runaway (i.e., before the lithium battery pressure relief valve opens or before violent combustion) is a significant challenge currently facing lithium battery safety management systems.

[0003] In existing technologies, early warning systems for lithium-ion battery thermal runaway primarily rely on BMS systems to monitor voltage, current, and temperature, as well as environmental monitoring based on gas or smoke sensors. However, these technologies have significant drawbacks and limitations in practical applications: First, temperature monitoring suffers from lag and spatial blind spots. Existing NTC thermistors are typically attached to the battery surface or module terminals, representing "point-based monitoring." Due to thermal resistance in heat conduction from the battery's interior to the surface, surface temperature changes often lag behind internal reactions. Traditional contact-type temperature sensors, limited by the number of points and heat conduction rate, struggle to capture early temperature rises inside the battery or in specific areas in real time, often triggering alarms only after thermal runaway has occurred. Furthermore, for large energy storage containers, the limited number of sensors makes it difficult to cover every cell, easily missing localized hotspots.

[0004] Second, the gas / smoke monitoring and early warning mechanism is too late. Gas or smoke detection requires waiting for the battery pressure relief valve to open and release a large amount of electrolyte vapor, at which point the battery is often already in an irreversible thermal runaway phase. Because the response only occurs after the electrolyte vapor is released, the early warning window is too short, making it difficult to take effective suppression measures.

[0005] Third, traditional acoustic monitoring has weak anti-interference capabilities and suffers from latency. Although some studies have attempted to use acoustic sensors to listen to the sound of the battery pressure relief valve opening, in actual operating environments, background noise such as the fan noise of the energy storage system and the vibration noise of the vehicle is extremely high. Moreover, the pressure relief sound of the safety valve itself has a wide characteristic frequency band, making it difficult for simple acoustic monitoring to effectively distinguish the battery pressure relief sound from environmental background noise such as vehicle operation and fan cooling, which can easily lead to false alarms or missed alarms and cannot meet the needs of accurate safety monitoring under complex operating conditions.

[0006] In summary, given the shortcomings of existing technologies, such as spatial coverage blind spots, time response lag, and weak anti-interference capabilities, there is an urgent need for an active early warning technology that can overcome the limitations of "point-based temperature measurement" and has strong anti-interference capabilities. This would solve the technical challenge of existing technologies failing to provide effective early warning before the thermal runaway critical point, thus gaining valuable time for the safe handling of the system. Summary of the Invention

[0007] To address the above problems, this invention proposes an early warning method for thermal runaway in lithium batteries based on thermoacoustic materials, comprising the following steps: S1, Construct a thermo-acoustic response layer and cover it on the surface of a battery cell; The response layer contains thermally expanding microcapsules, which rupture when the battery surface temperature reaches a preset threshold and actively emit pulsed sound wave signals with specific acoustic characteristics; S2, through a distributed sound acquisition unit deployed inside the battery module, collects mixed composite sound wave signals inside the battery box in real time and transmits them to the preprocessing module; S3, preprocesses the collected sound signal and constructs its time-frequency features, and inputs it into the AI ​​intelligent signal processing unit based on a lightweight convolutional neural network; the AI ​​intelligent signal processing unit identifies specific acoustic features corresponding to microcapsule rupture from the mixed composite sound wave signal and outputs the confidence level of the thermal anomaly event; S4, the early warning decision unit makes a logical judgment based on the confidence level of the thermal anomaly event. When the number of popping sounds per unit time exceeds the set threshold, it is determined to be an early sign of thermal runaway and sends a linkage control and cut-off command to the battery management system.

[0008] Preferably, the construction of the thermo-acoustic response layer in S1 is specifically as follows: The thermo-acoustic response layer comprises a polymer matrix and thermally expandable microcapsules dispersed therein; the microcapsules consist of a thermoplastic polymer shell composed of vinylidene chloride, acrylonitrile and methyl methacrylate copolymerized and a low-boiling-point alkane core material encapsulated inside; trimethylolpropane trimethacrylate is added as a crosslinking agent during the copolymerization process to limit the slippage of polymer chains at high temperatures and ensure that the microcapsules undergo transient brittle fracture upon rupture; Based on the principle of acoustic impedance mismatch, the microcapsules are dispersed in a low-viscosity alicyclic epoxy acrylate matrix resin at a volume fraction of not less than 40% to form a slurry, and then cured on the battery surface by an ultraviolet curing system to form a semi-exposed "island-sea" acoustic structure of microcapsules.

[0009] Preferably, the transient rupture behavior of the microcapsule at the preset initial sound emission temperature satisfies the following critical sound emission mathematical constraint model: The critical condition for microcapsule rupture is given by the formula Decide; in, The dynamic circumferential stress of the microcapsule shell satisfies the following formula: ; This represents the real-time absolute temperature of the battery surface. Due to external environmental pressures, The thickness of the microcapsule shell, The radius of the microcapsule is... For low-boiling-point alkane core materials at temperature The saturated vapor pressure at this temperature follows the Antoine equation: ; Where A, B, and C are the antoine constants of isooctane.

[0010] For the polymer shell at temperature The yield strength at this point follows a linear softening decay model: ; in Reference temperature The initial yield strength is below. The temperature softening coefficient of the material; Preferably, the acquisition method of the distributed sound acquisition unit in S2 is as follows: A capacitive MEMS microphone array is selected and arranged on the inner wall of the side plate of the battery module; the physical gap between the battery cells is used as a parallel plate acoustic waveguide channel, and the sound inlet of the microphone faces the outlet of the gap to receive high-frequency pulse sound waves propagating through total internal reflection via the acoustic waveguide channel. Establish a discrete convolutional acoustic channel model based on room impulse response, microphone Acquired mixed discrete signal Expressed as: ; in, For microphone At the point of time The acquired discrete-time mixing signal values, This represents the time point number of the current discrete sampling. The number of delay-sum index steps for discrete convolution. Given a finite length of the room impulse response sequence, For the explosion sound source at a certain point in time The original signal value, For the source of the explosion The signal value after discrete delay. For at a certain point in time Broadband Gaussian noise in the environment For sound source to microphone The interval corresponds to the delay The impulse response coefficient of the step.

[0011] Preferably, in step S3, the preprocessing and time-frequency feature construction of the acquired sound signal specifically includes: The digital audio signal in the mixed composite sound wave signal is pre-emphasized and framed; the power spectrum of each audio frame is obtained by performing a fast Fourier transform. And the power spectrum is mapped onto the Mel filter bank, linear frequency With Mel scale The conversion formula is: ; Subsequently, the logarithm of the Mel energy is taken to obtain the Log-Mel spectrogram characterizing the transient energy pulse as a two-dimensional feature map.

[0012] Preferably, the AI ​​intelligent signal processing unit recognition process based on a lightweight convolutional neural network in S3 specifically includes: The generated Log-Mel spectrogram slices are used as model input; the network backbone extracts pop sound features through a depthwise separable convolutional structure, specifically including single-channel depthwise convolution to extract vertical acoustic boundary, and pointwise convolution to achieve cross-channel feature fusion. The formula for single-channel depthwise convolution is: ; The formula for pointwise convolution is: ; in, The input feature map in spatial coordinates And the Feature values ​​at each channel For single-channel depthwise convolution kernels in the 1st... Individual channels, local coordinates within the kernel The weighting coefficient at the location, The spatial side length of a single-channel depthwise convolution kernel. and These are the indices for the summation of the horizontal and vertical coordinates within a single-channel depthwise convolution kernel. and These are the global spatial coordinate indices of the feature map in the horizontal and vertical directions, respectively. The output feature map in spatial coordinates after single-channel depthwise convolution. And the Feature values ​​at each channel The total number of channels in the input feature map. The channel summation index of the input feature map. In a 1×1 pointwise convolution kernel, the connection of the first... The input channel and the first The weighting coefficients of each output channel, This refers to the channel index of the output feature map. The final output feature map, after pointwise convolution, is in spatial coordinates. And the The feature values ​​at each channel.

[0013] The extracted high-dimensional acoustic features are transformed into feature vectors through a global average pooling layer, and the sigmoid function of the fully connected layer outputs the probability confidence of the presence of a thermally induced burst signal within the current time window. .

[0014] Preferably, the AI ​​intelligent signal processing unit employs an optimization strategy for extreme imbalance between positive and negative samples during the training phase, specifically: The Mixup data augmentation algorithm was used to enhance the purity of the microcapsule burst sound samples. negative samples with background noise According to the random mixing coefficients that follow a Beta distribution Linear stacking is performed to generate mixed training samples; during model optimization, the Focal Loss loss function with adaptive weighting for difficult samples is used. ; in, The value of the Focal Loss function during the model optimization process. For the true class labels of the training samples, This represents a positive sample (with the sound of microcapsules bursting). This represents a negative sample (background noise). As a balance factor, As a focusing factor, by suppressing the weight contribution of massive simple negative samples to the loss function, the model focuses on the identification of weak burst signals with low signal-to-noise ratio.

[0015] Preferably, the logical judgment and instruction sending by the early warning decision unit in S4 specifically involves: The early warning decision unit introduces a time-sliding window-based approach. The event density integral logic defines the explosion event density function within the window. for: ; in, For the current moment The burst event density function value, The set time sliding window length, For the current system time, For time integration variables, For at any time The probability confidence level of the existence of a thermally induced burst signal in the output. This is an indicator function that takes the value 1 if the condition is met and 0 otherwise. The confidence threshold for a single frame is the hard decision threshold.

[0016] When density function Exceeding the preset event density threshold At that time, it was confirmed as a genuine early sign of thermal runaway; the early warning decision unit then sent a circuit cut-off command to the battery management system and triggered the closed-loop action of the linkage fire-fighting cooling system.

[0017] Compared with the prior art, the present invention has the following innovative features and beneficial effects: (1) Breaking through the spatial blind spots and time lag of traditional perception, constructing a fully covered active acoustic waveguide sensing network: Existing technologies heavily rely on "point-type" contact temperature measurement, which has obvious thermal conduction lag and local monitoring blind spots; or rely on gas / smoke detection, and the early warning time is often in the late stage of thermal runaway and irreversible. This invention innovatively uses a passive thermal expansion microcapsule coating as a distributed sensing array, combined with the parallel plate natural acoustic waveguide formed by the gaps between battery cells, to forcibly convert the hidden microscopic local thermal anomalies into macroscopic high-frequency pulse acoustic wave signals. This full-coverage sensing capability enables the system to keenly capture the very early signs of danger before the critical point of thermal runaway, greatly expanding the golden time window for safe escape and emergency response; (2) Deeply integrate passive physical phase transition and edge-side lightweight AI algorithm to achieve highly interference-resistant and accurate voiceprint recognition: In response to the pain point that traditional simple acoustic monitoring is easily interfered with by complex environmental background noise such as energy storage wind turbines and vehicle vibration, resulting in false alarms, this invention proposes a dual mechanism of "physical customized sound generation + data-driven recognition". First, a critical constraint model based on thermodynamics and thin-wall elasticity is established at the material level to ensure the transient brittle sound generation of microcapsules at a preset dangerous temperature from a physical mechanism perspective; second, a lightweight deep separable convolutional network is introduced at the edge side, and Focal Loss and Mixup strategies are used in training to overcome the extreme imbalance of positive and negative samples, accurately extracting weak popping sound features from massive stray noise background, and realizing high confidence and high interpretability thermal anomaly event assessment with low computing power consumption; (3) Introducing time density integral verification and system-level multi-dimensional linkage to construct a closed-loop safety net mechanism from early diagnosis to physical intervention: Addressing the pain point that single-point abnormal signals easily trigger false alarms leading to system malfunctions, this invention designs an event density integral logic based on a time sliding window at the early warning decision-making level. The system does not rely on a single instantaneous threshold breakthrough, but dynamically assesses the detonation trend within the window. Only when the high-frequency detonation event density exceeds the safety red line is the actual thermal runaway chain reaction finally confirmed. Once the system confirms the diagnosis, it immediately sends an over-level command to the battery management system to implement electrical-level active forced blocking, and simultaneously triggers the physical closed-loop action of the fire thermal management system, achieving seamless connection and physical safety net from very early hazard assessment to energy blocking intervention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0020] Figure 2 A flowchart illustrating the lightweight CNN network architecture and signal processing.

[0021] Figure 3 This is a schematic diagram of microscopic sound generation from a microcapsule.

[0022] Figure 4 This is a schematic diagram of the waveguide inside a lithium battery module.

[0023] Figure 5 Comparison of frequency domain characteristics of microcapsule bursting signal and ambient noise floor. Detailed Implementation

[0024] This invention proposes an early warning method for thermal runaway in lithium batteries based on thermo-acoustic materials. The overall technical flowchart is shown below. Figure 1 As shown, the specific steps are as follows: S1, Construction of the thermo-acoustic response layer: Addressing the issues of hysteresis and blind spots in existing contact-type temperature sensors, this invention utilizes a passive distributed sensor by covering the surface of a battery cell with a composite material layer comprising a polymer matrix and thermally expandable microcapsules. When the battery surface abnormally heats up to a preset detonation temperature, the core material inside the microcapsule vaporizes, causing a violent expansion and rupture, actively releasing a short-duration, high-frequency pulsed acoustic signal.

[0025] S2, a distributed sound acquisition and processing system, utilizes a MEMS microphone array placed inside the battery module, with the gaps between individual battery cells serving as acoustic waveguide channels, to acquire mixed acoustic signals within the battery pack in real time. The acquired analog / digital signals undergo bandpass filtering preprocessing to remove low-frequency interference from road bumps and wind turbines, while retaining the high-frequency characteristics generated by microcapsule bursts.

[0026] S3 is an AI-powered intelligent signal processing and feature recognition algorithm design. The preprocessed audio data from S2 is converted into a two-dimensional spectrogram feature through time-frequency transformation and then input into a lightweight convolutional neural network (CNN). A depthwise separable convolutional structure is used to extract transient vertical texture features of popping sound patterns. Mixup data augmentation and Focal Loss loss functions are introduced during model training to address the imbalance between positive and negative samples, ultimately outputting the confidence level of thermal anomaly events.

[0027] S4, Early Warning Decision and Safety Closed-Loop Execution: The early warning decision unit receives the identification results output from S3 and performs logical judgment. When the detected explosion sound density exceeds the set unit time counting threshold, it is confirmed as a true early sign of thermal runaway. The controller immediately sends a circuit breaker command to the BMS management system and simultaneously triggers the fire cooling system, realizing closed-loop control from early warning to physical backup.

[0028] The specific implementation process of the present invention will be described in detail below with reference to specific embodiments.

[0029] S1. Construction of Thermally Induced Acoustic Response Layer and Mathematical Modeling of Critical Sound Generation In response to the hidden nature of abnormal battery temperature rise, this step aims to construct a passive sensing array that can forcibly transform microscopic thermal evolution into macroscopic acoustic abrupt changes through interdisciplinary modeling of materials synthesis and mechanics-thermodynamics.

[0030] Customized synthesis of S1-1 narrow-distribution thermally expandable microcapsules: To ensure a high degree of consistency in acoustic response frequencies, this embodiment employs suspension polymerization to prepare core-shell structured thermally expandable microcapsules with strictly uniform particle sizes. The preparation process includes oil phase preparation, polymerization, and morphology control.

[0031] (1) Oil phase preparation: Vinylidene chloride (VDC), acrylonitrile (AN), and methyl methacrylate (MMA) were selected as shell comonomers, with a mass ratio of VDC:AN:MMA = 60:30:10. The tight arrangement of VDC molecular chain segments provides high barrier properties, preventing the core material from permeating and leaking during long-term cycling at high and low temperatures. The strong polar nitrile groups of AN enhance the tensile strength of the shell. Isooctane with a standard boiling point of 99℃ was selected as the core material. The key innovation is the addition of 0.5wt% of trimethylolpropane trimethacrylate (TMPTMA) as a crosslinking agent to form a three-dimensional network structure. This significantly restricts the slippage of polymer chains at high temperatures, ensuring that the microcapsules undergo transient brittle fracture (generating high-frequency crisp sound waves) when the critical pressure is reached, rather than ductile yielding (generating low-frequency dull sound waves or silent deformation).

[0032] (2) Polymerization and Appearance Control: Colloidal silica was added to the aqueous phase as a solid stabilizer for the Pickering emulsion (pH=4.0). After mixing the oil and water, the mixture was stirred in a high-shear homogenizer at 3000 rpm to form an emulsion, which was then reacted at 60°C for 20 hours. After washing and drying, microcapsules with a very narrow normal distribution of particle size were obtained, with an average particle size of 25 μm. Shell thickness 2 .

[0033] S1-2 Joint Modeling of Critical Sound Emission Based on Thermodynamics and Thin-Wall Elasticity To ensure the microcapsules burst precisely at the preset 95℃, this invention establishes a thermo-mechanical coupled rupture threshold model. The microcapsule can be approximated as an ideal spherical thin-walled pressure vessel; its rupture is essentially due to the circumferential stress generated by the internal vapor pressure exceeding the ultimate yield strength of the shell material. Based on the generalized Hooke's law and thin-walled vessel theory, the dynamic circumferential stress of the shell layer... The calculation is as follows: ; in, This represents the real-time absolute temperature of the battery surface. Due to external environmental pressures, For shell thickness, The outer radius of the microcapsule. The saturated vapor pressure of the low-boiling-point isooctane core material. The nonlinear, rapid increase in temperature behavior follows Antoine's gas-liquid equilibrium equation: ; Where A, B, and C are the empirical antoine constants for isooctane.

[0034] Meanwhile, the yield strength of the polymer shell Softening occurs with increasing temperature, following a linear softening decay model: ; in, Reference temperature The initial yield strength is below. This is the temperature softening coefficient of the material.

[0035] Combining the above formulas, the extreme critical condition for microcapsules to rupture and produce sound is: This model can be used to guide the wall thickness selection during the microcapsule synthesis stage. By setting the ratio of the crosslinking agent, personalized and precise control of the alarm temperature (such as 80℃, 90℃, or 100℃) can be achieved, enabling microscopic sound generation from the microcapsules. Figure 3 As shown.

[0036] S1-3 Acoustic Impedance Matching and Functional Coating Composite The coating was prepared using a UV curing system, with a low-viscosity alicyclic epoxy acrylate as the base resin. To maximize the radiation efficiency of sound waves into the air, the coating design incorporated the principle of acoustic impedance mismatch. If the surrounding area is completely encapsulated by high-density resin at the moment of microcapsule bursting, a large amount of sound energy will be absorbed by the resin. Therefore, by dispersing microcapsules at a high proportion (volume fraction ≥40%) in the resin at 1500 rpm, a rough "island-sea" structure with numerous semi-exposed microcapsules is formed on the microscopic surface of the coating after spraying and curing. This structure significantly reduces the interface reflection loss of sound waves from the high-density solid phase to the low-density gas phase, enhancing the source-level intensity of acoustic emission.

[0037] S2. Construction of Distributed Sound Acquisition Network and Modeling of Complex Waveguide Channels This step involves building a high-fidelity sound capture system specifically designed for environments with high voltage, strong electromagnetic interference, and strong background mechanical noise.

[0038] S2-1 Array Layout and Utilization of Natural Waveguides: A capacitive MEMS digital microphone with an ultra-high signal-to-noise ratio (SNR≥65dB) is selected. Due to the extremely compact internal space of large energy storage modules, this invention cleverly models the approximately 1-2mm heat dissipation gap between battery cells as a parallel-plate acoustic waveguide. When high-frequency pulsed sound waves propagate in the gap, high-efficiency total internal reflection occurs when their wavelength is much smaller than the gap width. The microphone's inlet is directly opposite the gap outlet, forming a distributed "pickup array." This not only enables monitoring of hotspots in the concealed central area of ​​the module but also provides physical spatial filtering, suppressing noise from non-waveguide directions, such as... Figure 4 As shown.

[0039] S2-2 Signal Conditioning and Channel Modeling Based on Room Impulse Response (RIR): The microphone directly outputs a PDM digital stream, which is transmitted through an isolator to prevent high-voltage side crosstalk. The DSP unit front-end is equipped with a high-order digital high-pass filter with a cutoff frequency of 1000Hz (to filter out the main frequency standing wave and low-frequency path noise of the air-cooled system). Considering that the battery compartment is a sealed metal chamber with severe reverberation, this embodiment establishes a discrete convolutional acoustic channel model based on the room impulse response (RIR). Microphone Acquired mixed discrete signal Expressed as: ; in, This represents the time point number of the current discrete sampling. The number of delay-sum index steps for discrete convolution. Given a finite length of the room impulse response sequence, For the explosion sound source at a certain point in time The original signal value, For the source of the explosion The signal value after discrete delay. For at a certain point in time Broadband Gaussian noise in the environment For sound source to microphone The interval corresponds to the delay The impulse response coefficient of the step.

[0040] S3, Edge-side AI Intelligent Signal Processing Unit and Very Early Feature Recognition The AI ​​intelligent signal processing unit runs on a BMS edge computing board with NPU computing power, achieving millisecond-level inference through a lightweight deep network. The network architecture and signal processing flow are as follows: Figure 2 As shown.

[0041] S3-1 High-Dimensional Time-Frequency Acoustic Spectrum Mapping Power spectrum obtained by performing FFT operation on digital audio frames To highlight the concentrated energy characteristics of the popping sound in the high-frequency range, the power spectrum was nonlinearly mapped in the frequency domain using a Mel filter bank with M=64 filters. (Linear frequency) With Mel scale The conversion formula is: ; Log-Mel spectrogram was then obtained through logarithmic calculation. ( (where the time frame is the number of frames), this process transforms a one-dimensional chaotic waveform into a two-dimensional image characterizing energy evolution. Microcapsule bursting appears in the image as a highly distinctive, bright, vertical, broadband line, such as... Figure 5 As shown.

[0042] S3-2 AI Intelligent Signal Processing Unit (Lightweight Deep Voiceprint Feature Extraction Network) To compress the parameter count to the hundreds of kilobytes level for compatibility with BMS chips, the network core uses depthwise separable convolutions instead of standard convolutions. The computational cost of a standard convolution is... .in, The spatial side length of the standard convolution kernel. The number of channels in the input feature map. The number of channels in the output feature map. The spatial side length of the input feature map is used. This invention breaks it down into two steps: First, a single-channel depthwise convolution is performed to capture the vertical acoustic signature boundary: ; Then, a 1×1 pointwise convolution is performed to achieve cross-channel feature fusion: ; in, The input feature map in spatial coordinates And the Feature values ​​at each channel For single-channel depthwise convolution kernels in the 1st... Individual channels, local coordinates within the kernel The weighting coefficient at the location, The spatial side length of a single-channel depthwise convolution kernel. and These are the summation indices of the horizontal and vertical coordinates within a single-channel depthwise convolution kernel. and These are the global spatial coordinate indices of the feature map in the horizontal and vertical directions, respectively. The output feature map in spatial coordinates after single-channel depthwise convolution. And the Feature values ​​at each channel The total number of channels in the input feature map. The channel summation index of the input feature map. In a 1×1 pointwise convolution kernel, the connection of the first... The input channel and the first The weighting coefficients of each output channel, This refers to the channel index of the output feature map. The final output feature map, after pointwise convolution, is in spatial coordinates. And the The feature values ​​at each channel.

[0043] The total computational cost is reduced to that of standard convolution. (in and (As defined above) While maintaining feature extraction accuracy above 98%, the inference latency is reduced to less than 5ms. The network ends are flattened by a global average pooling layer, and the probability confidence that a single frame of audio is a popping sound is output by the Sigmoid function. .

[0044] S3-3 Cost-sensitive loss function for handling extreme imbalances: In real-world scenarios, samples with no abnormal background noise account for over 99.9%. To prevent the model from getting trapped in local optima where all predictions are negative, a Focal Loss approach with adaptive weighting for difficult samples is employed. ; in, The value of the Focal Loss function during the model optimization process. For the true class labels of the training samples, This represents a positive sample (with the sound of microcapsules bursting). This represents a negative sample (background noise). As a balance factor, As a focusing factor, by suppressing the weight contribution of massive simple negative samples to the loss function, the model focuses on the identification of weak burst signals with low signal-to-noise ratio.

[0045] When weak bursting signals are difficult to identify ( When the value is relatively small (approaching 0), Approaching 1, maintaining a large gradient backpropagation; when background noise is easily identifiable ( When it approaches 0, Rapid decay forcefully suppresses the meaningless contribution of massive simple negative samples to the loss function, significantly improving the system's sensitivity to very early and weak alarm signals.

[0046] S4, Early Warning Logic Decision-Making and System-Level Multidimensional Closed-Loop Control This step eliminates false alarms using dynamic counting logic and implements a tiered response strategy.

[0047] S4-1 Time density integral discrimination based on sliding window: The sound of stones striking the ground as a car drives over potholes can lead to high-confidence outputs in a single frame. To completely eliminate single-point false alarms, the early warning decision unit introduces a time-sliding window-based approach. The event density integral logic. Define the explosion event density function within the window. : ; in, For the current moment The burst event density function value, The set time sliding window length, For the current system time, For time integration variables, For at any time The probability confidence level of the existence of a thermally induced burst signal in the output. This is an indicator function that takes the value 1 if the condition is met and 0 otherwise. This is the hard-decision threshold for single-frame confidence. It applies only if the density function... Exceeding the preset event density threshold Only then does the system confirm, at the algorithm level, that a real thermal runaway chain reaction has started, and issue a very early physical-level alarm.

[0048] S4-2 Physical Interception and Linkage Control: Once an early warning is triggered, the controller immediately sends a command to the BMS main control board via a high-quality CAN frame.

[0049] (1) Electrical interruption: cut off the high voltage main relay, stop the charging and discharging cycle between the cells, and cut off the continuous Joule heat input caused by the internal short circuit.

[0050] (2) Thermal management level intervention: Instantly increase the power of the liquid cooling water pump of the battery thermal management system (BTMS) to 100%, or open the directional fire sprinkler valve of a specific module. Thanks to this method, which achieves accurate early warning at a very early stage (temperature is only about 95°C, far below the critical point of 180°C~200°C for thermal runaway), the BTMS has a sufficient time window (usually several minutes) to remove the local hot spot, thereby nipping a potential catastrophic explosion accident in the bud at the lowest cost.

[0051] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0052] While the above description illustrates specific embodiments of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for early warning of thermal runaway in lithium batteries based on thermoacoustic materials, characterized in that, Includes the following processes: S1, Construct a thermo-acoustic response layer and cover it onto the surface of a battery cell; the response layer contains thermally expanding microcapsules, which rupture when the battery surface temperature reaches a preset threshold and actively emit pulsed sound wave signals with specific acoustic signature characteristics; the construction of the thermo-acoustic response layer is specifically as follows: The thermo-acoustic response layer comprises a polymer matrix and thermally expandable microcapsules dispersed therein; the microcapsules consist of a thermoplastic polymer shell composed of vinylidene chloride, acrylonitrile and methyl methacrylate copolymerized and a low-boiling-point alkane core material encapsulated inside; trimethylolpropane trimethacrylate is added as a crosslinking agent during the copolymerization process to limit the slippage of polymer chains at high temperatures and ensure that the microcapsules undergo transient brittle fracture upon rupture; Based on the principle of acoustic impedance mismatch, the microcapsules are dispersed in a low-viscosity alicyclic epoxy acrylate matrix resin at a volume fraction of not less than 40% to form a slurry, and then cured on the battery surface by an ultraviolet curing system to form a semi-exposed "island-sea" acoustic structure of microcapsules. S2, through a distributed sound acquisition unit deployed inside the battery module, collects mixed composite sound wave signals inside the battery box in real time and transmits them to the preprocessing module; S3, preprocesses the collected sound signal and constructs its time-frequency features, and inputs it into the AI ​​intelligent signal processing unit based on a lightweight convolutional neural network; the AI ​​intelligent signal processing unit identifies specific acoustic features corresponding to microcapsule rupture from the mixed composite sound wave signal and outputs the confidence level of the thermal anomaly event; S4, the early warning decision unit makes a logical judgment based on the confidence level of the thermal anomaly event. When the number of popping sounds per unit time exceeds the set threshold, it is determined to be an early sign of thermal runaway and sends a linkage control and cut-off command to the battery management system.

2. The method for early warning of thermal runaway in lithium batteries based on thermoacoustic materials as described in claim 1, characterized in that: The transient rupture behavior of the microcapsule at the preset initial sound emission temperature satisfies the following critical sound emission mathematical constraint model: The critical condition for microcapsule rupture is given by the formula Decide; in, The dynamic circumferential stress of the microcapsule shell satisfies the following formula: ； This represents the real-time absolute temperature of the battery surface. Due to external environmental pressures, The thickness of the microcapsule shell, The radius of the microcapsule is... For low-boiling-point alkane core materials at temperature The saturated vapor pressure at this temperature follows the Antoine equation: ； Where A, B, and C are the Antoine constants of isooctane; For the polymer shell at temperature The yield strength at this point follows a linear softening decay model: ； in Reference temperature The initial yield strength is below. This is the temperature softening coefficient of the material.

3. The method for early warning of thermal runaway in lithium batteries based on thermoacoustic materials as described in claim 1, characterized in that: The specific acquisition method of the distributed sound acquisition unit in S2 is as follows: A capacitive MEMS microphone array is selected and arranged on the inner wall of the side plate of the battery module; the physical gap between the battery cells is used as a parallel plate acoustic waveguide channel, and the sound inlet of the microphone faces the outlet of the gap to receive high-frequency pulse sound waves propagating through total internal reflection via the acoustic waveguide channel. Establish a discrete convolutional acoustic channel model based on room impulse response, microphone Acquired mixed discrete signal Expressed as: ； in, For microphone At the point of time The acquired discrete-time mixing signal values, This represents the time point number of the current discrete sampling. The number of delay-sum index steps for discrete convolution. Given a finite length of the room impulse response sequence, For the explosion sound source at a certain point in time The original signal value, For the source of the explosion The signal value after discrete delay. For at a certain point in time Broadband Gaussian noise in the environment For sound source to microphone The interval corresponds to the delay The impulse response coefficient of the step.

4. The method for early warning of thermal runaway in lithium batteries based on thermoacoustic materials as described in claim 1, characterized in that: In step S3, the collected sound signal undergoes preprocessing and time-frequency feature construction, specifically as follows: The digital audio signal in the mixed composite sound wave signal is pre-emphasized and framed; the power spectrum of each audio frame is obtained by performing a fast Fourier transform. And the power spectrum is mapped onto the Mel filter bank, linear frequency With Mel scale The conversion formula is: ； Subsequently, the logarithm of the Mel energy is taken to obtain the Log-Mel spectrogram characterizing the transient energy pulse as a two-dimensional feature map.

5. The method for early warning of thermal runaway in lithium batteries based on thermoacoustic materials as described in claim 4, characterized in that: The specific identification process of the AI ​​intelligent signal processing unit based on the lightweight convolutional neural network in S3 is as follows: The generated Log-Mel spectrogram slices are used as model input; the network backbone extracts pop sound features through a depthwise separable convolutional structure, specifically including single-channel depthwise convolution to extract vertical acoustic boundary, and pointwise convolution to achieve cross-channel feature fusion. The formula for single-channel depthwise convolution is: ; The formula for pointwise convolution is: ; in, The input feature map in spatial coordinates And the Feature values ​​at each channel For single-channel depthwise convolution kernels in the 1st... Individual channels, local coordinates within the kernel The weighting coefficient at the location, The spatial side length of a single-channel depthwise convolution kernel. and These are the indices for the summation of the horizontal and vertical coordinates within a single-channel depthwise convolution kernel. and These are the global spatial coordinate indices of the feature map in the horizontal and vertical directions, respectively. The output feature map in spatial coordinates after single-channel depthwise convolution. And the Feature values ​​at each channel The total number of channels in the input feature map. The channel summation index of the input feature map. In a 1×1 pointwise convolution kernel, the connection of the first... The input channel and the first The weighting coefficients of each output channel, This refers to the channel index of the output feature map. The final output feature map, after pointwise convolution, is in spatial coordinates. And the Feature values ​​at each channel; The extracted high-dimensional acoustic features are transformed into feature vectors through a global average pooling layer, and the sigmoid function of the fully connected layer outputs the probability confidence of the presence of a thermally induced burst signal within the current time window. .

6. The method for early warning of thermal runaway in lithium batteries based on thermoacoustic materials as described in claim 5, characterized in that: The AI ​​intelligent signal processing unit employs an optimization strategy for extreme imbalance between positive and negative samples during the training phase, specifically: The Mixup data augmentation algorithm was used to enhance the purity of the microcapsule burst sound samples. negative samples with background noise According to the random mixing coefficients that follow a Beta distribution Perform linear superposition to generate mixed training samples; The model optimization process employs the Focal Loss loss function, which uses adaptive weighting for difficult samples. ； in, The value of the Focal Loss function during the model optimization process. For the true class labels of the training samples, This represents a positive sample, and the sound of microcapsules bursting is present. This represents a negative sample, which is background noise. As a balance factor, As a focusing factor, by suppressing the weight contribution of massive simple negative samples to the loss function, the model focuses on the identification of weak burst signals with low signal-to-noise ratio.

7. The method for early warning of thermal runaway in lithium batteries based on thermoacoustic materials as described in claim 1, characterized in that: The logical judgment and instruction sending of the early warning decision unit in S4 are specifically as follows: The early warning decision unit introduces a time-sliding window-based approach. The event density integral logic defines the explosion event density function within the window. for: ； in, For the current moment The burst event density function value, The set time sliding window length, For the current system time, For time integration variables, For at any time The probability confidence level of the existence of a thermally induced burst signal in the output. This is an indicator function that takes the value 1 if the condition is met and 0 otherwise. The confidence threshold for a single frame is set as the hard decision threshold. When density function Greater than or equal to the preset event density threshold At that time, it was confirmed as a genuine early sign of thermal runaway; the early warning decision unit then sent a circuit cut-off command to the battery management system and triggered the closed-loop action of the linkage fire-fighting cooling system.

Images