Power battery thermal runaway risk early warning method and system

Abstract

The invention relates to the technical field of new energy automobile and battery safety, and discloses a power battery thermal runaway risk early warning method, which comprises the following steps: synchronously acquiring battery internal resistance dynamic change, characteristic gas concentration and temperature gradient distribution through a multi-physics field coupling sensor network, and constructing an electrochemical-thermal-gas ternary early criterion; the internal resistance mutation rate is calculated through self-adaptive sliding window differential, the gas precipitation rate and the temperature space heterogeneity index are combined and input into a multi-branch fusion neural network containing a gating attention mechanism, the continuous thermal runaway risk probability is output, and corresponding early warning is triggered according to grading threshold values. The system comprises an internal resistance, gas and temperature monitoring module and a multi-modal fusion evaluation and grading early warning execution module. According to the method, the traditional early warning logic singly depending on a temperature threshold value is abandoned, a ternary early criterion system with the internal resistance abrupt change rate, the characteristic gas precipitation rate and the temperature space heterogeneity index as the core is constructed for the first time, and electrochemical and physical field anomalies caused by thermal runaway precursor events such as lithium dendrite piercing of a diaphragm and SEI membrane rupture can be effectively captured.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy vehicles and battery safety, and particularly relates to a power battery thermal runaway risk early warning method and system. BACKGROUND

[0002] With the rapid development of new energy vehicles and large-scale energy storage systems, the safety of power batteries has increasingly become a key bottleneck restricting the progress of the industry. Power batteries may experience thermal runaway under the inducement of overcharging, mechanical damage, or manufacturing defects, and the process has the characteristics of strong burst, fast spread, and huge energy release, which can easily cause fire or even explosion accidents. Therefore, establishing an efficient and reliable risk early warning mechanism is a core link to ensure the safe operation of the battery system. The current mainstream thermal runaway early warning method is mainly based on temperature sensors to monitor the surface temperature of single cells or modules, and sets a fixed threshold as the alarm trigger condition. This strategy relies on the macroscopic temperature rise signal after the thermal runaway has entered the obvious heat release stage.

[0003] The early evolution of power battery thermal runaway involves complex electrochemical microprocesses, including lithium dendrite growth piercing the separator, local micro-short circuit causing internal resistance to suddenly increase, and electrolyte decomposition to produce characteristic gas at abnormal potential. These physical and chemical changes often occur before significant temperature rise, constituting the initial inducement of the thermal runaway chain reaction. However, the existing early warning system generally ignores the perception and fusion analysis of the above early signs, and only uses temperature as a single criterion, making it difficult to capture the subtle but critical abnormal signals in the early stage of thermal runaway.

[0004] Over-reliance on temperature threshold trigger mechanism cannot respond in time due to weak temperature rise in the early stage of thermal runaway, resulting in a high false negative rate; at the same time, the dynamic changes in internal resistance and the emission rules of trace characteristic gas are not effectively included in the risk assessment model, making the early warning system lack sensitivity to electrochemical abnormalities.

[0005] Especially in the context of the widespread application of high-energy density batteries, the starting point of thermal runaway is more concealed and the development is more rapid, and the lag and one-sidedness of traditional methods further amplify the safety risks. SUMMARY

[0006] To solve the above technical problems, embodiments of the present application provide a power battery thermal runaway risk early warning method and system, which synchronously collects the internal resistance dynamic change, electrolyte decomposition characteristic gas concentration and local temperature gradient distribution data of the battery monomer through a multi-physical field coupling sensing network, constructs an early abnormal criterion system based on electrochemical-thermal-gas three-element coupling characteristics; the internal resistance mutation rate is calculated in real time by using an adaptive sliding window, combined with the gas precipitation rate and the temperature spatial heterogeneity index, input into a pre-trained multi-branch fusion neural network model for non-linear correlation analysis; the model dynamically weights the contribution of each physical quantity in different thermal runaway evolution stages through a gated attention mechanism, outputs a continuous thermal runaway risk probability value, and triggers a corresponding level of early warning response according to a preset grading threshold, thereby realizing high sensitivity and low false alarm rate risk identification in the early stage of thermal runaway.

[0007] In one aspect of the present application, a power battery thermal runaway risk early warning method is provided, which comprises: A four-wire micro-ohmmeter module embedded in the battery monomer or close to the pole ear position is used to obtain battery DC internal resistance sequence data in real time at a sampling frequency of no less than 10 times per second; A micro gas sensor array arranged in the sealed cavity of the battery shell is used to monitor the concentration of characteristic gas components generated by electrolyte decomposition in real time, the characteristic gas components including carbon monoxide, hydrogen, methane and hydrogen fluoride; A distributed optical fiber temperature measurement unit or an infrared thermal imaging array is used to obtain temperature data of multiple preset measurement points on the surface or inside the battery, forming a spatial temperature distribution matrix; The DC internal resistance sequence data is subjected to adaptive sliding window differential processing, and the internal resistance change rate per unit time is calculated, and when the internal resistance change rate exceeds a first preset threshold, it is marked as an internal resistance abnormal event; After the characteristic gas component concentration data is normalized, the time derivative of each gas concentration is calculated to obtain a gas precipitation rate vector; Based on the spatial temperature distribution matrix, the ratio of the maximum temperature difference to the average temperature of the battery surface is calculated as a temperature spatial heterogeneity index; The internal resistance change rate, gas precipitation rate vector and temperature spatial heterogeneity index are jointly input into a pre-trained multi-branch fusion neural network model, the multi-branch fusion neural network model comprising three parallel input branches corresponding to the internal resistance feature channel, gas feature channel and temperature feature channel, respectively, and each channel end is dynamically fused through a learnable gating weight vector to output a thermal runaway risk probability value; When the probability value of thermal runaway risk is greater than the second preset threshold and less than the third preset threshold, a first-level warning signal is triggered, prompting a power reduction operation in the charging strategy; when the probability value of thermal runaway risk is greater than or equal to the third preset threshold, a second-level warning signal is triggered, initiating the battery isolation and active cooling program.

[0008] This invention provides a power battery thermal runaway risk early warning system, which includes: The internal resistance real-time monitoring module is used to acquire the battery DC internal resistance sequence data in real time at a sampling frequency of no less than 10 times per second by a four-wire micro-ohmmeter module embedded inside the battery cell or close to the electrode tab. The characteristic gas concentration monitoring module is used to monitor the concentration of characteristic gas components generated by electrolyte decomposition in real time through a miniature gas sensor array set in the sealed cavity of the battery casing. The characteristic gas components include carbon monoxide, hydrogen, methane and hydrogen fluoride. The temperature field distribution sensing module is used to acquire temperature data of multiple preset measuring points on or inside the battery surface through a distributed fiber optic temperature measurement unit or an infrared thermal imaging array, forming a spatial temperature distribution matrix. The internal resistance mutation rate calculation module is used to perform adaptive sliding window differentiation processing on the DC internal resistance sequence data, calculate the internal resistance change rate per unit time, and mark an internal resistance abnormal event when the internal resistance change rate exceeds a first preset threshold. The gas evolution rate calculation module is used to normalize the concentration data of the characteristic gas components, calculate the time derivative of each gas concentration, and obtain the gas evolution rate vector. The temperature spatial heterogeneity index calculation module is used to calculate the ratio of the maximum temperature difference to the average temperature on the battery surface based on the spatial temperature distribution matrix, as the temperature spatial heterogeneity index. The multimodal feature fusion risk assessment module is used to input the internal resistance change rate, gas evolution rate vector and temperature spatial heterogeneity index into a pre-trained multi-branch fusion neural network model. The multi-branch fusion neural network model contains three parallel input branches, corresponding to the internal resistance feature channel, gas feature channel and temperature feature channel, respectively. The ends of each channel are dynamically fused through a learnable gating weight vector to output the probability value of thermal runaway risk. The graded early warning execution module is used to trigger a first-level early warning signal when the probability value of thermal runaway risk is greater than a second preset threshold and less than a third preset threshold, prompting a power reduction operation of the charging strategy; and to trigger a second-level early warning signal when the probability value of thermal runaway risk is greater than or equal to the third preset threshold, initiating the battery isolation and active cooling program.

[0009] In one embodiment of the present invention, the four-wire micro-ohmmeter module consists of a constant current source, a voltage sampling circuit, an analog switch array, and an analog-to-digital converter. The constant current source outputs a current amplitude of 5 amperes. The voltage sampling circuit adopts a differential amplification structure with a common-mode rejection ratio of not less than 120dB. The analog-to-digital converter has a resolution of 24 bits and a sampling rate of 1000 times per second. After the original voltage signal is processed by moving average through a digital filter, the DC internal resistance is calculated in combination with the known constant current value.

[0010] In one embodiment of the present invention, the miniature gas sensor array consists of four independent metal oxide semiconductor gas sensing units, which respectively provide specific responses to carbon monoxide, hydrogen, methane and hydrogen fluoride. The operating temperature of each sensing unit is precisely controlled at 350 degrees Celsius by a miniature heater. The resistance change signal is converted into a voltage signal by a transimpedance amplifier and then acquired by a 16-bit analog-to-digital converter with a sampling period of 500 milliseconds.

[0011] In one embodiment of the present invention, the distributed optical fiber temperature measurement unit adopts the Raman scattering principle. The optical fiber is laid in a serpentine path along the length of the battery cell between the electrodes, with a spatial resolution of 5 cm, a temperature measurement accuracy of ±0.5 degrees Celsius, and a sampling frequency of twice per second. The infrared thermal imaging array is composed of a 64-pixel × 64-pixel uncooled microbolometer focal plane array, with a field of view covering the entire surface of the battery cell, a frame rate of 10 frames per second, and a temperature resolution of 0.1 degrees Celsius.

[0012] As one embodiment of the present invention, the specific steps of the adaptive sliding window differential processing are as follows: The initial window width is set to 10 milliseconds; when the standard deviation of the internal resistance values ​​of 5 consecutive sampling points is less than 0.5 milliohms, the window width is expanded to 50 milliseconds; when the standard deviation is greater than or equal to 0.5 milliohms, the window width is shrunk to 5 milliseconds; the rate of change of internal resistance is calculated within the window using the central difference formula, the formula being... ,in The sampling interval is denoted as .

[0013] In one embodiment of the present invention, the internal resistance feature channel of the multi-branch fusion neural network model consists of three one-dimensional convolutional layers with kernel sizes of 5, 3, and 3 respectively, followed by batch normalization and modified linear unit activation functions after each layer; the gas feature channel consists of two fully connected layers with 64 and 32 hidden units respectively; the temperature feature channel consists of a two-dimensional convolutional layer and a global average pooling layer; the output vectors of the three channels are concatenated and input to the gated fusion layer, the weight vector of the gated fusion layer is generated by two independent fully connected networks, its input is the concatenated vector of the outputs of the three channels, and the output is normalized by the softmax function and used as the fusion weight of each channel.

[0014] As one embodiment of the present invention, the training process of the multi-branch fusion neural network model includes: collecting a large amount of battery internal resistance, gas and temperature time-series data under abuse conditions such as accelerated aging, overcharging, needle penetration and squeezing; labeling different time windows before thermal runaway as high-risk samples and normal operating condition data as low-risk samples; using the cross-entropy loss function for end-to-end training, with Adam as the optimizer, an initial learning rate of 0.001, a batch size of 32, and 200 training rounds; after training, the model has a false negative rate of less than 0.5% and a false positive rate of less than 2% on the validation set.

[0015] In one embodiment of the present invention, the first preset threshold is 0.8 milliohms per second, the second preset threshold is 70%, and the third preset threshold is 90%; the first-level warning signal is sent to the main control unit of the battery management system through the controller local area network bus, instructing it to limit the maximum charging current to 50% of the rated value; the second-level warning signal triggers the relay to disconnect the high-voltage circuit of the battery pack and start the cooling pump and cooling fan in the liquid cooling plate.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention abandons the traditional early warning logic that relies solely on temperature thresholds and for the first time constructs a ternary early warning criterion system with internal resistance mutation rate, characteristic gas evolution rate and temperature spatial heterogeneity index as the core. It can effectively capture electrochemical and physical field anomalies caused by thermal runaway precursor events such as lithium dendrite piercing the separator and SEI film rupture. 2. The instantaneous change of internal resistance is accurately quantified by the adaptive sliding window differential algorithm, avoiding insufficient sensitivity or noise interference caused by the fixed window; the multi-branch fusion neural network model introduces a gated attention mechanism to realize the dynamic weight allocation of different physical quantities in different stages of thermal runaway, which significantly improves the model's ability to identify weak early signals. 3. In actual testing, the system can issue reliable early warnings 15 to 30 minutes in advance, reducing the false alarm rate to below 0.5%, providing sufficient time window for battery safety protection and effectively ensuring the operational safety of electric vehicles and energy storage systems. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of a power battery thermal runaway risk early warning method and system proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-branch fusion neural network model based on the electrochemical-thermal-gas ternary coupling characteristics in this invention; Figure 3 This is a logical flowchart of the multi-physics field coupled sensing data acquisition and preprocessing stage in this invention. Figure 4This is a flowchart illustrating the logical flow of the adaptive sliding window differential calculation and feature extraction stages in this invention. Figure 5 This is a flowchart illustrating the logical flow of the dynamic fusion of internal resistance, gas, and temperature three-modal characteristics and the risk probability output stage in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal perception layer, the edge computing layer and the battery management system in this invention. Detailed Implementation

[0018] Please refer to Figures 1 to 6 This invention provides a method and system for early warning of thermal runaway risk in power batteries. Its core lies in constructing an early anomaly criterion system based on the ternary coupling characteristics of electrochemistry, heat, and gas. Through multi-physics field synchronous sensing, adaptive differential feature extraction, and multi-modal neural network fusion analysis, it achieves highly sensitive identification of precursory thermal runaway events. The specific embodiments of this invention are described in detail below with reference to the accompanying drawings.

[0019] The method for early warning of thermal runaway risk in power batteries includes the following steps: S1, through a four-wire micro-ohmmeter module embedded inside the battery cell or close to the electrode, acquires the battery DC internal resistance sequence data in real time at a sampling frequency of no less than 10 times per second. S2, by using a miniature gas sensor array set in the sealed cavity of the battery casing, the concentration of characteristic gas components generated by the decomposition of the electrolyte is monitored in real time. The characteristic gas components include carbon monoxide, hydrogen, methane and hydrogen fluoride. S3 acquires temperature data from multiple preset measurement points on or inside the battery surface through a distributed fiber optic temperature measurement unit or an infrared thermal imaging array, forming a spatial temperature distribution matrix. S4, perform adaptive sliding window differentiation processing on the DC internal resistance sequence data, calculate the rate of change of internal resistance per unit time, and mark it as an internal resistance abnormal event when the rate of change of internal resistance exceeds a first preset threshold. S5. After normalizing the concentration data of the characteristic gas components, calculate the time derivative of each gas concentration to obtain the gas evolution rate vector. S6. Based on the spatial temperature distribution matrix, calculate the ratio of the maximum temperature difference to the average temperature on the battery surface, as the temperature spatial heterogeneity index. S7, The internal resistance change rate, gas evolution rate vector and temperature spatial heterogeneity index are input together into the pre-trained multi-branch fusion neural network model. The multi-branch fusion neural network model contains 3 parallel input branches, which correspond to the internal resistance feature channel, gas feature channel and temperature feature channel respectively. The ends of each channel are dynamically fused through a learnable gating weight vector to output the probability value of thermal runaway risk. S8. When the probability value of thermal runaway risk is greater than the second preset threshold and less than the third preset threshold, a first-level warning signal is triggered to prompt the charging strategy to reduce power. When the probability value of thermal runaway risk is greater than or equal to the third preset threshold, a second-level warning signal is triggered to start the battery isolation and active cooling program.

[0020] In step S1, the four-wire micro-ohmmeter module consists of a constant current source, a voltage sampling circuit, an analog switch array, and a 24-bit analog-to-digital converter. The constant current source outputs a fixed current amplitude of 5 amperes, which is sufficient to generate a measurable voltage drop without causing significant Joule heating. The voltage sampling circuit employs a differential amplifier structure with a high common-mode rejection ratio (CMRR) of not less than 120 dB to effectively suppress high-frequency electromagnetic interference generated during battery operation. The analog switch array is used to quickly switch measurement paths between multiple battery cells, with a switching period of no more than 10 milliseconds.

[0021] The analog-to-digital converter has a sampling rate of 1000 times per second. The raw voltage signal is processed by a digital moving average filter with a window length of 10 sampling points to suppress high-frequency noise. DC internal resistance... The calculation formula is: ,in This is the filtered voltage measurement value. The constant current source outputs a current of 5 amps. This internal resistance sequence is updated at a frequency of no less than 10 times per second and cached in the circular buffer of the edge computing unit with a buffer depth of 1000 sampling points, corresponding to 100 seconds of historical data.

[0022] In step S2, the miniature gas sensor array consists of four independent metal-oxide-semiconductor gas sensing units integrated on the same ceramic substrate. Each sensing unit is coated with a catalytic material that selectively adsorbs specific gases. The first sensing unit responds to carbon monoxide, with a sensitive layer of platinum-doped tin oxide; the second sensing unit responds to hydrogen, with a sensitive layer of palladium-modified zinc oxide; the third sensing unit responds to methane, with a sensitive layer of nickel-doped indium oxide; and the fourth sensing unit responds to hydrogen fluoride, with a sensitive layer of lanthanide-modified tungsten oxide.

[0023] Each sensing unit integrates a micro-heater at its bottom, and its operating temperature is precisely maintained at 350 degrees Celsius by a closed-loop temperature controller to ensure optimal response sensitivity and recovery speed. The resistance change of the sensing unit is converted into a 0-5V voltage signal by a transimpedance amplifier, and then acquired by a 16-bit analog-to-digital converter at a 500ms cycle. The raw concentration data undergoes temperature compensation and cross-sensitivity correction algorithms, ultimately outputting the volumetric concentration values ​​of the four characteristic gases in ppm. All concentration data is normalized before being input into subsequent processing modules; the normalization formula is as follows: ,in and These are the minimum and maximum historical concentrations of the gas under normal operating conditions. This represents the normalized gas concentration value. This is the original, corrected gas concentration value.

[0024] In step S3, the distributed fiber optic temperature measurement unit operates based on the Raman scattering principle. The sensing fiber is laid in a serpentine path along the length of the battery cell within the separator layer between the positive and negative electrodes. The fiber cladding is made of polyimide material resistant to electrolyte corrosion. A laser pulse is injected into the fiber with a width of 10 nanoseconds. The backscattered light is separated into Stokes light and anti-Stokes light by a wavelength division multiplexer. The intensity ratio of the two is monotonically related to the local temperature. The system has a spatial resolution of 5 centimeters, meaning that each 5-centimeter fiber segment corresponds to one temperature measurement point. Typically, 20 to 40 measurement points are arranged on the entire battery cell.

[0025] The temperature measurement accuracy is ±0.5 degrees Celsius, with a sampling frequency of twice per second. An alternative solution uses an infrared thermal imaging array for non-contact temperature measurement. This array consists of a 64-pixel × 64-pixel uncooled microbolometer focal plane, encapsulated behind the observation window on top of the battery module, covering the entire surface of the unit. Infrared images are output at a rate of 10 frames per second, with each frame containing 4096 temperature pixels, resulting in a temperature resolution of 0.1 degrees Celsius. Regardless of the measurement method used, an N-dimensional temperature vector is ultimately generated. ,in This vector, representing the effective number of measurement points, forms the basis of the spatial temperature distribution matrix.

[0026] In step S4, the DC internal resistance sequence undergoes adaptive sliding window differentiation. The initial window width is set to 10 milliseconds, corresponding to the sampling interval. The system calculates the standard deviation of the internal resistance values ​​of 5 consecutive sampling points within the current window in real time. .like If the value is less than 0.5 milliohms, the system is considered to be in a stable state, and the window width is expanded to 50 milliseconds to improve the noise resistance of the differential calculation; if... If the resistance is ≥ 0.5 milliohms, a potential abrupt change is identified, and the window width is reduced to 5 milliseconds to improve temporal resolution. Within the defined window, the rate of change of internal resistance is calculated using the central difference formula. Its expression is: ; in The sampling interval is in seconds. and These are the adjacent internal resistance samples before and after the current time. The rate of change is output once per second. When the first preset threshold (set to 0.8 milliohms per second) is reached, the system marks the internal resistance abnormality event and records the time of occurrence, peak change rate and duration of the event.

[0027] In step S5, the time derivative is calculated for the normalized characteristic gas concentration data. Let... For the first The gas at time Given the normalized concentration, its precipitation rate is: ; in The gas sampling period is 500 milliseconds. This yields a 4D gas evolution rate vector. This vector directly reflects the instantaneous intensity of the electrolyte decomposition reaction and is a key indicator for judging SEI film rupture or internal short circuit.

[0028] In step S6, the spatial heterogeneity index of temperature is calculated based on the spatial temperature distribution matrix. Let... This represents the highest temperature among all measuring points at the current moment. Lowest temperature, The maximum temperature difference is the arithmetic mean of the temperatures at all measuring points. Temperature spatial heterogeneity index Defined as This index quantifies the unevenness of heat distribution inside the battery; the presence of localized overheating areas will lead to... Significantly higher, even though the overall average temperature did not reach dangerous levels.

[0029] In step S7, the rate of change of internal resistance is... Gas evolution rate vector and temperature spatial heterogeneity index The common inputs are fed into a pre-trained multi-branch fusion neural network model.

[0030] The model contains three parallel input branches: the internal resistance characteristic channel receives a scalar. After processing through three one-dimensional convolutional layers with kernel sizes of 5, 3, and 3 respectively, and each layer having 32 output channels, batch normalization and modified linear unit activation functions are applied, ultimately outputting a 32-dimensional feature vector; the gas feature channel receives a 4-dimensional vector. After processing through two fully connected layers, the number of hidden units is 64 and 32 respectively, outputting a 32-dimensional feature vector; the temperature feature channel receives a scalar. First, it is expanded into a two-dimensional tensor, processed by a two-dimensional convolutional layer with a 3×3 convolutional kernel, and then compressed into a 32-dimensional feature vector through a global average pooling layer.

[0031] The output vectors of the three channels are concatenated to form a 96-dimensional joint feature vector. This vector is simultaneously input to both the main decoder and the gated weight generator. The gated weight generator consists of two fully connected layers with 64 hidden units and 3 output units. Its output is normalized using the softmax function to obtain the fused weights for the three channels. ,satisfy Final fusion features This fused feature is input to the final fully connected layer, outputting a single thermal runaway risk probability value. .

[0032] In step S8, according to The value triggers a tiered warning. The second preset threshold is set at 70%, and the third preset threshold is set at 90%. > 0.7 and When the current is less than 0.9, the system sends a Level 1 warning signal to the main control unit of the battery management system via the controller area network bus, instructing it to limit the maximum charging current to 50% of the rated value and disable fast charging mode. When the value is ≥ 0.9, the system immediately triggers a level 2 warning signal, controls the high-voltage relay to disconnect the main circuit of the battery pack, and simultaneously starts the cooling pump and cooling fan of the liquid cooling system to execute a forced cooling procedure. All warning events are recorded in non-volatile memory, including timestamps, each input feature value, and the final risk probability.

[0033] The power battery thermal runaway risk early warning system includes a real-time internal resistance monitoring module, a characteristic gas concentration monitoring module, a temperature field distribution sensing module, an internal resistance mutation rate calculation module, a gas evolution rate calculation module, a temperature spatial heterogeneity index calculation module, a multi-modal feature fusion risk assessment module, and a graded early warning execution module. The real-time internal resistance monitoring module is integrated at the battery cell tab, and its four-wire micro-ohmmeter hardware circuit communicates with the battery management system main control board via a high-speed serial interface. The characteristic gas concentration monitoring module's miniature gas sensor array is sealed and installed in a dedicated chamber on the top of the battery casing. The chamber is connected to the battery interior through a microporous membrane, allowing gas molecule diffusion but preventing electrolyte leakage. The temperature field distribution sensing module selects a distributed fiber optic or infrared thermal imaging scheme based on the battery packaging form, and its output data is sent to the edge computing unit via a dedicated interface. The internal resistance mutation rate calculation module, the gas evolution rate calculation module, and the temperature spatial heterogeneity index calculation module are all deployed as software function blocks in the real-time operating system of the edge computing unit, with an execution cycle of 100 milliseconds. The multimodal feature fusion risk assessment module runs on the neural network inference engine of the edge computing unit. The model is stored in TensorRT format, with an inference latency of less than 10 milliseconds. The graded warning execution module includes a digital output drive circuit that directly controls the power switches of the relays and the cooling pump, and interacts with the vehicle control system via the controller area network bus.

[0034] This embodiment achieves early and highly sensitive early warning of thermal runaway risk in power batteries through the aforementioned method and system. Actual measurements show that in a needle penetration test, the system can detect a sudden increase in internal resistance and hydrogen evolution 15 minutes before the temperature rises, with a risk probability value reaching over 80% 30 minutes in advance and a false negative rate of less than 0.5%, significantly outperforming the traditional temperature threshold method.

Claims

1. A method for power battery thermal runaway risk early warning, characterized in that, The application relates to a battery thermal runaway risk prediction method based on multi-branch fusion neural network model. The method comprises the following steps: embedding a four-wire micro-ohmmeter module inside the battery monomer or close to the position of the pole ear; monitoring the concentration of characteristic gas components generated by electrolyte decomposition in real time through a micro gas sensor array arranged in the sealed cavity of the battery shell; obtaining temperature data of multiple preset measuring points on the surface or inside of the battery through a distributed optical fiber temperature measurement unit or an infrared thermal imaging array, and forming a spatial temperature distribution matrix; performing adaptive sliding window differential processing on the direct current internal resistance sequence data to calculate the internal resistance change rate per unit time, and marking an internal resistance abnormal event when the internal resistance change rate exceeds a first preset threshold value; calculating the time derivative of each gas concentration to obtain a gas evolution rate vector after normalizing the characteristic gas component concentration data; calculating the ratio of the maximum temperature difference to the average temperature on the surface of the battery based on the spatial temperature distribution matrix as a temperature spatial heterogeneity index; inputting the internal resistance change rate, the gas evolution rate vector and the temperature spatial heterogeneity index into a pre-trained multi-branch fusion neural network model; triggering a first warning signal when the thermal runaway risk probability value is greater than a second preset threshold value and less than a third preset threshold value, and prompting a charging strategy power reduction operation; 2. The method of claim 1, wherein, triggering a second warning signal when the thermal runaway risk probability value is greater than or equal to the third preset threshold value, and starting a battery isolation and active cooling program.

3. The method of claim 2, wherein the step of determining the thermal runaway risk level of the power battery comprises: The multi-branch fusion neural network model comprises three parallel input branches corresponding to internal resistance feature channels, gas feature channels and temperature feature channels, and each channel end is dynamically fused through a learnable gating weight vector to output a thermal runaway risk probability value. The adaptive sliding window differential processing on the direct current internal resistance sequence data to calculate the internal resistance change rate per unit time comprises the following steps: setting an initial window width of 10 milliseconds to calculate the standard deviation of the internal resistance values of the continuous 5 sampling points in the current window in real time; extending the window width to 50 milliseconds when the standard deviation is less than 0.5 milliohm, and contracting the window width to 5 milliseconds when the standard deviation is greater than or equal to 0.5 milliohm; 4.The power battery thermal runaway risk pre-warning method of claim 3, wherein, calculating the internal resistance change rate in the determined window by using a central difference formula. The method for calculating the time derivative of each gas concentration to obtain a gas evolution rate vector after normalizing the characteristic gas component concentration data comprises the following steps: The corrected concentration value is calculated according to the formula wherein and are the minimum and maximum historical concentration values of the gas under normal working conditions, respectively, is the normalized gas concentration value, is the original corrected gas concentration value; performing temperature compensation and cross-sensitivity correction on the original concentration data of each characteristic gas; 5. The power battery thermal runaway risk early warning method according to claim 4, characterized in that, calculating the time derivative of each gas concentration based on the normalized concentration value and a gas sampling period of 500 milliseconds to form a 4-dimensional gas evolution rate vector. extracting all the temperature values of the measuring points from the spatial temperature distribution matrix, determining the highest temperature , the lowest temperature and the average temperature ; calculating the maximum temperature difference ; Temperature spatial heterogeneity index defined as .

6. The power battery thermal runaway risk early warning method according to claim 5, characterized in that, The method for calculating the ratio of the maximum temperature difference to the average temperature on the surface of the battery based on the spatial temperature distribution matrix as a temperature spatial heterogeneity index comprises the following steps: The method for inputting the internal resistance change rate, the gas evolution rate vector and the temperature spatial heterogeneity index into a pre-trained multi-branch fusion neural network model comprises the following steps: inputting the internal resistance change rate into the internal resistance feature channel, processing through three one-dimensional convolution layers, and sequentially setting the convolution kernel size to be 5, 3 and 3, and connecting batch normalization and a rectified linear unit activation function after each layer to output a 32-dimensional internal resistance feature vector; The gas evolution rate vector is input into the gas feature channel, processed by two fully connected layers with 64 and 32 hidden units respectively, and a 32-dimensional gas feature vector is output; The temperature spatial heterogeneity index is input into the temperature feature channel, first expanded into a two-dimensional tensor, then processed by a two-dimensional convolution layer with a 3*3 kernel, and finally compressed into a 32-dimensional temperature feature vector by a global average pooling layer; The feature vectors output by the three channels are spliced into a 96-dimensional joint feature vector.

7. The power battery thermal runaway risk early warning method according to claim 6, characterized in that, The end of each channel is dynamically fused by a learnable gating weight vector, and a thermal runaway risk probability value is output, including: The 96-dimensional joint feature vector is input into the gating weight generator, which is composed of two fully connected networks, and the number of units in the output layer is 3. After normalization by the softmax function, the internal resistance channel weight, the gas channel weight and the temperature channel weight are obtained; Based on the three channel weights, the three 32-dimensional feature vectors are weighted and summed to obtain a fused feature vector; The fused feature vector is input into the output fully connected layer to generate a thermal runaway risk probability value.

8. The power battery thermal runaway risk early warning method according to claim 7, characterized in that, The training process of the multi-branch fusion neural network model includes: Collecting internal resistance, gas and temperature time series data of the battery under accelerated aging, overcharging, needle puncture and extrusion conditions, labeling different time windows before thermal runaway as high-risk samples, and normal working condition data as low-risk samples; An end-to-end training is performed using a cross-entropy loss function, the optimizer is Adam, the initial learning rate is 0.001, the batch size is 32, and the number of training rounds is 200.

9. The power battery thermal runaway risk early warning method according to claim 8, characterized in that, The first preset threshold is 0.8 milliohm per second, the second preset threshold is 70%, and the third preset threshold is 90%; The first-level warning signal is sent to the battery management system master control unit through the controller area network bus, instructing it to limit the maximum charging current to 50% of the rated value; The second-level warning signal triggers a relay to disconnect the high-voltage loop of the battery pack and starts the cooling pump and cooling fan in the liquid cooling plate.

10. The method of claim 1, wherein, The micro gas sensor array is composed of four independent metal oxide semiconductor gas sensing units, which respond specifically to carbon monoxide, hydrogen, methane and hydrogen fluoride respectively; The working temperature of each sensing unit is accurately controlled at 350 degrees Celsius by a micro heater; The resistance change signal is converted into a voltage signal by a transimpedance amplifier and then collected by a 16-bit analog-to-digital converter at a 500 millisecond cycle.

Images