Multi-component gas early warning system based on sensing, storage and calculation integration and early warning method thereof

Abstract

The invention discloses a multi-component gas early warning system based on sensing, storage and calculation integration and an early warning method thereof, the system comprises the following modules: a sensor array module configured to measure different gas concentrations in a battery compartment by using a sensor array capable of identifying different gases; the data storage processing module is configured to process, cache and store the gas data collected by the sensor; the analysis and prediction module is configured to perform gas concentration and type analysis on the preprocessed gas data according to a deep learning algorithm, and predict the change of the gas concentration; and the early warning module is configured to send an early warning signal to the battery and take safety measures according to the predicted gas concentration change. According to the invention, the six characteristic gases and the characteristic proportion thereof in the thermal runaway process of the battery can be accurately detected through the multi-component gas sensor array, the response is quick, and the false alarm rate is lower than 5%. The potential thermal runaway risk can be quickly identified, and early warning can be performed in time.

Description

Technical Field

[0001] This invention relates to the field of gas identification technology, and in particular to a multi-component gas early warning system and method based on integrated sensing, storage, and computing. Background Technology

[0002] With the widespread application of new energy vehicles and energy storage systems, the safety issues of lithium-ion batteries have become increasingly prominent, especially the frequent fires and explosions caused by battery thermal runaway. Existing technologies mainly focus on the detection of single gases using gas sensors, and most fail to consider the characteristics of multiple gases involved in the thermal runaway process.

[0003] For example:

[0004] Patent No. CN102607739A (Gas Sensor and Manufacturing Method Thereof): This patent relates to a metal oxide semiconductor gas sensor, which is suitable for detecting gas concentration. However, due to its single gas detection and slow response time, it cannot meet the rapid early warning requirements such as battery thermal runaway.

[0005] Patent No. CN117571647A (Multi-gas sensor and its usage method): This patent introduces the basic design of a multi-gas sensor, but it is not specifically optimized for the gas characteristics (such as carbon monoxide, hydrogen, etc.) of thermal runaway processes, and it lacks the identification of gas ratio relationships.

[0006] The main drawbacks of existing technologies:

[0007] 1. Slow response speed: Most gas sensors have a response time of more than 10 seconds, which cannot meet the needs of early warning in the early stage of battery thermal runaway.

[0008] 2. Single gas detection: Traditional gas sensors can usually only detect a single gas and cannot comprehensively analyze the relationship between multiple gases, which leads to the inability to accurately identify the occurrence of thermal runaway.

[0009] 3. High false alarm rate: Existing sensors are easily affected by environmental interference, resulting in a high false alarm rate and affecting the reliability of the system.

[0010] 4. The gas ratio characteristics are not fully utilized: During battery thermal runaway, carbon monoxide, hydrogen and hydrocarbon gases (methane, ethylene, acetylene, propylene) have specific ratio relationships, and existing technologies have failed to make comprehensive judgments based on these ratio relationships. Summary of the Invention

[0011] In view of the shortcomings of the prior art, the technical problem to be solved by the present invention is: how to...

[0012] To address the aforementioned technical problems, this invention designs a multi-component gas early warning system based on integrated sensing, storage, and computing. This system can accurately detect six characteristic gases (CO, hydrogen, methane, ethylene, acetylene, and propylene) and their characteristic proportions during battery thermal runaway through a multi-component gas sensor array, responding within 6 seconds with a false alarm rate of less than 5%. It can quickly identify potential thermal runaway risks and provide timely warnings.

[0013] One technical solution adopted by this invention is to provide a multi-component gas early warning system based on sensing, storage, and computing integration, which includes the following modules:

[0014] The sensor array module is configured to measure the concentration of different gases in the battery compartment using a sensor array capable of identifying different gases.

[0015] The data storage and processing module is configured to process, cache, and store the gas data collected by the sensor.

[0016] The analysis and prediction module is configured to analyze the gas concentration and type of the preprocessed gas data based on a deep learning algorithm, and to predict changes in gas concentration.

[0017] The early warning module is configured to send an early warning signal to the battery and take safety measures based on predicted changes in gas concentration.

[0018] Furthermore, the data storage processing module includes:

[0019] The preprocessing submodule is configured to perform Kalman filtering and drift correction on the acquired gas data;

[0020] The temperature compensation model construction submodule is configured to construct a temperature compensation model that relates the temperature drift and temperature change of the sensor resistance to the temperature compensation coefficient.

[0021] The linear fitting submodule is configured to fit the temperature compensation model based on the collected temperature drift and actual temperature change values ​​of the sensor resistance, and calculate the temperature compensation coefficient.

[0022] The temperature compensation submodule is configured to map the temperature compensation coefficient into the temperature compensation model and compensate for all collected gas data.

[0023] Furthermore, the temperature compensation model construction submodule includes:

[0024] The temperature drift calculation unit is configured to set different temperature points in the constant temperature chamber, maintain a constant temperature at each temperature point for a preset constant temperature time, measure the average value of multiple original resistance values ​​of the sensor at each temperature point at a fixed frequency, and calculate the temperature drift of the sensor resistance at each temperature point relative to the reference temperature point.

[0025] The third-order temperature compensation model building unit is configured to construct a temperature compensation model using a third-order polynomial:

[0026] R {comp} =R {raw} -[α·(TT {0} )+β·(TT {0} ) 2 +γ·(TT {0} ) 3 ];

[0027] Among them, R {comp} R represents the sensor resistance after temperature compensation. {raw} The original resistance of the sensor before temperature compensation is represented by α, β, and γ, which represent the temperature compensation coefficients, and T represents the ambient temperature. {0} Indicates the reference temperature.

[0028] Furthermore, the linear fitting submodule includes:

[0029] The fitting equation construction unit is configured to construct a fitting equation based on the acquired temperature drift and actual temperature change values ​​of the sensor resistance:

[0030]

[0031] Wherein, ΔR(T) {i} ) represents the temperature drift of the sensor resistance at the i-th temperature point, x {i} =T {i} -T {0} This represents the temperature difference between the i-th temperature point and the reference temperature.

[0032] Furthermore, the linear fitting submodule also includes:

[0033] The error verification unit is configured to verify the effectiveness of the temperature compensation model using the root mean square error.

[0034]

[0035] Where M represents the total number of data points, R j R represents the j-th measured value of the sensor resistance. refj This indicates the reference value corresponding to the measured value.

[0036] Furthermore, the analysis and prediction module includes:

[0037] The dataset construction submodule is configured to build training and validation sets based on simulated gas data collected by sensors in a simulated battery thermal runaway environment, and to standardize the training and validation sets.

[0038] The prediction model building submodule is configured as a deep separable convolutional neural network structure and utilizes the SE attention module to perform channel enhancement on the convolutional output;

[0039] The prediction model training submodule is configured to input the training set into the prediction model and minimize the sparse cross-entropy loss through the Adam optimizer to obtain the trained prediction model.

[0040] Furthermore, the dataset construction submodule includes:

[0041] The dataset partitioning unit is configured to divide the collected simulated gas data into 8 channels and 180 time steps per sample, and to divide the training set and validation set in a ratio of 85%:15%, and to maintain the consistency of the distribution of various types of samples in the training set and validation set by using stratify=y;

[0042] The normalization unit is configured to perform z-score normalization on the collected dataset:

[0043]

[0044] in, μ represents the data point of channel c at time t after normalization of the original sensor sequence. c σ c represents the mean and standard deviation of the training set on the c channel, respectively.

[0045] Furthermore, the prediction model construction submodule includes:

[0046] The depthwise convolutional unit is configured to perform temporal convolutions within each channel using one-dimensional convolutional kernels to extract temporal features.

[0047]

[0048] Among them, w c,i This represents the i-th convolution weight in channel c;

[0049] Pointwise convolutional units are configured to linearly combine the outputs of depthwise convolutions along the channel direction:

[0050]

[0051] Among them, v j,c This represents the 1×1 convolution weight between output channel j and input channel c;

[0052] The attention calibration unit is configured to recalibrate the channel weights using the SE attention module:

[0053] z' c,t=σ(W2ReLU(W1s))·z c,t

[0054] Where W1 and W2 represent the weight matrices of the fully connected layer, s represents the channel descriptor, and σ represents the sigmoid activation function;

[0055] The classification unit is configured to calculate class probabilities using Softmax:

[0056]

[0057] Among them, w k b k Let p represent the weight vector and bias of the k-th class, respectively, where K represents the number of classes. k ∈[0,1] and

[0058] Furthermore, the prediction model training submodule includes:

[0059] The loss function building unit is configured to construct the loss function based on the sparse cross-entropy loss:

[0060]

[0061] Where y represents the target label index, p y This represents the predicted probability corresponding to the target label index y;

[0062] The iteration unit is configured to iterate the sparse cross-entropy loss function a preset number of times using the Adam optimizer;

[0063] The accelerated convergence unit is configured to use the learning rate optimizer to verify whether the sparse cross-entropy loss function has stalled for more than a first preset number of rounds. If it has, the learning rate is halved.

[0064] The training termination unit is configured to use an early stopping training strategy to verify whether the sparse cross-entropy loss function no longer decreases in the second preset round. If so, the model training is stopped.

[0065] To solve the above-mentioned technical problems, the second technical solution adopted by the present invention is: to provide a multi-component gas early warning method based on sensing, storage, and computing integration, the method comprising the following steps:

[0066] S1: Utilize a sensor array capable of identifying different gases to measure the concentration of different gases within the battery compartment;

[0067] S2: Process the gas data collected by the sensor, and cache and store it;

[0068] S3: Based on deep learning algorithms, analyze the gas concentration and types of the preprocessed gas data, and predict changes in gas concentration.

[0069] S4: Based on the predicted changes in gas concentration, send a warning signal to the battery and take safety measures.

[0070] The multi-component gas early warning system and method based on integrated sensing, storage, and computing of the present invention have at least the following beneficial effects: 1. High sensitivity and rapid response: Through a multi-component gas sensor array, the system can accurately detect six characteristic gases (CO, hydrogen, methane, ethylene, acetylene, and propylene) and their characteristic proportions during battery thermal runaway, and respond within 6 seconds with a false alarm rate of less than 5%. It can quickly identify potential thermal runaway risks and provide timely warnings. 2. High integration and miniaturization: Using advanced packaging technology, the gas sensing, signal processing, and storage computing functions are highly integrated, resulting in a compact and low-power system suitable for new energy vehicles and energy storage power stations with limited space. 3. Intelligent linkage and intelligent identification: The system is linked with the BMS via a CAN bus to transmit gas identification results in real time, and combines deep learning algorithms to analyze gas ratios, accurately warning of thermal runaway events and improving battery safety. It also has data storage capabilities, supporting fault diagnosis and post-analysis. 4. Wide range of applications: This system can be widely applied in new energy vehicles, energy storage power stations, fire protection and industrial safety, smart homes, and other fields to achieve real-time gas monitoring and risk warning, ensuring safety. Attached Figure Description

[0071] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0072] Figure 1 This is a structural block diagram of one embodiment of the multi-component gas early warning system based on sensing, storage, and computing of the present invention.

[0073] Figure 2 This is a structural block diagram of the data storage and processing module of the present invention.

[0074] Figure 3 This is a structural block diagram of the temperature compensation model construction submodule of the present invention.

[0075] Figure 4 This is a structural block diagram of the linear fitting submodule of the present invention.

[0076] Figure 5 This is a structural block diagram of the analysis and prediction module of the present invention.

[0077] Figure 6 The structural block diagram for constructing sub-modules for the dataset of this invention.

[0078] Figure 7 This is a structural block diagram of the prediction model construction submodule of the present invention.

[0079] Figure 8 This is a structural block diagram of the prediction model training submodule of the present invention.

[0080] Figure 9 This is a flowchart of one embodiment of the multi-component gas early warning method based on sensing, storage, and computing integration of the present invention. Detailed Implementation

[0081] The invention will now be further described with reference to the accompanying drawings.

[0082] Please see Figure 1 This is a flowchart of an embodiment of the multi-component gas early warning system based on sensor-memory-computing integration of the present invention. This embodiment of the system may specifically include the following modules:

[0083] The sensor array module 100 is configured to measure the concentration of different gases within the battery compartment using a sensor array capable of identifying different gases. This sensor array module 100 consists of multiple miniature gas sensor units manufactured using MEMS technology. Each sensor unit contains a sensitive material for a different gas, enabling simultaneous detection of multiple gases. Each sensor unit employs a different gas-sensitive material to improve sensitivity to gases such as carbon monoxide, hydrogen, methane, ethylene, acetylene, and propylene. Each sensor unit in the array uses a different gas-sensitive material designed for the characteristic adsorption properties of different gases. The gas-sensitive materials are precisely arranged on electrodes. When gas molecules adsorb onto the surface of the gas-sensitive material, a change in resistance is caused. Based on the degree of resistance change, the sensor can accurately identify the concentration and type of gas. Its manufacturing process includes cantilever beam fabrication, heater deposition, sensitive material deposition, electrode fabrication, and encapsulation.

[0084] The data storage and processing module 200 is configured to process, cache, and store the gas data acquired by the sensor. This module acquires the sensor output signal through a signal reading circuit (which may include modules such as a MUX, resistor bridge, filter, PGA, and ADC), and performs amplification and filtering to ensure signal quality and accuracy. The sensor output signal data is stored in memory (ROM and RAM) to provide a basis for subsequent analysis and calculation. Furthermore, a microprocessor can be used for data preprocessing, such as noise reduction and feature extraction.

[0085] Please see Figure 2 This data storage processing module 200 may include the following sub-modules:

[0086] Preprocessing submodule 210 is configured to perform Kalman filtering and drift correction on the acquired gas data;

[0087] The temperature compensation model construction submodule 220 is configured to construct a temperature compensation model relating the temperature drift and temperature change of the sensor resistance to the temperature compensation coefficient. Since the sensor resistance will drift with temperature changes, in order to calculate the concentration of each gas component more accurately, this temperature compensation model construction submodule 220 constructs a temperature compensation model, which can use the temperature coefficient to compensate for the temperature drift of the sensor resistance caused by temperature changes, thereby reducing the calculation error of the concentration of each gas component and improving the accuracy of the system.

[0088] Please see Figure 3 The temperature compensation model construction submodule 220 may include the following units:

[0089] The temperature drift calculation unit 221 is configured to set different temperature points in a constant temperature chamber, maintain a constant temperature at each temperature point for a preset time, measure the average of multiple original resistance values ​​of the sensor at each temperature point at a fixed frequency, and calculate the temperature drift of the sensor resistance at each temperature point relative to a reference temperature point. This temperature drift calculation unit 221 measures the temperature drift of the sensor by performing a temperature rise test on the sensor in a constant temperature chamber. The specific test process is as follows: First, the sensor is placed in a constant temperature chamber without simulating battery thermal runaway gas injection; the sensor resistance value is recorded only in air. Temperature points are set at intervals of 5°C from 0°C to 60°C, and each temperature point is maintained at a constant temperature for 3 minutes. The original resistance value of the sensor at each temperature point is measured at a frequency of 1 Hz. The average of the recorded sensor resistance values ​​at each temperature point is taken to obtain the relationship between the sensor resistance and temperature. A reference temperature, such as 25°C, can then be preset, and the difference between the sensor resistance value at each temperature point and the sensor resistance value at the reference temperature point can be calculated. This difference is the temperature drift curve of the sensor.

[0090] Third-order temperature compensation model building unit 222 is configured to build a temperature compensation model using a third-order polynomial:

[0091] R {comp} =R {raw} -[α·(TT {0} )+β·(TT {0} ) 2 +γ·(TT {0} ) 3 ];

[0092] Among them, R {comp} R represents the sensor resistance after temperature compensation. {raw} The original resistance of the sensor before temperature compensation is represented by α, β, and γ, which represent the temperature compensation coefficients, and T represents the ambient temperature. {0}The reference temperature is represented here. To eliminate the influence of temperature on the sensor resistance, it can be assumed that there is a continuous and smooth curve between temperature drift and temperature difference. Therefore, a third-order polynomial can be used to approximate this relationship. That is, a temperature compensation model is constructed using the above third-order polynomial to represent the relationship between the sensor resistance and the temperature difference before and after compensation.

[0093] The linear fitting submodule 230 is configured to fit the temperature compensation model based on the collected temperature drift and actual temperature change values ​​of the sensor resistance, and calculate the temperature compensation coefficient.

[0094] Please see Figure 4 This linear fitting submodule 230 may include the following units:

[0095] The fitting equation construction unit 231 is configured to construct a fitting equation based on the actual values ​​of the temperature drift and temperature change of the acquired sensor resistance:

[0096]

[0097] Wherein, ΔR(T) {i} ) represents the temperature drift of the sensor resistance at the i-th temperature point, x {i} =T {i} -T {0} This represents the temperature difference between the i-th temperature point and the reference temperature. To solve for the three temperature compensation coefficients α, β, and γ, the fitting equation construction unit 231 constructs the above fitting equation by taking the temperature drift as a known quantity. The specific solution to this fitting equation can be obtained by constructing a matrix relating the sensor resistance, temperature drift, and compensation coefficients using the least squares method, and then obtaining the three temperature compensation coefficients α, β, and γ by solving a closed-form equation using the least squares method. Furthermore, a linear or logarithmic model can be constructed for the sensor resistance and gas concentration. Substituting the temperature-compensated sensor resistance into the gas concentration model yields the gas concentration value.

[0098] Error verification unit 232 is configured to verify the effectiveness of the temperature compensation model using root mean square error:

[0099]

[0100] Where M represents the total number of data points, R j R represents the j-th measured value of the sensor resistance. refj This represents the reference value corresponding to the measured value. To quantify the deviation between the sensor resistance and the reference value before and after the temperature compensation algorithm, the root mean square error can be used for verification, as shown in the root mean square error verification formula given in this error verification unit 232.

[0101] The temperature compensation submodule 240 is configured to map the temperature compensation coefficient into the temperature compensation model and compensate for all collected gas data.

[0102] The analysis and prediction module 300 is configured to analyze the gas concentration and type of preprocessed gas data using a deep learning algorithm, and to predict changes in gas concentration. This module 300 employs a Deep Separable Convolutional Neural Network (DS-CNN) algorithm, combined with multidimensional time-series signals output from a gas sensor array, to extract features and identify gas types and concentrations. It utilizes deep learning algorithms, particularly CNN and Deepwise Separable Convolution techniques, to reduce computational complexity and improve recognition efficiency. The system learns the proportional relationships between gases during thermal runaway (e.g., specific ratios of carbon monoxide to hydrogen, methane to ethylene) to achieve early warning based on these proportional relationships. The feature proportion model is trained and optimized through data, enabling the system to promptly identify and issue warnings when gas proportion changes conform to thermal runaway characteristics.

[0103] Please see Figure 5 The analysis and prediction module 300 may include the following sub-modules:

[0104] The dataset construction submodule 310 is configured to construct training and validation sets based on simulated gas data collected by sensors in a simulated battery thermal runaway environment, and to standardize the training and validation sets. This dataset construction submodule 310 can simulate the typical characteristic gases (CO, H2, C2H4, etc.) released during lithium battery thermal runaway. A battery is placed in an 80L stainless steel sealed controllable heating chamber, and the temperature inside the chamber is increased from 25°C by a hot plate at a rate of 5°C / min to simulate thermal runaway and continuous gas release. During this process, the sensor array can sample in real time to acquire the dataset.

[0105] Please see Figure 6 The dataset construction submodule 310 may include the following units:

[0106] Dataset partitioning unit 311 is configured to divide the collected simulated gas data into 8 channels and 180 time steps per sample, and to divide the training set and validation set in a ratio of 85%:15%, and to maintain the consistency of the distribution of each type of sample in the training set and validation set by using stratify=y; in addition, the dataset collected by this dataset partitioning unit 311 contains 4 types of gases, so the text labels can be mapped to integers 0-3.

[0107] Normalization unit 312 is configured to perform z-score normalization on the collected dataset:

[0108]

[0109] in, μ represents the data point of channel c at time t after normalization of the original sensor sequence. c σ c These represent the mean and standard deviation of the training set across the c channels, respectively. This method allows for z-score standardization for each channel: first, (sample, channel, time step) is flattened to (sample, channel × time step), and the mean and standard deviation of the entire training set are calculated; then, the training and validation sets are transformed back to (sample, time step, channel) format to match the input requirements of Keras Conv1D. This accelerates model convergence and eliminates differences in the dimensions of different channels.

[0110] The prediction model construction submodule 320 is configured as a deepwise separable convolutional neural network structure and utilizes the SE attention module to enhance the channels of the convolutional output. The neural network of this prediction model construction submodule 320 is based on a Depthwise Separable CNN (DS-CNN) structure, with Squeeze-and-Excitation (SE) attention modules added at key locations to achieve lightweight and efficient channel recalibration. The specific process is as follows:

[0111] The input layer receives a preprocessed signal with a shape of (T = 180 time steps, C = 8 channels);

[0112] DepthwiseConv1D+PointwiseConv1D: First, use DepthwiseConv1D (convolution kernel length 5) to extract temporal features for each signal, and then fuse the information between channels through convolution.

[0113] BatchNorm+ReLU: Standardization and non-linear activation are performed immediately after each convolution to accelerate convergence and suppress internal covariate shift.

[0114] SE Attention Module: Performs global average pooling on the depthwise convolution output, generates channel-level attention weights, multiplies them with the original features, and dynamically enhances key channels;

[0115] Several DS-CNN Block+SE: Stack two depthwise separable convolutional modules in sequence, and perform MaxPool1D(2) downsampling after the second module, followed by SE weighting;

[0116] GlobalAveragePooling1D: Global pooling along the time dimension to generate a C=64 two-dimensional feature vector;

[0117] Dense(64)+ReLU→Dense(4)+Softmax: First reduce the dimensionality and then map it to the probability distribution of the four gas categories.

[0118] This design maximizes the computational and parameter efficiency of DS-CNN, as well as the adaptive compensation of channel correlation by the SE module, so that the model can meet the lightweight requirements of edge devices while maintaining classification performance.

[0119] Please see Figure 7 The prediction model construction submodule 320 may include the following units:

[0120] The depthwise convolutional unit 321 is configured to perform temporal convolutions within each channel using one-dimensional convolutional kernels to extract temporal features.

[0121]

[0122] Among them, w c,i This represents the i-th convolution weight in channel c;

[0123] Pointwise convolutional unit 322 is configured to linearly combine the outputs of depthwise convolutions along the channel direction:

[0124]

[0125] Among them, v j,c This represents the 1×1 convolution weight between output channel j and input channel c;

[0126] Attention calibration unit 323 is configured to recalibrate channel weights using the SE attention module:

[0127] z' c,t =σ(W2ReLU(W1s))·z c,t

[0128] Where W1 and W2 represent the weight matrices of the fully connected layer, s represents the channel descriptor, and σ represents the sigmoid activation function;

[0129] Classification unit 324 is configured to calculate class probabilities using Softmax:

[0130]

[0131] Among them, w k b k Let p represent the weight vector and bias of the k-th class, respectively, where K represents the number of classes. k ∈[0,1] and

[0132] The prediction model training submodule 330 is configured to input the training set into the prediction model and minimize the sparse cross-entropy loss through the Adam optimizer to obtain the trained prediction model.

[0133] Please see Figure 8 The training submodule 330 of this prediction model may include the following units:

[0134] Loss function building unit 331 is configured to construct a loss function based on sparse cross-entropy loss:

[0135]

[0136] Where y represents the target label index, p y This represents the predicted probability corresponding to the target label index y;

[0137] Iteration unit 332 is configured to iterate the sparse cross-entropy loss function a preset number of times using the Adam optimizer;

[0138] Accelerated convergence unit 333 is configured to use a learning rate optimizer to verify whether the sparse cross-entropy loss function has stalled for more than a first preset number of rounds. If it has, the learning rate is halved. Specifically, accelerated convergence unit 333 can use the ReduceLROnPlateau() optimizer, and the first preset number of rounds can be set to 5. That is, after verifying that the loss has stalled for 5 rounds, the learning rate is automatically halved to accelerate subsequent convergence.

[0139] Training termination unit 334 is configured to use an early stopping training strategy to verify whether the sparse cross-entropy loss function no longer decreases in the second preset round. If so, model training is stopped. Specifically, training termination unit 334 can use the EarlyStopping() callback function, and the second preset round can be set to 10 times. That is, if the loss no longer decreases after more than 10 rounds, the training is terminated early and the optimal weights are restored.

[0140] The early warning module 400 is configured to send an early warning signal to the battery and take safety measures based on predicted changes in gas concentration. When the system detects a change in the concentration of a characteristic gas and it reaches a set threshold, the early warning module 400 can transmit an alarm signal in real time to the battery management system (BMS) via the CAN bus. The BMS then takes appropriate safety measures based on the early warning information, such as cutting off battery power or activating the cooling system.

[0141] Please see Figure 9This is a flowchart illustrating an embodiment of the multi-component gas early warning method based on sensor-memory-computation integration of the present invention. The multi-component gas early warning method based on sensor-memory-computation integration of this embodiment is used in the multi-component gas early warning system based on sensor-memory-computation integration described in the above embodiment. Specifically, the multi-component gas early warning method based on sensor-memory-computation integration of this embodiment includes the following steps:

[0142] S1: Utilize a sensor array capable of identifying different gases to measure the concentration of different gases within the battery compartment;

[0143] S2: Process the gas data collected by the sensor, and cache and store it;

[0144] S3: Based on deep learning algorithms, analyze the gas concentration and type of the preprocessed gas data, and predict changes in gas concentration.

[0145] S4: Based on the predicted changes in gas concentration, send a warning signal to the battery and take safety measures.

[0146] This invention utilizes a multi-component gas sensor array to accurately detect six characteristic gases (CO, hydrogen, methane, ethylene, acetylene, and propylene) and their proportions during battery thermal runaway, responding within 6 seconds with a false alarm rate of less than 5%. It can quickly identify potential thermal runaway risks and provide timely warnings. Employing advanced packaging technology, it highly integrates gas sensing, signal processing, and in-memory computing functions, resulting in a compact, low-power system suitable for space-constrained new energy vehicles and energy storage power stations. It can also link with a BMS via a CAN bus to transmit gas identification results in real time and combine deep learning algorithms to analyze gas ratios, accurately predicting thermal runaway events and improving battery safety. It has data storage capabilities, supporting fault diagnosis and post-analysis. This system can be widely applied in new energy vehicles, energy storage power stations, fire protection and industrial safety, smart homes, and other fields to achieve real-time gas monitoring and risk warnings, ensuring safety. Furthermore, the modules in this solution can be packaged using multi-layer vertical integration, with the sensor array module at the top, the signal processing circuit in the middle, and the microprocessor and I / O interfaces at the bottom. TSV and RDL technologies connect functional modules at different levels through vertical vias, ensuring system miniaturization and efficient transmission.

[0147] The above description merely illustrates preferred embodiments of the present invention and is quite specific and detailed; however, it should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A multi-component gas early warning system based on integrated sensing, storage, and computing, characterized in that, include: The sensor array module is configured to measure the concentration of different gases in the battery compartment using a sensor array capable of identifying different gases. The data storage and processing module is configured to process, cache, and store the gas data collected by the sensor. The analysis and prediction module is configured to analyze the gas concentration and type of the preprocessed gas data based on a deep learning algorithm, and to predict changes in gas concentration. The early warning module is configured to send an early warning signal to the battery and take safety measures based on predicted changes in gas concentration.

2. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 1, characterized in that, The data storage processing module includes: The preprocessing submodule is configured to perform Kalman filtering and drift correction on the acquired gas data; The temperature compensation model construction submodule is configured to construct a temperature compensation model that relates the temperature drift and temperature change of the sensor resistance to the temperature compensation coefficient. The linear fitting submodule is configured to fit the temperature compensation model based on the collected temperature drift and actual temperature change values ​​of the sensor resistance, and calculate the temperature compensation coefficient. The temperature compensation submodule is configured to map the temperature compensation coefficient into the temperature compensation model and compensate for all collected gas data.

3. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 2, characterized in that, The temperature compensation model construction submodule includes: The temperature drift calculation unit is configured to set different temperature points in the constant temperature chamber, maintain a constant temperature at each temperature point for a preset constant temperature time, measure the average value of multiple original resistance values ​​of the sensor at each temperature point at a fixed frequency, and calculate the temperature drift of the sensor resistance at each temperature point relative to the reference temperature point. The third-order temperature compensation model building unit is configured to construct a temperature compensation model using a third-order polynomial: R {comp} =R {raw} -[α·(T-T {0} )+β·(T-T {0} ) 2 +γ·(T-T {0} ) 3 ]; Among them, R {comp} R represents the sensor resistance after temperature compensation. {raw} The original resistance of the sensor before temperature compensation is represented by α, β, and γ, which represent the temperature compensation coefficients, and T represents the ambient temperature. {0} Indicates the reference temperature.

4. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 3, characterized in that, The linear fitting submodule includes: The fitting equation construction unit is configured to construct a fitting equation based on the acquired temperature drift and actual temperature change values ​​of the sensor resistance: Wherein, ΔR(T) {i} ) represents the temperature drift of the sensor resistance at the i-th temperature point, x {i} =T {i} -T {0} This represents the temperature difference between the i-th temperature point and the reference temperature.

5. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 4, characterized in that, The linear fitting submodule also includes: The error verification unit is configured to verify the effectiveness of the temperature compensation model using the root mean square error. Where M represents the total number of data points, R j R represents the j-th measured value of the sensor resistance. refj This indicates the reference value corresponding to the measured value.

6. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 1, characterized in that, The analysis and prediction module includes: The dataset construction submodule is configured to build training and validation sets based on simulated gas data collected by sensors in a simulated battery thermal runaway environment, and to standardize the training and validation sets. The prediction model building submodule is configured as a deep separable convolutional neural network structure and utilizes the SE attention module to perform channel enhancement on the convolutional output; The prediction model training submodule is configured to input the training set into the prediction model and minimize the sparse cross-entropy loss through the Adam optimizer to obtain the trained prediction model.

7. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 6, characterized in that, The dataset construction submodule includes: The dataset partitioning unit is configured to divide the collected simulated gas data into 8 channels and 180 time steps per sample, and to divide the training set and validation set in a ratio of 85%:15%, and to maintain the consistency of the distribution of various types of samples in the training set and validation set by using stratify=y; The normalization unit is configured to perform z-score normalization on the collected dataset: in, μ represents the data point of channel c at time t after normalization of the original sensor sequence. c σ c represents the mean and standard deviation of the training set on the c channel, respectively.

8. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 6, characterized in that, The prediction model construction submodule includes: The depthwise convolutional unit is configured to perform temporal convolutions within each channel using one-dimensional convolutional kernels to extract temporal features. Among them, w c,i This represents the convolution weight at the i-th position of channel c; Pointwise convolutional units are configured to linearly combine the outputs of depthwise convolutions along the channel direction: Among them, v j,c This represents the 1×1 convolution weight between output channel j and input channel c; The attention calibration unit is configured to recalibrate the channel weights using the SE attention module: With' c,t =σ(W2ReLU(W1s))·z c,t Where W1 and W2 represent the weight matrices of the fully connected layer, s represents the channel descriptor, and σ represents the sigmoid activation function; The classification unit is configured to calculate class probabilities using Softmax: Among them, w k b k Let p represent the weight vector and bias of the k-th class, respectively, where K represents the number of classes. k ∈[0,1 and 9. The multi-component gas early warning system based on sensor-memory-computing integration as described in claim 6, characterized in that, The prediction model training submodule includes: The loss function building unit is configured to construct the loss function based on the sparse cross-entropy loss: Where y represents the target label index, p y This represents the predicted probability corresponding to the target label index y; The iteration unit is configured to iterate the sparse cross-entropy loss function a preset number of times using the Adam optimizer; The accelerated convergence unit is configured to use a learning rate optimizer to verify whether the sparse cross-entropy loss function has stalled for more than a first preset number of rounds. If it has, the learning rate is halved. The training termination unit is configured to use an early stopping training strategy to verify whether the sparse cross-entropy loss function no longer decreases in the second preset round. If so, the model training is stopped.

10. A multi-component gas early warning method based on integrated sensing, storage, and computing, characterized in that, Includes the following steps: S1: Utilize a sensor array capable of identifying different gases to measure the concentration of different gases within the battery compartment; S2: Process the gas data collected by the sensor, and cache and store it; S3: Based on deep learning algorithms, analyze the gas concentration and type of the preprocessed gas data, and predict changes in gas concentration. S4: Based on the predicted changes in gas concentration, send a warning signal to the battery and take safety measures.

Images