Gas classification system and method based on MEMS sensor array

Abstract

The invention discloses a gas classification system and method based on an MEMS sensor array. The MEMS sensor array comprises a bearing substrate, a plurality of heating elements mounted on the bearing substrate in an array manner, and a plurality of MEMS gas sensors correspondingly mounted on the plurality of heating elements; the plurality of voltage followers are correspondingly connected with the plurality of MEMS gas sensors; the input end of the microcontroller is connected with the plurality of voltage followers through the ADC module, and the output end of the microcontroller is connected with the plurality of heating elements through the DAC module and the power amplification module in sequence. The method comprises the following steps: acquiring more gas response data by changing the heating condition of the MEMS gas sensor, and constructing a gas data set; training a gas classification model by using the gas data set; the gas is tested by using the MEMS gas sensor to obtain response data, the data characteristics of the response data are extracted, and the gas is classified through the gas classification model. According to the system and the method, different gases or gas mixtures can be identified more accurately at low cost.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas intelligent recognition, and specifically relates to a gas classification system and method based on a MEMS sensor array. Background Art

[0002] Currently, most electronic nose systems based on MEMS gas sensors adopt a hardware design scheme of multi-sensor integration, aiming to provide as much raw data as possible for the subsequent recognition algorithm. However, when the number of physical sensors increases significantly, the overall system loses the characteristics of miniaturization, low power consumption, and low cost, thereby limiting the application scope of the gas classification system.

[0003] In the prior art, in order to enhance the selectivity of MEMS gas sensors, the MEMS gas sensors usually use a pulse heating modulation scheme. This modulation scheme has the following two main advantages. On the one hand, high-frequency pulse modulation can reduce power consumption while maintaining the stable working condition of the sensor. On the other hand, low-frequency pulse modulation can obtain a "gas dynamic response", thereby improving the classification accuracy. Obviously, using low-frequency pulse modulation also has the characteristic of reducing power consumption, and it has been verified that low-frequency pulses have a beneficial effect on improving classification accuracy. For high-frequency pulse modulation, its modulation scheme is usually relatively single, and it often needs to use square waves, triangular waves, or sine waves with fixed amplitudes. Although heating with these waveforms can increase the amount of effective data, it cannot significantly reduce the complexity of the network model. In view of the problems of limited gas response data sources and inaccurate gas classification effects in the prior art, there is an urgent need to provide a new type of intelligent gas classification system and method. Summary of the Invention

[0004] Aiming at the problems existing in the above-mentioned prior art, the present invention provides a gas classification system and method based on a MEMS sensor array. The system has a simple structure and low manufacturing cost. It can obtain the response effect of a large number of MEMS gas sensors under the condition of a limited number of MEMS gas sensors, can significantly enhance the source response data, can greatly increase the information volume of the source response data, and helps to achieve accurate gas classification. The method has a simple implementation process and high intelligence. It can obtain more gas response data by changing the heating conditions of gas sensors under limited resource conditions, and can more accurately identify different gases or gas mixtures.

[0005] To achieve the above object, the present invention provides a gas classification system based on a MEMS sensor array, including a MEMS sensor array, a voltage follower, an ADC module, a DAC module, a power amplifier module, and a microcontroller;

[0006] The MEMS sensor array includes a carrier substrate, a plurality of heating elements mounted on the carrier substrate in an array, and a plurality of MEMS gas sensors correspondingly mounted on the plurality of heating elements;

[0007] A plurality of voltage followers are correspondingly connected to the plurality of MEMS gas sensors;

[0008] The input end of the microcontroller is connected to the plurality of voltage followers through an ADC module, and the output end of the microcontroller is sequentially connected to the plurality of heating elements through a DAC module and a power amplification module.

[0009] As a preference, the microcontroller is a ZYNQ chip, which includes a PS part and a PL part. The PS part includes an ARM dual-core processor and a DDR storage module, and the PL part includes an FPGA; the ARM dual-core processor and the FPGA are connected through an AXI bus and perform data interaction.

[0010] As a preference, the ZYNQ chip is Xilinx zynq7020, and the ARM dual-core processor module is a Cortex-A9 ARM core; the DDR storage module is DDR3.

[0011] As a preference, the MEMS gas sensor is a carbon monoxide sensor or a hydrogen sensor.

[0012] As a preference, the model of the carbon monoxide sensor is GM-702B.

[0013] As a preference, the model of the hydrogen sensor is GMV-2021B.

[0014] Furthermore, in order to facilitate driving the MEMS gas sensor with a low-frequency signal, a low-pass filter module is further included, and the low-pass filter module is arranged between the DAC module and the power amplification module. Since the complete gas response of the MEMS gas sensor occurs on a time scale of several minutes, the amplitude-modulated driving signal of the sensor should be a low-frequency signal. Therefore, the amplitude-modulated driving signal of the MEMS gas sensor should be a low-frequency signal.

[0015] In the present invention, a MEMS gas sensor is mounted on a heating element of a carrier substrate, which facilitates changing the operating ambient temperature of the MEMS gas sensor by controlling the heating element. Since the sensitivity of the MEMS gas sensor varies under different temperature conditions, the sensitivity of the MEMS gas sensor can be conveniently changed. Thus, by varying the heating conditions of the sensor, more information about gas response can be obtained, and response data under different temperature conditions can be acquired, which is beneficial to increasing the diversity and richness of the sample data set. Furthermore, a gas classification model with higher classification accuracy can be trained, which helps to more accurately identify different gases or gas mixtures through this technical means. By arranging a power amplification module between the DAC module and the heating element, it is convenient to amplify the voltage drive signal sent by the microprocessor using the power amplification module, thereby enhancing the energy of the signal. Thus, the amplitude and voltage of the signal received by each sensor can be controlled, and the heating element can be more efficiently driven to change the operating temperature of the MEMS gas sensor. By equipping each MEMS gas sensor with a voltage follower, each MEMS gas sensor can have an independent dedicated data acquisition channel. On this basis, by connecting the microcontroller to the voltage follower through the ADC module, it is convenient to obtain the response data of the MEMS gas sensor in real time. The system has a simple structure and low manufacturing cost. It can conveniently change the operating temperatures of multiple MEMS gas sensors, and thus change the sensitivities of multiple MEMS gas sensors. Therefore, under the condition of a limited number of MEMS gas sensors, the response effect equivalent to that of a huge number of MEMS gas sensors can be obtained, achieving a significant enhancement of the source response data, greatly increasing the information volume of the source response data. Furthermore, it is beneficial to obtain a gas classification model with higher classification accuracy under low-cost conditions, which helps to achieve accurate gas classification.

[0016] The present invention also provides a gas classification method based on a MEMS sensor array, which uses a gas classification system based on a MEMS sensor array, and includes the following steps;

[0017] Step 1: Obtain a gas data set;

[0018] S11: Place the MEMS sensor array in a test chamber and prepare a mixed gas sample according to the set component concentration information;

[0019] S12: The microcontroller adopts multi - waveform modulation technology to generate k voltage drive signals based on k preset waveform drive signals. Among them, the voltage drive signal is an amplitude - adjustable periodic signal, and a unique drive state identification signal is set for each voltage drive signal, and then sent to the DAC module. The DAC module sends the received k voltage drive signals to the low - pass filter module; the low - pass filter module filters the voltage drive signals and then sends them to the power amplifier module; the power amplifier module amplifies the input voltage drive signals and then outputs them to the heating element to drive the corresponding heating element to perform heating operations using different voltage drive signals, so that the corresponding MEMS gas sensor works under different temperature conditions; after the working temperature condition of the MEMS gas sensor is stable, inject the mixed gas with the prepared component concentration data into the measurement chamber;

[0020] S13: Make multiple MEMS gas sensors perform response actions under the condition of the mixed gas with the set component concentration data. When multiple MEMS gas sensors reach a steady state, use a voltage follower to collect the voltage signals of the corresponding MEMS gas sensors in real - time and send them to the ADC module. After the ADC module converts the received analog voltage signals into digital voltage signals, it sends the digital voltage signals to the microcontroller; the microcontroller obtains the response data of the MEMS gas sensor according to the received digital voltage signals. The microcontroller synchronously records the drive state identification signal and the corresponding response data as gas original sample data, attaches the corresponding gas label, and then stores it in the storage module;

[0021] S14: Inject clean air into the measurement chamber and displace the mixed gas to the outside until the measurement chamber is filled with clean air;

[0022] S15: Repeat S12 to S14 multiple times to obtain a large amount of gas original sample data;

[0023] S16: Denoise the gas original sample data, and then separate the gas original data according to the drive state identification signals of different voltage drive signals. After separation, the response data under each voltage drive signal is obtained; for the separated response data, use the bisection peak - seeking algorithm and local inflection point algorithm to extract data features, then use the sample interpolation algorithm to expand the sample dimension, and then obtain the characteristic curve through feature point fitting; merge the different characteristic curves belonging to the same response process to form the characteristic sample data of a single gas, and multiple gas labels correspond to generate a set of characteristic sample data, forming a gas data set, thus completing the construction of the gas data set;

[0024] S17: Divide the gas data set into a training set, a test set, and a validation set according to a set ratio;

[0025] Step Two: Construct a gas classification model;

[0026] S21: Build an MLP-based classification model in the microcontroller, where the weights and biases of the classification model are set in the storage module of the PS part, the part of the classification model other than the weights and biases is set in the PL part, and the PL part communicates with the storage module through DMA;

[0027] S22: Train the classification model using the training set. During the training process, continuously update the weights and biases through the backpropagation algorithm to reduce the value of the loss function and implement the process of optimizing the model parameters. Then, test the trained classification model using the test set and validate the trained classification model using the validation set to finally obtain the gas classification model;

[0028] Step 3: Classify the gas;

[0029] S31: Place the MEMS gas sensor in the environment to be tested and obtain the response data of the MEMS gas sensor to the gas in the environment to be tested;

[0030] S32: First, perform denoising processing on the response data, and then use the bisection peak-finding algorithm and the local inflection point algorithm to extract data features;

[0031] S33: Use the data features as input data and input them into the gas classification model. The gas classification model compares the current data features with the known response feature data, and then identifies and outputs the gas classification information.

[0032] As an optimization, in S15 of Step 1, 70% of the gas dataset is divided into the training set, 15% of the gas dataset is divided into the test set, and 15% of the gas dataset is divided into the validation set.

[0033] Furthermore, in order to better avoid the loss of important feature information, in S16 of Step 1, during the process of expanding the sample dimension using the sample interpolation algorithm, expand the dimension to 2000.

[0034] The present invention proposes a gas classification method based on hybrid waveform modulation technology. By using multi-waveform modulation technology to drive the heating elements of gas sensors in a MEMS sensor array, the sensitivity of semiconductors to temperature can be fully utilized, enabling the MEMS gas sensors to operate under different temperature conditions. As a result, the data features for distinguishing gas types contained in the response process of the gas sensors are rich enough, effectively increasing the feature dimension of gas response data and significantly reducing the complexity of the network model. The voltage drive signal used is an adjustable amplitude periodic signal, and the response data of the MEMS gas sensors are collected under the driving conditions of each voltage drive signal, which can effectively capture the reaction of the sensors to the gas under specific heating conditions. These data can then be used as gas sample data. By analyzing these data, different types of gases can be detected and classified. In this way, the effect of increasing the sensors can be achieved without actually increasing the number of sensors, enhancing the source data at the source of data generation, greatly increasing the information volume of gas data, and obtaining a more accurate gas classification effect. In each data collection process of the present invention, multiple waveform drive signals are used to generate multiple voltage drive signals. Thus, it is convenient to increase the data features as much as possible by the method of superimposing multiple waveforms for heating. During the feature extraction process, the bisection peak-seeking algorithm and the local inflection point algorithm are used to extract data features. By setting appropriate thresholds, peaks, valleys, and inflection points can be identified, enhancing the feature extraction process and providing a more comprehensive representation of the sensor response. Compared with traditional technologies, the present invention further increases the absorption peaks of gases, improves the reaction rate of gases, can provide more data features for the subsequent classification model, and thus can significantly reduce the complexity of the classification model, achieving the effect of reducing the complexity of the subsequent MLP neural network structure. At the same time, by using the sample interpolation algorithm to expand the sample dimension, the present invention can effectively prevent the actual peaks from being filtered out as noise in the subsequent filter bank process, avoiding the loss of important feature information and further ensuring the accuracy of the classification model. In addition, during the construction of the classification model, the weights and biases are stored in the storage module in the PS part, and at the same time, the PL part calls the DMA for high-speed data transmission, which can significantly reduce the utilization rate of the FPGA, thereby reducing the power consumption. This technical means can effectively balance the hardware structure and the complexity of the algorithm, which is beneficial to improving the target accuracy of the classification task.

[0035] The implementation process of this method is simple and has a high degree of intelligence. By changing the heating conditions of the gas sensors under limited resource conditions, more data on gas responses can be obtained, thereby increasing the diversity of the dataset. With richer data, a classification model with higher classification accuracy can be obtained, enabling more accurate identification of different gases or gas mixtures. Description of the Drawings

[0036] Figure 1 is a schematic structural diagram of the gas classification system in the present invention;

[0037] Figure 2 is a partial enlarged view of the characteristic points in the response curve in an embodiment of the present invention;

[0038] Figure 3 is a waveform diagram of the drive signal adopted in an embodiment of the present invention;

[0039] Figure 4 is a schematic structural diagram of the ZYNQ chip in the present invention;

[0040] Figure 5 is a schematic diagram of the original gas sample data in the present invention;

[0041] Figure 6 is a data feature map extracted by using the dichotomy peak search algorithm and the local inflection point algorithm in the present invention.

[0042] In the figure: 1, microcontroller; 2, DAC module; 3, power amplification module; 4, ADC module; 5, MEMS sensor array; 6, MEMS gas sensor; 7, carrier substrate. Detailed implementation manners

[0043] The present invention will be further described below with reference to the accompanying drawings.

[0044] As Figures 1 to 6 shown, the present invention provides a gas classification system based on an MEMS sensor array, including an MEMS sensor array 5, a voltage follower, an ADC module 4, a DAC module 2, a power amplification module 3, and a microcontroller 1;

[0045] The MEMS sensor array 5 includes a carrier substrate 7, a plurality of heating elements arranged in an array on the carrier substrate 7, and a plurality of MEMS gas sensors 6 correspondingly installed on the plurality of heating elements;

[0046] A plurality of voltage followers are correspondingly connected to the plurality of MEMS gas sensors 6;

[0047] The number of heating elements, MEMS gas sensors 6, and voltage followers can be determined according to the expected target of the classification task and the performance indexes of the optional sensors. As a preference, the number of heating elements, MEMS gas sensors 6, and voltage followers is four.

[0048] The input end of the microcontroller 1 is connected to the plurality of voltage followers through the ADC module 4, and the output end of the microcontroller 1 is sequentially connected to the plurality of heating elements through the DAC module 2 and the power amplification module 3.

[0049] As a preference, the microcontroller 1 is a ZYNQ chip, which includes a PS part and a PL part. The PS part includes an ARM dual-core processor and a DDR storage module, and the PL part includes an FPGA; the ARM dual-core processor and the FPGA are connected through an AXI bus and perform data interaction, as Figure 4 shown.

[0050] As a preference, the ZYNQ chip is Xilinx zynq7020, and the ARM dual-core processor module is a Cortex-A9 ARM core; the DDR storage module is DDR3.

[0051] As a preference, the MEMS gas sensor 6 is a carbon monoxide sensor or a hydrogen sensor.

[0052] As a preference, the model of the carbon monoxide sensor is GM-702B.

[0053] As a preference, the model of the hydrogen sensor is GMV-2021B.

[0054] To facilitate driving the MEMS gas sensor with a low-frequency signal, a low-pass filter module is further included, and the low-pass filter module is arranged between the DAC module 2 and the power amplifier module 3. Since the complete gas response of the MEMS gas sensor occurs on a time scale of several minutes, the amplitude-modulated drive signal of the sensor should be a low-frequency signal. Therefore, the amplitude-modulated drive signal of the MEMS gas sensor should be a low-frequency signal (usually effective when the frequency ≤ 1 Hz). As a preference, the low-pass filter module is a 7th-order Butterworth low-pass filter.

[0055] In the present invention, the MEMS gas sensor is mounted on the heating element of the carrier substrate, which facilitates changing the operating ambient temperature of the MEMS gas sensor by controlling the heating element. Since the MEMS gas sensor has different sensitivities under different temperature conditions, the sensitivity of the MEMS gas sensor can be conveniently changed. Thus, by varying the heating conditions of the sensor, more information about the gas response can be obtained, and the response data under different temperature conditions can be obtained, which is beneficial to increasing the diversity and richness of the sample data set. Furthermore, a gas classification model with higher classification accuracy can be trained, which helps to more accurately identify different gases or gas mixtures through this technical means. By arranging a power amplification module between the DAC module and the heating element, it is convenient to amplify the voltage drive signal sent by the microprocessor using the power amplification module, thereby enhancing the energy of the signal. Thus, the amplitude and voltage of the signal received by each sensor can be controlled, and the heating element can be more efficiently driven to change the operating temperature of the MEMS gas sensor. By equipping each MEMS gas sensor with a voltage follower, each MEMS gas sensor can have an independent dedicated data acquisition channel. On this basis, connecting the microcontroller to the voltage follower through the ADC module facilitates obtaining the response data of the MEMS gas sensor in real time. The system has a simple structure and low manufacturing cost. It can conveniently change the operating temperatures of multiple MEMS gas sensors, thereby changing the sensitivities of multiple MEMS gas sensors. Thus, under the condition of limited number of MEMS gas sensors, the response effect of a large number of MEMS gas sensors can be obtained, realizing a significant enhancement of the source response data, greatly increasing the information volume of the source response data, which is beneficial to obtaining a gas classification model with higher classification accuracy under low-cost conditions and helps to achieve accurate gas classification.

[0056] The present invention also provides a gas classification method based on a MEMS sensor array, which adopts a gas classification system based on a MEMS sensor array, including the following steps;

[0057] Step 1: Obtain a gas data set;

[0058] S11: Place the MEMS sensor array 5 in the test chamber and prepare a mixed gas sample according to the set component concentration information;

[0059] As a preference, in an embodiment of the present invention, the mixed gas sample can be prepared according to the component concentration information table in Table 1, and 47 mixed gas samples are prepared;

[0060] Table 1: Component Concentration Information Table

[0061]

[0062]

[0063]

[0064] S12: The microcontroller 1 adopts multi - waveform modulation technology to generate k voltage drive signals based on k preset waveform drive signals, and sets a unique drive state identification signal for each voltage drive signal. When changing the voltage drive signal, the corresponding drive state identification signal is generated synchronously, as Figure 3 shown;

[0065] The k voltage drive signals represent the drive voltages of the heating element of the gas sensor in different working states. Among them, each different waveform drive signal corresponds to a different voltage change. The voltage drive signal is an adjustable - amplitude periodic signal, that is, its amplitude (the maximum value of the voltage) and period (the frequency of signal repetition) can be adjusted as needed. These voltage drive signals are used to activate the heating element of the gas sensor. The heating element will generate different temperature conditions under different voltage drives, so that the response characteristics of the sensor to the gas can be changed;

[0066] Then, the k voltage drive signals are sent to the DAC module 2, and the DAC module 2 sends the received k voltage drive signals to the low - pass filter module; the low - pass filter module filters the voltage drive signals to obtain a smooth waveform control signal, and then sends it to the power amplifier module 3; the power amplifier module 3 amplifies the input voltage drive signal and then outputs it to the heating element. In this way, sufficient energy can be generated to drive the corresponding heating element for heating operation using different voltage drive signals, so that the corresponding MEMS gas sensor 6 works under different temperature conditions; when the working temperature condition of the MEMS gas sensor 6 is stable, a mixed gas with prepared component concentration data is injected into the measurement chamber;

[0067] As a preference, the frequency of the voltage drive signal is 50 Hz;

[0068] S13: Make multiple MEMS gas sensors 6 perform response actions under the condition of a mixed gas with set component concentration data. When multiple MEMS gas sensors 6 reach a steady state, use a voltage follower to collect the voltage signals of the corresponding MEMS gas sensors 6 in real - time and send them to the ADC module 4. After the ADC module 4 converts the received analog voltage signal into a digital voltage signal, it sends the digital voltage signal to the microcontroller 1; the microcontroller 1 obtains the response data of the MEMS gas sensor 6 according to the received digital voltage signal. The microcontroller 1 synchronously records the drive state identification signal and the corresponding response data as the gas original sample data, marks the corresponding gas label, and then stores it in the storage module;

[0069] S14: Inject clean air into the measurement chamber and displace the mixed gas to the outside until the measurement chamber is filled with clean air;

[0070] S15: Repeat S12 to S14 multiple times to obtain a large amount of original gas sample data;

[0071] S16: Denoise the original gas sample data to make the original gas data sample more accurate, and then separate the original gas data according to the driving state identification signals of different voltage driving signals. After separation, the response data under each voltage driving signal is obtained;

[0072] For the separated response data, use the bisection peak-finding algorithm and the local inflection point algorithm to extract data features. As Figure 6 shown, the points that can be highlighted in color represent the feature points of the response curve. These color-coded points represent key features such as peaks, valleys, and inflection points, providing a clear visual representation of the important features in the data.

[0073] The specific above-mentioned feature extraction algorithm is as follows. For a set of discrete data y1, y2,..., y n , for each data point y i (2 ≤ i ≤ n - 1): If y i > y i-1 + w and y i > y i+1 + b, then y i is a peak value. If y i < y i-1 - w ′ and y i < y i+1 - b ′ , then y i is a valley value. If |k i-1 | < E and |k i+1 | > E ′ , then y i is an inflection point. Where w, b, w ′ , b ′ represent the determination parameters of peak values and valley values, E represents the determination parameter of inflection points, k i-1 represents the slope between y i and y i-1 , k i+1 represents the slope between y i and y i+1 . The size and step length of the slope sliding window will affect the accuracy of algorithm recognition. From the characteristics and change trends of the data itself, for peak points, the data interval on the left side is smaller and the data interval on the right side is larger. Therefore, when selecting parameters, ensure that w < b. For valley points, it is exactly the opposite, and it is necessary to ensure that w′ >b ′ Use the grid search method to determine the window length and step size. After multiple rounds of experiments, the parameter values are determined: w = 0.01305, b = 0.01633, w ′ = 0.00954, b ′ = 0.00538, E = 0.03, E ′ = 0.1, the window length is 300, and the step size is 10;

[0074] Then, use the sample interpolation algorithm to expand the sample dimension, and then obtain the feature curve through feature point fitting; merge different feature curves belonging to the same response process to form the feature sample data of a single gas. Multiple gas labels correspond to generate a set of feature sample data, which constitutes the gas data set, and then completes the construction of the gas data set, as Figure 5 shown;

[0075] S17: Divide the gas data set into a training set, a test set, and a validation set according to the set ratio;

[0076] Step 2: Construct a gas classification model;

[0077] S21: Construct a classification model based on MLP in the microcontroller 1. Among them, the weights and biases of the classification model are set in the storage module of the PS part, and the part of the classification model other than the weights and biases is set in the PL part. And the PL part communicates with the storage module through DMA. In this way, the use of hardware logic can be significantly reduced;

[0078] As an optimization, according to the principle of MLP, the weight matrices are defined as follows: the first-layer weight matrix is weights1

[50]

[2000] , the second-layer weight matrix is weights2

[50]

[50] , and the third-layer (output layer) weight matrix is weights3

[50] [8]. The bias matrices are defined as biases1

[50] , biases2

[50] , and biases3[8]. Each array element is a signed floating-point number. If these weights are implemented in FPGA hardware, they will consume a large part of the BRAM (Block Random Access Memory) resources.

[0079] S22: Use the training set to train the classification model. During the training process, continuously update the weights and biases through the backpropagation algorithm to reduce the value of the loss function and achieve the process of optimizing the model parameters. Then, use the test set to test the trained classification model, and use the validation set to validate the trained classification model. The accuracy of the gas classification model through validation is shown in Table 2, and finally, the gas classification model is obtained;

[0080] Table 2: Accuracy Ranking Table of Gas Classification Model

[0081]

[0082] Step 3: Classify the gas;

[0083] S31: Place the MEMS gas sensor 6 in the environment to be tested, and obtain the response data of the MEMS gas sensor 6 to the gas in the environment to be tested;

[0084] S32: First, perform denoising processing on the response data, and then use the bisection peak-seeking algorithm and the local inflection point algorithm to extract data features;

[0085] S33: Take the data features as input data and input them into the gas classification model. The gas classification model compares the current data features with the known response feature data, and then identifies and outputs the gas classification information.

[0086] As an optimization, in S15 of Step 1, 70% of the gas dataset is divided into the training set, 15% of the gas dataset is divided into the test set, and 15% of the gas dataset is divided into the validation set.

[0087] In order to better avoid the loss of important feature information, in S16 of Step 1, during the process of expanding the sample dimension using the sample interpolation algorithm, the dimension is expanded to 2000.

[0088] The present invention proposes a gas classification method based on hybrid waveform modulation technology. By using multi-waveform modulation technology to drive the heating elements of gas sensors in the MEMS sensor array, the sensitivity of semiconductors to temperature can be fully utilized, enabling the MEMS gas sensors to operate under different temperature conditions. As a result, the data features for distinguishing gas types contained in the response process of the gas sensors are rich enough, effectively increasing the feature dimension of the gas response data and significantly reducing the complexity of the network model. The voltage drive signal adopted is an adjustable amplitude periodic signal, and the response data of the MEMS gas sensors are collected under the driving conditions of each voltage drive signal, which can effectively capture the reaction of the sensors to the gas under specific heating conditions and can be used as gas sample data. By analyzing these data, different types of gases can be detected and classified. In this way, the effect of increasing the sensors can be achieved without increasing the actual number of sensors, and the enhancement of the source data can be realized at the source of data generation, greatly increasing the information volume of the gas data and obtaining a more accurate gas classification effect. In each data collection process of the present invention, multiple waveform drive signals are used to generate multiple voltage drive signals. Thus, it is convenient to increase the data features as much as possible by the method of superposition heating with multiple waveforms. During the feature extraction process, the dichotomy peak-seeking algorithm and the local inflection point algorithm are used to extract data features. By setting appropriate thresholds, peaks, valleys, and inflection points can be identified, enhancing the feature extraction process and providing a more comprehensive representation of the sensor response. Compared with traditional technologies, the present invention further increases the absorption peaks of the gas and improves the reaction rate of the gas, providing more data features for the subsequent classification model and significantly reducing the complexity of the classification model, achieving the effect of reducing the complexity of the subsequent MLP neural network structure. At the same time, by using the sample interpolation algorithm to expand the sample dimension, the present invention can effectively prevent the actual peak from being filtered out as noise in the subsequent filter bank process, avoiding the loss of important feature information and further ensuring the accuracy of the classification model classification. In addition, during the construction of the classification model, the weights and biases are stored in the storage module in the PS part, and at the same time, the PL part calls the DMA for high-speed data transmission, which can significantly reduce the utilization rate of the FPGA and thus reduce the power consumption. This technical means can effectively balance the hardware structure and the complexity of the algorithm, which is beneficial to improving the target accuracy of the classification task.

[0089] The implementation process of this method is simple and has a high degree of intelligence. By changing the heating conditions of the gas sensors under limited resource conditions, more data on gas responses can be obtained, thereby increasing the diversity of the data set. More abundant data can obtain a classification model with higher classification accuracy, enabling more accurate identification of different gases or gas mixtures.

Claims

1. A gas classification system based on a MEMS sensor array, comprising a MEMS sensor array (5), a voltage follower, an ADC module (4), a DAC module (2), a power amplifier module (3) and a microcontroller (1); The MEMS sensor array (5) comprises a carrier substrate (7), a plurality of heating elements mounted in an array on the carrier substrate (7), and a plurality of MEMS gas sensors (6) mounted correspondingly on the plurality of heating elements; A plurality of voltage followers are correspondingly connected to a plurality of MEMS gas sensors (6); The input end of the microcontroller (1) is connected to a plurality of voltage followers via an ADC module (4), and the output end of the microcontroller (1) is connected to a plurality of heating elements via a DAC module (2) and a power amplifier module (3) in sequence.

2. A gas classification system based on a MEMS sensor array according to claim 1, characterized in that: The microcontroller (1) is a ZYNQ chip, which comprises a PS part and a PL part. The PS part comprises an ARM dual-core processor and a DDR storage module, and the PL part comprises an FPGA. The ARM dual-core processor and the FPGA are connected via an AXI bus and perform data exchange.

3. A gas classification system based on a MEMS sensor array according to claim 2, characterized in that: The ZYNQ chip is Xilinx zynq7020, the ARM dual-core processor module is Cortex-A9 ARM core; and the DDR storage module is DDR3.

4. A gas classification system based on a MEMS sensor array according to claim 2, characterized in that: The MEMS gas sensor (6) is a carbon monoxide sensor or a hydrogen sensor.

5. A gas classification system based on a MEMS sensor array according to claim 4, characterized in that: The model of the carbon monoxide sensor is GM-702B.

6. A gas classification system based on a MEMS sensor array according to claim 4, characterized in that: The model of the hydrogen sensor is GMV-2021B.

7. A gas classification system based on a MEMS sensor array according to claim 4, characterized in that: It also includes a low-pass filter module, which is arranged between the DAC module (2) and the power amplification module (3).

8. A gas classification method based on a MEMS sensor array, using a gas classification system based on a MEMS sensor array as claimed in claim 7, characterized in that: The steps include: Step 1: Obtain gas data set; S11: placing the MEMS sensor array (5) in a test chamber, and preparing a mixed gas sample according to set component concentration information; S12: The microcontroller (1) uses a multi-waveform modulation technology to generate k voltage drive signals based on k preset waveform drive signals, wherein the voltage drive signal is an amplitude-adjustable periodic signal, and a unique drive state identification signal is set for each voltage drive signal, and then sent to the DAC module (2). The DAC module (2) sends the received k voltage drive signals to the low-pass filter module; the low-pass filter module filters the voltage drive signals and sends them to the power amplifier module (3); the power amplifier module (3) amplifies the input voltage drive signals and outputs them to the heating element, so as to drive the corresponding heating element to perform heating operation using different voltage drive signals, so that the corresponding MEMS gas sensor (6) works under different temperature conditions; when the working temperature condition of the MEMS gas sensor (6) is stable, a mixed gas prepared with component concentration data is injected into the measuring chamber; S13: causing the plurality of MEMS gas sensors (6) to respond under mixed gas conditions with set component concentration data; when the plurality of MEMS gas sensors (6) reach a steady state, using a voltage follower to collect voltage signals of corresponding MEMS gas sensors (6) in real time and sending them to an ADC module (4); the ADC module (4) converts the received analog voltage signals into digital voltage signals and then sends the digital voltage signals to a microcontroller (1); the microcontroller (1) obtains response data of the MEMS gas sensor (6) based on the received digital voltage signals; the microcontroller (1) synchronously records the drive state identification signal and the corresponding response data as gas original sample data, adds corresponding gas labels, and stores them in a storage module; S14: injecting clean air into the measuring chamber and displacing the mixed gas to the outside until the measuring chamber is filled with clean air; S15: Repeat S12 to S14 multiple times to obtain a large amount of gas raw sample data; S16: De-noising the original gas sample data, and then separating the original gas data according to the driving state identification signals of different voltage driving signals, and obtaining the response data under each voltage driving signal after separation; for the separated response data, using the binary peak search algorithm and the local inflection point algorithm to extract data features, and then using the sample interpolation algorithm to expand the sample dimension, and then obtaining the characteristic curve through feature point fitting; merging different characteristic curves belonging to the same response process to form the characteristic sample data of a single gas, and generating a set of characteristic sample data corresponding to multiple gas labels to form a gas data set, thereby completing the construction of the gas data set; S17: Divide the gas dataset into a training set, a test set, and a validation set according to a set ratio; Step 2: Construct a gas classification model; S21: constructing a classification model based on MLP in the microcontroller (1), wherein weights and biases of the classification model are set in a storage module of the PS part, and parts of the classification model other than the weights and biases are set in the PL part, and the PL part communicates with the storage module via DMA; S22: train the classification model using the training set. During the training process, the weights and biases are continuously updated through the back propagation algorithm to reduce the value of the loss function and realize the process of model parameter tuning. Then, the trained classification model is tested using the test set and verified using the verification set, and finally a gas classification model is obtained. Step 3: Classify the gas; S31: placing the MEMS gas sensor (6) in an environment to be tested, and obtaining response data of the MEMS gas sensor (6) to the gas in the environment to be tested; S32: firstly perform denoising on the response data, and then use the binary peak search algorithm and the local inflection point algorithm to extract data features; S33: Input the data features as input data into the gas classification model. The gas classification model compares the current data features with known response feature data, and then identifies and outputs gas classification information.

9. A gas classification method based on a MEMS sensor array according to claim 7, characterized in that: In step 1 S15, 70% of the gas data set is divided into a training set, 15% of the gas data set is divided into a test set, and 15% of the gas data set is divided into a validation set.

10. A gas classification method based on a MEMS sensor array according to claim 7, characterized in that: In step 1 S16, in the process of expanding the sample dimension using the sample interpolation algorithm, the dimension is expanded to 2000.

Images