A deep neural network gas identification method and system based on an internet-of-things electronic nose

Abstract

The application discloses a kind of deep neural network gas identification method and system based on internet of things electronic nose, gas sample response signal obtained by gas sensor array collection is sent to cloud data center by terminal equipment with sensor array and internet of things module.In the pre-processing of data to microcontroller MCU, gram angle field transformation is carried out, so that the two-dimensional sensor response data received by host computer becomes three-dimensional data that can be input into convolutional neural network after dimensionality increasing, and the gas label is output through classifier, which can represent gas category and concentration grade, realizes the real-time, anti-interference and high detection accuracy of remote monitoring.Gives play to the advantages of strong feature extraction ability, fast model convergence and high recognition accuracy of convolutional neural network.It can be widely applied in environmental detection, industrial production, medical treatment and safety fields.

Description

Technical Field

[0001] This invention relates to the field of gas analysis, and more particularly to a deep neural network gas recognition method based on an Internet of Things (IoT) electronic nose. Background Technology

[0002] With societal development, people's daily lives are increasingly reliant on natural gas. While gas, such as methane and carbon monoxide, brings convenience, it also poses hidden dangers due to leaks and explosions. Therefore, the detection and identification of combustible gases is receiving increasing attention. In recent years, driven by advancements in high technology, combustible gas detection technology has also seen significant development. "Electronic nose" technology, also known as machine olfaction technology, is a computer-aided sensor array technology that has emerged with the development of pattern recognition technology. First proposed by Gardner in the 1990s, it plays a crucial role in numerous fields, including food safety, environmental monitoring, and medicine. Compared to conventional gas chromatography-mass spectrometry (GC-MS) analysis methods, it is faster, simpler, and less expensive, showing great promise. There are many types of gas sensors available today. Among them, metal-oxide semiconductor (MOS) sensors have been widely used in the field of gas detection due to their low cost, small size, high sensitivity and long service life. However, MOS sensors have poor selectivity and exhibit cross-sensitivity to multiple gases. Identifying multiple gases based on a single sensor without other sensing technologies remains a huge challenge.

[0003] In this context, employing computer-aided methods is crucial. Currently, research on the classification of mixed gases is extensive. A 2019 paper published in *Sensors*, "Combustible Gas Classification Modeling using Support Vector Machine and Pairing Plot Scheme," constructed a supervised learning method for gas classification based on support vector machines. The proposed model achieved 100% accuracy in classifying all concentration levels of CH4 and CO gases above the sensor's minimum detection limit. Since the time-series response signals of gas sensors are difficult to process using traditional gas identification algorithms, a manual feature extraction step is required to enable machine learning algorithms to process the signals. However, this step requires repeated verification of the effectiveness of the extracted features, making it complex and tedious. A 2020 paper published in *Sensors and Actuators A: Physical*, "An optimized deep convolutional neural network for dendrobium classification based on electronicnose," proposed using a one-dimensional deep convolutional neural network to automatically extract features; however, this method does not utilize the multi-dimensional time-series signal correlation of the sensor array response signals. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by enabling real-time detection and early warning of gas type and concentration level within a specified area. To address the complex steps required by traditional machine recognition methods, such as manual feature extraction, this invention proposes a deep residual convolutional neural network algorithm that combines Gram angle fields to achieve dimensionality enhancement of the response signal and automatic feature extraction.

[0005] The objective of this invention is achieved through the following technical solution: a deep neural network gas recognition method based on an IoT electronic nose, characterized in that the method includes the following steps:

[0006] Step 1: Based on the sensor array terminal device with integrated network communication module, the collected gas sample response data is transmitted to a remote cloud data center via network protocol;

[0007] Step 2: Combining the deep residual convolutional neural network algorithm of Gram Corner Field (GAF), perform data upscaling on the sensor response data in the cloud data center to serve as input data for the deep residual convolutional neural network; complete the identification of the type and concentration level of the gas to be measured.

[0008] 2. The deep neural network gas recognition method based on an IoT electronic nose according to claim 1, characterized in that information interaction with the cloud data center is completed through the network communication module on the terminal device; specifically including:

[0009] (1) The terminal device can upload data packets containing gas concentration to the cloud data center and can parse the instructions issued by the data center;

[0010] (2) Parse and generate data frames; The terminal device parses the data information in the data frames returned by the sensor array and generates upload data frames with timestamp information;

[0011] (3) Receive parameter calibration from the host computer; gas data is received via USB serial port to receive configuration parameters from the host computer software, including the IP address and port number of the cloud data center, the terminal device address and group number, as well as the transmission mode and interval time;

[0012] (4) Global Navigation Satellite System (GNSS) positioning; by returning the geographical location information of the sensor terminal device in real time, the target gas concentration in a specific area can be monitored;

[0013] (5) Simultaneously collect the temperature, humidity, and air pressure of the hazardous source gas environment and send them to the microcontroller unit (MCU) on the sensor terminal device.

[0014] (6) Save working parameters; The terminal device saves the working parameters so that the sensor array can continue to collect gas sample response data.

[0015] 3. The deep neural network gas recognition method based on an IoT electronic nose according to claim 2, characterized in that the data upscaling operation on the sensor response data in the cloud data center includes the following steps:

[0016] (1) The gas sample data collected by the sensor array is sent to the MCU after signal amplification and AD sampling preprocessing;

[0017] (2) The microcontroller unit (MCU) preprocesses the raw gas sample data and performs Gram angle field transformation as input data for the deep residual convolutional neural network, specifically including:

[0018] (2.1) Since the baseline of each sensor response curve is different, the original sensor response data received by the host computer is represented as a two-dimensional matrix X, where the data in the i-th row and j-th column is the response value of the j-th sampling point of the i-th sensor. The data in the i-th row needs to be subtracted from the sensor baseline according to the following formula to obtain matrix X′. This matrix represents the matrix obtained by normalizing the minimum value of each row of the original sensor response data matrix to 0.

[0019] x′ ij =x ij -x i (baseline),x′ ij ∈X′

[0020] Where i represents the number of sensors, and j represents the j-th response signal of each sensor;

[0021] (2.2) Find the maximum value of X′, and divide each element of X′ by the maximum value to obtain the matrix. This matrix represents the matrix obtained by normalizing the minimum value of each row of the original sensor response data matrix to 0, and then reducing its numerical range to [0,1].

[0022]

[0023] (2.3) Using matrices Find φ from the elements in ij , Then, the Gramian Summation Angular Field (GASF) matrix of each sensor response data is calculated:

[0024]

[0025] Where i represents the data in the i-th row (i.e., the number of sensors), and n represents the number of response signals from each sensor.

[0026] (2.4) For each gas sample, a Gram transformation is performed, and the dimension of each sample becomes i×j×j; after the Gram angle field transformation is performed on each odor sample, it is used as the input feature map matrix of the deep residual convolutional neural network.

[0027] 4. The deep neural network gas recognition method based on an IoT electronic nose according to claim 1, characterized in that the deep residual convolutional neural network for recognizing and detecting the gas to be tested is as follows:

[0028] (1) Construct a deep residual convolutional neural network for gas identification; the deep residual convolutional neural network consists of multiple residual blocks, each of which consists of multiple convolutional layers and a skip connection channel; the calculation formula used in the convolutional layers is as follows:

[0029]

[0030] (i,j)∈{0,1,…,D l+1}

[0031]

[0032] Z l and Z l+1 Let Z represent the feature map matrices of the l-th and (l+1)-th convolutional layers, respectively. l+1 (i,j) represents the pixel value in the i-th row and j-th column of the feature map matrix of the (l+1)-th convolutional layer. Let D represent the kernel weight of the (l+1)th convolutional layer with the k-th kernel, b be the kernel bias, and D be the weight of the convolutional kernel. l+1 For Z l+1 Size, D l For Z l The dimensions of the feature map are the same, with the length and width being equal, corresponding to the pixel value in the i-th row and j-th column of the feature map, K. l s0 is the number of channels in the feature map, p is the number of layers with zero padding, f is the size of the convolution kernel, and s0 is the convolution stride. During the calculation of the convolutional layer, the convolution kernel sweeps across the input features, performs element-wise matrix multiplication and summation on the input features within the receptive field (the size of the convolution kernel), and adds the bias to calculate the output feature map of the convolutional layer.

[0033] (2) The classifier of the deep residual deep neural network outputs a classification code to represent the gas category and concentration level to be identified, thus completing the identification of the gas to be tested.

[0034] 5. The deep neural network gas recognition method based on IoT electronic nose according to claim 4, characterized in that the classifier of the deep residual deep neural network uses k-hot encoding to convert the gas category value and concentration level into a binary vector, and assigns each specific category and concentration level of gas sample to only one encoded category, that is, only the corresponding feature is 1, and the rest are 0.

[0035] 6. A deep neural network gas recognition method based on an IoT electronic nose according to claim 4, characterized in that: the first convolutional layer in the residual block uses m convolutional kernels of size 3×3 with a stride of 2, which can reduce the size of the feature map; for the k-th residual block, its first convolutional layer has m = 2×k + 5, where k∈[1,5]; the second convolutional layer in the residual block uses m / 2 convolutional kernels of size 1×1, which can reduce the number of channels in the feature map and increase the nonlinearity of the network; the third convolutional layer in the residual block uses m convolutional kernels of size 3×3 to recover the number of channels in the feature map; the stride of the convolutional kernels in the second and third convolutional layers is 1; skip connections are used to add the outputs of the first and third convolutional layers in the residual block, so that the network can learn the residual instead of the entire mapping function during training.

[0036] 7. A deep neural network gas recognition system based on an IoT electronic nose, which implements the deep neural network gas recognition method based on an IoT electronic nose as described in any one of claims 1-6, characterized in that the system consists of a gas sensor module, a gas parameter acquisition module, a host computer algorithm module, an IoT communication module, and a supporting remote IoT cloud platform.

[0037] The gas sensor module is used to select the sensor model of the sensor array according to the sensor manual of the target gas to be detected, send the gas sample signal detected by the sensor array to the gas parameter acquisition module, and transmit the temperature, humidity and pressure signals of the sensor to the host computer algorithm module.

[0038] The gas parameter acquisition module is used to send the gas sample signal to the host computer algorithm module via an AD sampling circuit;

[0039] The host computer algorithm module is used to perform algorithm processing through a microcontroller, identify the target gas, and send the results to the Internet of Things communication module.

[0040] The IoT communication module is used to send the temperature, humidity and air pressure signals from the sensor, along with the target gas identification results processed by the host computer algorithm module, to the IoT cloud platform to achieve remote monitoring and early warning.

[0041] In summary, the technical solutions conceived in this invention have the following beneficial effects compared with the prior art:

[0042] 1. This invention combines a traditional electronic nose system based on a sensor array with an Internet of Things (IoT) cloud platform, taking advantage of the speed, simplicity, and low cost of the electronic nose system, as well as the remoteness and real-time capability of the IoT cloud platform.

[0043] 2. This invention enables automatic feature extraction and odor identification from time-series data output by sensors without sacrificing effective features. Compared to other algorithms, this algorithm has a simpler training process, higher feature extraction efficiency, and higher odor identification accuracy.

[0044] 3. Unlike traditional dimensionality reduction-based data processing methods, GAF transforms inconspicuous one-dimensional features into obvious two-dimensional features, converting two-dimensional sensor response curve data into three-dimensional data, thereby increasing the dimensionality of the data and laying a good data foundation for the feature extraction steps of subsequent deep residual convolutional neural networks.

[0045] 4. The features extracted by this invention have excellent robustness to drift, which can effectively improve the accuracy of gas detection and identification under drift and extend the service life of the equipment. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a system structure diagram of a deep neural network gas recognition method based on an IoT electronic nose provided by the present invention.

[0048] Figure 2 This is a diagram of a deep neural network model. Detailed Implementation

[0049] The following description, in conjunction with the accompanying drawings, further illustrates embodiments of the present invention.

[0050] This invention provides a deep neural network gas recognition system based on an IoT electronic nose. The system structure diagram of the IoT electronic nose is attached. Figure 1 As shown, the system consists of a gas sensor module, a gas parameter acquisition module, a host computer algorithm module, an IoT communication module, and a supporting remote IoT cloud platform.

[0051] The gas sensor module is used to detect the concentration range of the target gas and interfering gases as needed. Based on the sensor datasheets, the MP series sensors from Zhengzhou Weisheng Electronics Technology Co., Ltd. and the TGS series sensors from FIGARO Corporation of Japan were selected. The selected MP series sensors include: MP-4 (sensitive to CH4), MP-9 (sensitive to both CH4 and CO), and MP503 (sensitive to CH2O). The selected TGS series sensors include: TGS821 (sensitive to H2), TGS816 (sensitive to both CH4 and H2), and TGS2602 (for air pollutant detection). For convenience, MP-9, MP-4, MP503, TGS821, TGS816, and TGS2602 are numbered S1-S6 respectively. Furthermore, considering the significant impact of temperature and humidity on the sensors, and to reduce circuit complexity, the SHT20 digital temperature and humidity sensor from Sensirion of Switzerland was selected as the temperature and humidity sensor in the sensor array. The gas sensor module sends the gas sample signal detected by the sensor array to the gas parameter acquisition module; and transmits the temperature, humidity and pressure signals of the sensor to the host computer algorithm module.

[0052] The gas parameter acquisition module is used to send the gas sample signal to the host computer algorithm module via the AD sampling circuit. The AD conversion chip is not an on-chip built-in chip, but an off-chip ADC128S052 is selected because the sensor output is in the range of 0 to 5V, so this chip can meet the requirements of sensor voltage detection.

[0053] The host computer algorithm module is used to process algorithms via a microcontroller, identify the target gas, and send the results to the IoT communication module. The core control chip of the host computer algorithm module's hardware system is the MSP430F169 microcontroller manufactured by Texas Instruments (TI). This chip is a 16-bit ultra-low-power microprocessor with abundant on-chip peripherals, including AD converters, DA converters, timers, USART, comparators, hardware multipliers, and an internal temperature sensor. Utilizing its I / O ports, 60kB Flash memory, and USART resources, the electronic nose hardware system can perform signal processing, data storage, and serial communication functions.

[0054] The IoT communication module transmits the temperature, humidity, and air pressure signals from the sensor, along with the target gas identification results processed by the host computer algorithm module, to the IoT cloud platform for remote monitoring and early warning. The IoT communication module utilizes the ME3616 NB-IoT module with a built-in BeiDou / GPS satellite positioning chip, developed by Gosuncn Technology. This module integrates dual core functions of "NB-IoT + satellite positioning" and is designed for low-speed, low-power, long-distance, and massively connected IoT applications. It supports multiple network protocols (CoAP, TCP / UDP, MQTT, OMA-LWM2M) and multiple low-power modes (PSM, eDRX), and also supports GNSS positioning. These protocols and functions enable its application in various IoT and M2M scenarios, such as smart meters, smart parking, asset tracking, and wearable devices.

[0055] The IoT cloud platform utilizes China Mobile's OneNET IoT platform, which supports and adapts to various network environments and protocol types. It enables rapid access to various sensors and smart hardware, providing rich APIs and application templates to support the development of various industry applications and smart hardware. This effectively reduces the development and deployment costs of IoT applications and meets the platform-level service needs of the IoT field, including device connectivity, protocol adaptation, data storage, data security, and big data analytics. Currently, OneNET has built an overall "cloud-network-edge-device" IoT architecture, extending downwards to terminal adaptation and access capabilities and upwards to integrate segmented industry applications. It provides basic device management capabilities such as device access and device management, as well as applications such as location positioning, remote upgrades, data visualization, and message queuing.

[0056] Based on relevant national documents, namely the "National Standard of the People's Republic of China for Indoor Air Quality" and the "National Standard of the People's Republic of China for Design Specifications of Detection Alarms for Combustible and Toxic Gases in the Petrochemical Industry," and in conjunction with the sensor datasheet, the detection ranges for target and interfering gases are set as follows: CH4 concentration range 0–10000 ppm, CO concentration range 0–1000 ppm, CH2O concentration range 0–50 ppm, and H2 concentration range 0–100 ppm. Target gases CH4 and CO are divided into six concentration levels within their respective ranges, increasing from Level 1 to Level 6 from left to right. Since formaldehyde can be considered as both a target and interfering gas in this study, its concentration levels are also divided into six levels accordingly. The selection of interfering gases is primarily based on the sensor datasheet; most gas sensors will show a significant response to H2.

[0057] The terminal's gas sensor module transmits collected data such as temperature, humidity, and air pressure directly to the terminal device's microcontroller (MCU) via the I2C bus. After signal amplification and AD sampling of the gas samples, the response signal is transmitted to the terminal device's MCU via the SPI bus. The final total number of samples is 411. All samples are divided into four groups based on whether they contain interfering gas samples and their types: a no-interference group (183 samples) and a mixed-interference group (183 + 228 samples). For the mixed-interference group, samples containing interfering gases account for 55.47% of the total samples, and samples containing two or more target gases or interfering gases account for 89.27% ​​of the total samples. This proportion is primarily to give the electronic nose a stronger ability to identify mixed gases, rather than just detecting a single gas. Algorithms are then used to process and identify the type and concentration level of combustible gases.

[0058] The raw gas sample data undergoes preprocessing and Gram angle field transformation as input data for the convolutional neural network, specifically including:

[0059] 1. Because the baseline of each sensor response curve is different, the raw sensor response data received by the host computer is represented as a two-dimensional matrix X, where the data in the i-th row and j-th column is the response value of the j-th sampling point of the i-th sensor. The sensor baseline needs to be subtracted from the data in the i-th row according to the following formula to obtain matrix X′. This matrix represents the matrix obtained by normalizing the minimum value of each row of the raw sensor response data matrix to 0.

[0060] x′ ij =x ij -x i (baseline),x′ ij ∈X′

[0061] (where 1≤i≤6, 1≤j≤170)

[0062] 2. Find the maximum value of X′, and divide each element of X′ by this maximum value to obtain the matrix. This matrix represents the matrix obtained by normalizing the minimum value of each row of the original sensor response data matrix to 0, and then reducing its numerical range to [0,1].

[0063]

[0064] 3. Using matrices Find φ from the elements in ij , Then, the Gramian Summation Angular Field (GASF) matrix of each sensor response data is calculated.

[0065]

[0066] 4. For each odor sample, since Gram transform has been performed, the dimension of each sample becomes 6×170×170. After Gram angular field transformation, each odor sample is used as the input feature map matrix of the deep residual convolutional neural network.

[0067] 5. Construct a deep convolutional network for gas identification. The network mainly consists of 5 residual blocks, each of which comprises 3 convolutional layers and one skip connection channel. The network structure of the algorithm model is attached. Figure 2 As shown, Figure 2 (a) in the diagram represents the structure of a deep residual convolutional neural network; Figure 2 (b) in the diagram represents the structure of the residual block. The calculation formula used in the convolutional layer is as follows:

[0068]

[0069] (i,j)∈{0,1,…,D l+1} Z l and Z l+1 Let Z represent the feature map matrices of the l-th and (l+1)-th convolutional layers, respectively. l+1 (i,j) represents the pixel value in the i-th row and j-th column of the feature map matrix of the (l+1)-th convolutional layer. Let D represent the kernel weight of the (l+1)th convolutional layer with the k-th kernel, b be the kernel bias, and D be the weight of the convolutional kernel. l+1 For Z l+1 Size, D l For Z l The dimensions of the feature map are the same, with the length and width being equal, corresponding to the pixel value in the i-th row and j-th column of the feature map, K. l s0 is the number of channels in the feature map, p is the number of layers with zero padding, f is the size of the convolution kernel, and s0 is the convolution stride. During the calculation of the convolutional layer, the convolution kernel sweeps across the input features, performs element-wise matrix multiplication and summation on the input features within the receptive field (the size of the convolution kernel), and adds the bias to calculate the output feature map of the convolutional layer.

[0070] 6. A deep neural network classifier outputs classification codes for the gas category and concentration level to complete the gas identification. Specifically, in the output binary vector, only the subscripts for the gas category and concentration level corresponding to the sample are 1, while the rest are 0. This algorithm achieves detection accuracies of 89.04% for CO and 90.03% for CH4 in the presence of interfering gases, demonstrating good anti-interference capabilities and detection accuracy.

[0071] 7. The ME3616 IoT module serializes and packages the algorithm recognition results and the concentration response data of the sensor array and then transmits them to the data center of the cloud platform (this case uses the OneNet cloud platform).

[0072] 8. After the IoT cloud platform completes the identification of the gas type and concentration level, it combines the location information uploaded by the sensor terminal to promptly take response measures such as gas alarm, helping managers to complete online early warning and monitoring of gas in designated areas.

[0073] This invention proposes a deep neural network gas identification method based on an IoT electronic nose. For identifying the types and concentrations of mixed gases in complex environments, it uses a deep residual convolutional network (CNN) with data dimensionality upscaling via Gram angle field analysis for classification. Even in the presence of interfering gases, the detection accuracy for CO and CH4 reaches 89.04% and 90.03%, respectively, demonstrating good anti-interference performance. This indicates that the algorithm can enhance the dimensionality of the data, amplify the differences between different sample data, and fully utilize the multi-channel feature extraction capabilities of CNNs. The algorithm adapts well to changes in sensor response data and boasts advantages such as strong feature extraction capabilities, fast model convergence, and high recognition accuracy. Furthermore, this invention integrates network communication between the terminal sensing device and the cloud platform, providing technical support for online detection and early warning of hazardous sources in communities, and has broad application prospects.

[0074] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A deep neural network gas recognition method based on an IoT-enabled electronic nose, characterized in that, The method includes the following steps: Step 1: Based on the sensor array terminal device with integrated network communication module, the collected gas sample response data is transmitted to a remote cloud data center via network protocol; Step 2: Perform data upscaling on the sensor response data in the cloud data center, including the following steps: (1) The gas sample data collected by the sensor array is amplified and preprocessed by AD sampling before being sent to the MCU; (2) The microcontroller unit (MCU) preprocesses the gas sample data and performs Gram angle field transformation as input data for the deep residual convolutional neural network, specifically including: (2.1) Represent the raw sensor response data received by the host computer as a two-dimensional matrix X, where the data in the i-th row and j-th column is the response value of the j-th sampling point of the i-th sensor. Subtract the sensor baseline from the data in the i-th row to obtain the matrix. This matrix represents the matrix obtained by normalizing the minimum value of each row of the original sensor response data matrix to 0; Where i represents the number of sensors, and j represents the j-th response signal of each sensor; (2.2) Find The maximum value, and Divide each element in the matrix by the maximum value to obtain the matrix. This matrix represents the normalization of the minimum value of each row of the original sensor response data matrix to 0, and the reduction of the numerical range to [0,1]. (2.3) Using matrices Find the elements in , Then, the Gram summation angle field GASF matrix of each sensor response data is calculated: Where i represents the data in the i-th row, i.e. the number of sensors, and n represents the number of response signals from each sensor; (2.4) For each gas sample, a Gram transformation is performed, and the dimension of each sample becomes i×j×j; after the Gram angular field transformation is performed on each odor sample, it is used as the input feature map matrix of the deep residual convolutional neural network. The classification and concentration level of the gas to be tested are identified using a deep residual convolutional neural network, as detailed below: (1) Construct a deep residual convolutional neural network for gas identification; the deep residual convolutional neural network consists of multiple residual blocks, each of which consists of multiple convolutional layers and a skip connection channel; the calculation formula in the convolutional layer is as follows: in and Let represent the feature map matrices of the l-th and (l+1)-th convolutional layers, respectively. This represents the pixel value in the i-th row and j-th column of the feature map matrix of the (l+1)-th convolutional layer. Let represent the kernel weights of the (l+1)th convolutional layer with the k-th kernel, and b be the kernel bias. for Size, for The dimensions are such that the length and width of the feature map are the same, corresponding to the pixel value in the i-th row and j-th column of the feature map. denoted as the number of channels in the feature map, p as the number of zero-padding layers, f as the size of the convolution kernel, and s0 as the convolution stride. During the calculation of the convolutional layer, the convolution kernel sweeps across the input features, performs element-wise matrix multiplication and summation on the input features within the size of the convolution kernel, and adds the bias to calculate the output feature map of the convolutional layer. (2) The classifier of the deep residual deep neural network outputs a classification code to represent the gas category and concentration level to be identified, thus completing the identification of the gas to be tested.

2. The deep neural network gas recognition method based on an IoT electronic nose according to claim 1, characterized in that, Information exchange with the cloud data center is completed through the network communication module on the terminal device; specifically including: (1) The terminal device can upload data packets containing gas concentration to the cloud data center and can parse the instructions issued by the data center; (2) Parse and generate data frames; the terminal device parses the data information in the data frames returned by the sensor array and generates upload data frames with timestamp information; (3) Receive parameter calibration from the host computer; gas data is received via USB serial port to receive configuration parameters from the host computer software, including the IP address and port number of the cloud data center, the terminal device address and group number, as well as the transmission mode and interval time; (4) Global Navigation Satellite System (GNSS) positioning; by returning the geographic location information of the sensor terminal device in real time, the target gas concentration in a specific area can be monitored; (5) Simultaneously collect the temperature, humidity, and air pressure of the hazardous source gas environment and send them to the microcontroller unit (MCU) on the sensor terminal device; (6) Save working parameters; The terminal device saves the working parameters so that the sensor array can continue to collect gas sample response data.

3. The deep neural network gas recognition method based on an IoT electronic nose according to claim 1, characterized in that, The classifier of the deep residual deep neural network uses k-hot encoding to convert the gas category value and concentration level into a binary vector. Each gas sample of a specific category and concentration level corresponds to only one encoded category, that is, only the corresponding feature is 1, and the rest are 0.

4. The deep neural network gas recognition method based on an IoT electronic nose according to claim 1, characterized in that, The first convolutional layer in the residual block uses m 3×3 kernels with a stride of 2. This convolutional layer reduces the size of the feature map. For the k-th residual block, its first convolutional layer has m = 2×k+5, where k∈[1, 5]. The second convolutional layer in the residual block uses m / 2 1×1 kernels. This convolutional layer reduces the number of channels in the feature map and increases the nonlinearity of the network. The third convolutional layer in the residual block uses m 3×3 kernels to recover the number of channels in the feature map. The stride of the kernels in the second and third convolutional layers is 1. Skip connections are used to add the outputs of the first and third convolutional layers in the residual block, so that the network can learn the residual instead of the entire mapping function during training.

5. A deep neural network gas recognition system based on an IoT electronic nose, implementing the deep neural network gas recognition method based on an IoT electronic nose according to any one of claims 1-4, characterized in that, The system consists of a gas sensor module, a gas parameter acquisition module, a host computer algorithm module, an IoT communication module, and a supporting remote IoT cloud platform. The gas sensor module is used to select the sensor model of the sensor array according to the sensor manual of the target gas to be detected, send the gas sample signal detected by the sensor array to the gas parameter acquisition module, and transmit the temperature, humidity and pressure signals of the sensor to the host computer algorithm module. The gas parameter acquisition module is used to send the gas sample signal to the host computer algorithm module via an AD sampling circuit; The host computer algorithm module is used to perform algorithm processing through a microcontroller, identify the target gas, and send the results to the Internet of Things communication module. The IoT communication module is used to send the temperature, humidity and air pressure signals from the sensor, along with the target gas identification results processed by the host computer algorithm module, to the IoT cloud platform to achieve remote monitoring and early warning.

Images