Rapid detection and classification method for potherb mustard producing areas

Abstract

The invention provides a rapid detection and classification method for potherb mustard producing areas, the method is based on an electronic nose system, the electronic nose system comprises an air pump, an air valve, a sensor array air chamber, an equipment acquisition card and an upper computer, the air chamber is internally provided with a sensor array, and the method comprises the following steps: S1, data acquisition; obtaining potherb mustard gas sample data through a sensor array in the electronic nose system; s2, preprocessing the data; s3, inputting into a WVCTS model for training; and S4, classifying the producing areas through a WVCTS model. The whole system is miniaturized and can be conveniently used in different scenes, meanwhile, operation can be carried out without professional technicians, the test cost is reduced, data are input into the optimized WVCTS model after producing areas are classified, and the classification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a method for rapid detection and classification of snow vegetable origins. Background Art

[0002] Gas Chromatography-Mass Spectrometry (GC-MS) is an analytical technique that combines gas chromatography (GC) and mass spectrometry (MS). It is widely used in chemistry, environmental science, food science, forensics, drug metabolism and clinical medicine to separate, identify and quantify volatile or semi-volatile compounds in complex mixtures. The sample is first injected into the gas chromatography column, where the compounds are separated in the column under the action of the carrier gas. Different compounds interact differently, resulting in different speeds of movement in the chromatographic column. The separated compounds come out of the gas chromatography column and enter the mass spectrometer, where the compounds are ionized and decomposed into fragment ions, which are separated and detected according to their mass and charge ratios. The mass spectrometer produces a mass spectrum showing ion peaks of different mass-to-charge ratios. These spectra can be compared with the database to determine the information of the compounds and finally realize the classification of the origin of the product.

[0003] However, the existing gas chromatography-mass spectrometry technology has the problems that the instruments used are large in size and expensive, cannot achieve the effect of real-time detection, require professional technicians to operate, and the detection time is long, which consumes a lot of manpower and material resources. Summary of the invention

[0004] The main purpose of the present invention is to provide a method for rapid detection and classification of the origin of snow vegetable, hoping to effectively solve the problems in the background technology.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for rapid detection and classification of snow vegetable origins is based on an electronic nose system, the electronic nose system includes an air pump, an air valve, a sensor array air chamber, an equipment acquisition card and a host computer, the air chamber is provided with a sensor array, and includes the following steps:

[0007] S1: Data collection, to obtain snow vegetable gas sample data through the sensor array in the electronic nose system;

[0008] S2: Data preprocessing;

[0009] S3: input into the WVCTS model for training;

[0010] The WVCTS model includes a WVC module, a GRU module and a FC fully connected layer;

[0011] The WVC module is used to process the input data, and the input data is converted into three parallel data streams, the first data stream is the original data, the second data stream is processed by two layers of one-dimensional convolution layers, and the third data stream is processed by three layers of one-dimensional convolution layers; after the data processing is completed, the output is summed to obtain the final output of the WVC module;

[0012] The GRU module receives the final output of the WVC module as input and performs sequence processing; the GRU module is composed of a double-layer GRU network, the first layer is a common GRU, and the second layer is a bidirectional GRU, which jointly processes sequence data and extracts sequence features;

[0013] The FC fully connected layer receives the output of the GRU module and maps it to the required output space;

[0014] S4: Origin classification using the WVCTS model.

[0015] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions:

[0016] As a preferred technical solution of the present invention: the step S1 specifically includes the following sub-steps:

[0017] S11: Start the electronic nose system and preheat the sensor array to reach normal operating temperature;

[0018] S12: Open the gas valve and the gas pump to purge the sensor array gas chamber with pure dry nitrogen;

[0019] S13: Turn off the air pump, switch the air valve, inject snow vegetable gas, let the snow vegetable gas and sensor array fully respond, and collect gas curve data through the equipment acquisition card to save it in the host computer;

[0020] S14: Repeatedly collect snow vegetable gas sample data.

[0021] As a preferred technical solution of the present invention: the sensor array includes 16 metal oxide sensors purchased from Figaro Company of Japan, and the sensor models are TGS813-1, TGS813-2, TGS816, TGS821, TGS822-1, TGS822-2, TGS826, TGS832, TGS2600, TGS2602, TG2609, TGS2610-1, TGS2610-2, TGS2611-1, TGS2611-2, and TGS2620.

[0022] As a preferred technical solution of the present invention: the step S2 specifically includes the following sub-steps:

[0023] S21: Median filtering denoising: median filtering denoising is performed on the collected data;

[0024] S22: Determine the data sample size: Determine the size of each data sample to be S×T, where S is the number of sensors, T is the number of time points, and determine the sampling frequency to be 10 Hz, that is, 10 data points are collected per second;

[0025] S23: Data normalization: Subtract the voltage response value of each sensor from its baseline value;

[0026] S24: Geometric downsampling: The data is downsampled in a geometric progression and compressed in the time dimension. Finally, the size of each sample data is S×100.

[0027] As a preferred technical solution of the present invention: in step S22, the number S of sensors should be no less than 3.

[0028] As a preferred technical solution of the present invention: the step S3 specifically includes the following sub-steps:

[0029] S31: Create WVC module: Create three parallel data streams. The first branch consists of the original input and does not require any processing. The second branch is two layers of one-dimensional convolution layers. The first layer parameters are: input channels = 16, output channels = 96, convolution kernel = 1, stride = 1, and padding is 'consistent'. The second layer parameters are: input channels = 96, output channels = 32, convolution kernel = 3, stride = 1, and padding = 'consistent'. The third branch is three layers of one-dimensional convolution layers. The first layer parameters are: input channels = 16, output channels = 96, convolution kernel = 1, stride = 1, and padding = 'consistent'. The second layer parameters are: input channels = 96, output channels = 96, convolution kernel = 3, stride = 1, and padding = 'consistent'. The second layer parameters are: input channels = 96, output channels = 32, convolution kernel = 1, stride = 1, and padding = 'consistent'. After passing through the three branches, sum all the outputs.

[0030] S32: Create GRU module: Create a two-layer GRU network, the first layer is a normal GRU with 80 hidden layer neurons; the second layer is a BiGRU with 64 hidden layer neurons;

[0031] S33: Connect a FC fully connected layer, then use Hardswish as the activation function, cross entropy as the loss function, and use the AdamW algorithm for network parameter optimization.

[0032] As a preferred technical solution of the present invention: a one-dimensional BatchNorm layer is connected after each convolution layer, and a Hardswish activation function is used after each convolution layer, except for the last convolution layer of each parallel route, in which the activation function is after the BatchNorm layer.

[0033] The present invention provides a method for rapid detection and classification of the origin of snow vegetable, which has the following beneficial effects: the entire system is miniaturized and convenient to use in different scenarios, and can be operated without the need for professional technicians, thereby reducing testing costs. After the origin is classified, the data is input into an optimized WVCTS model to improve the accuracy of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a schematic diagram of the electronic nose system involved in the present invention.

[0035] Figure 2 This is a structural diagram of the WVCTS model for the rapid detection and classification method of the origin of snow vegetable provided by the present invention.

[0036] Figure 3 This is the WVC module structure diagram.

[0037] Figure 4 This is the GRU module structure diagram.

[0038] Figure 5 This is the snow vegetable gas curve after data preprocessing.

[0039] Figure 6 This is the classification result of the optimized WVCTS model test set. DETAILED DESCRIPTION

[0040] The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

[0041] The experimental data collection environment of this embodiment has a room temperature of 25°C and a humidity of 60-70% RH. Each time, 5 ml of snow vegetable gas is extracted from a headspace bottle placed in a constant temperature box at 60°C for 30 minutes using the headspace extraction method and injected into the gas chamber of a homemade electronic nose detection device for detection. Each headspace bottle is sampled 3 times. Before the experiment, the electronic nose sensor needs to be preheated for 1 hour to ensure that the metal oxide sensor is at the working temperature. The detection sampling rate of the electronic nose is 10Hz. Snow vegetables are collected from six production areas, namely Jiaxing, Ningbo Zhanqi Town, Ningbo Xianxiang Town, Suzhou, Xuzhou, and Xuancheng. 120 samples are collected from each production area, and a total of 784 (redundant) data are input into the WVCTS-Net model for training.

[0042] A rapid detection and classification method for the origin of snow vegetable, based on an electronic nose system, such as Figure 1As shown, the electronic nose system includes an air pump, a three-way valve, an air chamber, an air pump, a data acquisition card and a host computer. A sensor array is arranged in the air chamber, and the following steps are included:

[0043] S1: Data collection, to obtain snow vegetable gas sample data through the sensor array in the electronic nose system;

[0044] S2: Data preprocessing;

[0045] S3: Input into the WVCTS model for training, such as Figure 2 As shown;

[0046] The WVCTS model includes a WVC module, a GRU module and a FC fully connected layer;

[0047] The WVC module is used to process the input data, and the input data is converted into three parallel data streams, the first data stream is the original data, the second data stream is processed by two layers of one-dimensional convolution layers, and the third data stream is processed by three layers of one-dimensional convolution layers; after the data processing is completed, the output is summed to obtain the final output of the WVC module;

[0048] The GRU module receives the final output of the WVC module as input and performs sequence processing; the GRU module is composed of a double-layer GRU network, the first layer is a common GRU, and the second layer is a bidirectional GRU, which jointly processes sequence data and extracts sequence features;

[0049] The FC fully connected layer receives the output of the GRU module and maps it to the required output space;

[0050] S4: Origin classification using the WVCTS model.

[0051] The sensor array includes 16 metal oxide sensors purchased from Figaro Company of Japan, which are placed in a closed gas chamber to eliminate interference from external gases. Metal oxide sensors are the most widely used type in electronic noses at present, with advantages such as stable performance, high sensitivity, small size, and cross-sensitivity and broad spectrum. The principle is that gas molecules undergo physical and chemical adsorption on the surface of metal oxides, causing changes in the conductivity of metal oxides, and converting the changes in conductivity into changes in voltage through voltage divider resistors. In actual operation, it is manifested as the concentration of Angelica gas causing changes in the sensor curve. Different sensors have their specific sensitive substances. The 16 sensors used in this invention are selected according to the volatile compound gases in Angelica, and their sensitive substances are shown in Table 1.

[0052] Step S1 specifically includes the following sub-steps:

[0053] S11: Start the electronic nose system and preheat the sensor array to reach normal operating temperature;

[0054] S12: Open the gas valve and the gas pump to purge the sensor array gas chamber with pure dry nitrogen;

[0055] S13: Turn off the air pump, switch the air valve, inject snow vegetable gas, let the snow vegetable gas and sensor array fully respond, and collect gas curve data through the equipment acquisition card to save it in the host computer;

[0056] S14: Repeatedly collect snow vegetable gas sample data.

[0057] Step S2 specifically includes the following sub-steps:

[0058] S21: Median filtering denoising: median filtering denoising is performed on the collected data;

[0059] S22: Determine the data sample size: Determine the size of each data sample as S×T, where S is the number of sensors, T is the number of time points, and the sampling frequency is determined to be 10 Hz, that is, 10 data points are collected per second, S is not less than 3, and the data curve after preprocessing is as follows Figure 5 As shown;

[0060] S23: Data normalization: Subtract the voltage response value of each sensor from its baseline value;

[0061] S24: Geometric downsampling: The data is downsampled in a geometric sequence and compressed in the time dimension. Finally, the size of each sample data is S×100. The data curve after preprocessing is as follows: Figure 5 shown.

[0062] Table 116 Sensor sensitive substances

[0063]

[0064] Step S3 specifically includes the following sub-steps:

[0065] S31: Create WVC module: Create three parallel data streams. The first branch consists of the original input and does not require any processing. The second branch is two layers of one-dimensional convolution layers. The first layer parameters are: input channel = 16, output channel = 96, convolution kernel = 1, stride = 1, and padding is 'consistent'. The second layer parameters are: input channel = 96, output channel = 32, convolution kernel = 3, stride = 1, and padding = 'consistent'. The third branch is three layers of one-dimensional convolution layers. The first layer parameters are: input channel = 16, output channel = 96, convolution kernel = 1, stride = 1, and padding = 'consistent'. The second layer parameters are: input channel = 96, output channel = 96, convolution kernel = 3, stride = 1, and padding = 'consistent'. The second layer parameters are: input channel = 96, output channel = 32, convolution kernel = 1, stride = 1, and padding = 'consistent'. After passing through the three branches, sum all the outputs, such as Figure 3 As shown;

[0066] S32: Create GRU module: Create a two-layer GRU network. The first layer is a normal GRU with 80 hidden layer neurons. The second layer is a BiGRU with 64 hidden layer neurons. Figure 4 As shown;

[0067] S33: Connect a FC fully connected layer, use cross entropy as the loss function, and use the AdamW algorithm to optimize the network parameters.

[0068] The dropout ratio is 0.15, and the activation function uses the Hardswish function. The cross entropy function can usually be used to determine the accuracy of the model in estimating the true probability distribution. The smaller the loss, the smaller the deviation between the model classification and the true category, and the more accurate the model. Therefore, cross entropy is used as the loss function. The AdamW algorithm is used for network parameter optimization. The AdamW algorithm can dynamically adjust the learning rate of each parameter. In each iteration, the learning step size of each parameter is limited to a certain range, so that the parameter value is relatively stable. The initial learning rate is set to 0.0005.

[0069] Each convolutional layer is connected to a one-dimensional BatchNorm layer, and the Hardswish activation function is used after each convolutional layer, except for the last convolutional layer of each parallel route, where the activation function is after the BatchNorm layer.

[0070] The training results are as follows Figure 6 As shown, the average accuracy was 98.49% ± 1.56.

[0071] The electronic nose system performs rapid detection and classification of the origin of Angelica sinensis based on any one of the methods for rapid detection and classification of the origin of Angelica sinensis described above.

[0072] The above-mentioned specific implementation methods are used to explain the present invention and are only preferred embodiments of the present invention, rather than limiting the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A rapid detection and classification method for the origin of potherb mustard, characterized in that: The method is based on an electronic nose system, which includes an air pump, an air valve, a sensor array air chamber, an equipment acquisition card and a host computer. The air chamber is provided with a sensor array, and includes the following steps: S1: Data collection, to obtain snow vegetable gas sample data through the sensor array in the electronic nose system; S2: Data preprocessing; S3: input into the WVCTS model for training; The WVCTS model includes a WVC module, a GRU module and a FC fully connected layer; The WVC module is used to process the input data, and the input data is converted into three parallel data streams, the first data stream is the original data, the second data stream is processed by two layers of one-dimensional convolution layers, and the third data stream is processed by three layers of one-dimensional convolution layers; After the data processing is completed, the output is summed to obtain the final output of the WVC module; The GRU module receives the final output of the WVC module as input and performs sequence processing; the GRU module is composed of a double-layer GRU network, the first layer is a common GRU, and the second layer is a bidirectional GRU, which jointly processes sequence data and extracts sequence features; The FC fully connected layer receives the output of the GRU module and maps it to the required output space; S4: Origin classification using the WVCTS model.

2. The rapid detection and classification method for the origin of potherb mustard according to claim 1, characterized in that: The step S1 specifically includes the following sub-steps: S11: Start the electronic nose system and preheat the sensor array to reach normal operating temperature; S12: Open the gas valve and the gas pump to purge the sensor array gas chamber with pure dry nitrogen; S13: Turn off the air pump, switch the air valve, inject snow vegetable gas, let the snow vegetable gas and sensor array fully respond, and collect gas curve data through the equipment acquisition card to save it in the host computer; S14: Repeatedly collect snow vegetable gas sample data.

3. The rapid detection and classification method for the origin of potherb mustard according to claim 1 or 2, characterized in that: The sensor array includes 16 metal oxide sensors purchased from Figaro Company of Japan, and the sensor models are TGS813-1, TGS813-2, TGS816, TGS821, TGS822-1, TGS822-2, TGS826, TGS832, TGS2600, TGS2602, TG2609, TGS2610-1, TGS2610-2, TGS2611-1, TGS2611-2, and TGS2620.

4. The rapid detection and classification method for the origin of potherb mustard according to claim 1, characterized in that: The step S2 specifically includes the following sub-steps: S21: Median filtering denoising: median filtering denoising is performed on the collected data; S22: Determine the data sample size: Determine the size of each data sample to be S×T, where: S is the number of sensors, T is the number of time points, and determine the sampling frequency to be 10 Hz, that is, 10 data points are collected per second; S23: Data normalization: Subtract the voltage response value of each sensor from its baseline value; S24: Geometric downsampling: The data is downsampled in a geometric progression and compressed in the time dimension. Finally, the size of each sample data is S×100.

5. The rapid detection and classification method for the origin of potherb mustard according to claim 4, characterized in that: In step S22, the number S of sensors is not less than 3.

6. The rapid detection and classification method for the origin of potherb mustard according to claim 1, characterized in that: The step S3 specifically includes the following sub-steps: S31: Create WVC module: Create three parallel data streams. The first branch consists of the original input and does not require any processing. The second branch is two layers of one-dimensional convolution layers. The first layer parameters are: input channels = 16, output channels = 96, convolution kernel = 1, stride = 1, and padding is 'consistent'. The second layer parameters are: input channels = 96, output channels = 32, convolution kernel = 3, stride = 1, and padding = 'consistent'. The third branch is three layers of one-dimensional convolutional layers. The first layer parameters are: input channels = 16, output channels = 96, convolution kernel = 1, stride = 1, padding = 'consistent'; the second layer parameters are: input channels = 96, output channels = 96, convolution kernel = 3, stride = 1, padding = 'consistent'; Second layer parameters: input channels = 96, output channels = 32, convolution kernel = 1, stride = 1, padding = uniform, after three branches, all outputs are summed; S32: Create GRU module: Create a two-layer GRU network, the first layer is a normal GRU with 80 hidden layer neurons; the second layer is a BiGRU with 64 hidden layer neurons; S33: Connect a FC fully connected layer, then use Hardswish as the activation function, cross entropy as the loss function, and use the AdamW algorithm for network parameter optimization.

7. The rapid detection and classification method for the origin of potherb mustard according to claim 6, characterized in that: Each convolutional layer is connected to a one-dimensional BatchNorm layer, and the Hardswish activation function is used after each convolutional layer, except for the last convolutional layer of each parallel route, where the activation function is after the BatchNorm layer.

Images