Method for detecting ultra-low concentration mixed organophosphorus pesticides based on cnt-fet and machine learning

Abstract

The application provides a kind of based on CNT-FET and machine learning's ultra-low concentration mixed organophosphorus pesticide detection method, the method comprises: preparation CNT-FET sensor is used to detect organophosphorus pesticide;Test CNT-FET sensor in different concentrations mixed organophosphorus pesticide in transfer curve, and record device baseline and liquid environment baseline;The device transfer characteristic is converted into image to construct image data set and is divided into training set and test set;Recognition model of CNN convolutional neural network of fusion CBAM double attention mechanism is constructed and is iteratively trained and verified, obtains the recognition model satisfying preset standard.Thereby, the recognition model obtained by training can identify and output the pesticide components and concentration in mixed organophosphorus pesticide.By the application, rapid and accurate identification and quantitative analysis of low concentration mixed organophosphorus pesticide components can be realized, which provides technical support for food safety supervision, environmental monitoring and pesticide residue control in agricultural production process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pesticide detection, in particular to a CNT-FET and machine learning-based method for detecting mixed organophosphorus pesticides at ultra-low concentration. BACKGROUND

[0002] In the modern agricultural system, scientific application of pesticides is of great significance to ensure crop yield and quality. Organophosphorus pesticides, with their high biological activity, can inhibit pest reproduction, block disease transmission, and enhance crop disease resistance and adaptability, laying a solid foundation for global food security and sustainable agricultural development. However, the unreasonable use and excessive residues of pesticides, such as excessive spraying or improper application methods, are causing a series of ecological and environmental problems, including but not limited to water pollution, soil fertility degradation, and disruption of ecological balance, which ultimately pose potential threats to human health.

[0003] In the field of pesticide residue detection technology, as the standards for food safety and ecological environmental protection become increasingly stringent, the demand for accurate and efficient detection of low-concentration pesticide residues is becoming increasingly prominent. In particular, for mixed pesticide samples, due to the structural similarity of different pesticide components, how to accurately distinguish and quantify each component in a complex matrix has become a major challenge for current detection technology.

[0004] Traditional analysis and detection technologies, such as gas chromatography-mass spectrometry (GC-MS) and high-performance liquid chromatography (HPLC), play an important role in the field of pesticide residue detection and can achieve high sensitivity and accuracy under certain conditions. However, when faced with low-concentration mixed pesticide samples, traditional GC-MS and HPLC detection technologies have many limitations. On the one hand, the sample pretreatment process is extremely tedious, usually requiring multiple steps such as extraction and purification, which not only consumes time and effort, but also prolongs the overall analysis period, increases the complexity of operation and detection cost. On the other hand, due to the structural similarity of some components in mixed organophosphorus pesticides, traditional techniques are difficult to effectively distinguish at low concentrations, which can easily lead to deviations in detection results. In addition, these traditional detection technologies usually rely on large laboratory equipment, making it difficult to achieve rapid detection on site, and unable to meet the actual needs of real-time monitoring and early warning, limiting their application in practical scenarios.

[0005] In recent years, carbon nanotube field effect transistor (CNT-FET) technology has been widely studied due to its unique electrical properties, high specific surface area, and excellent chemical stability, exhibiting outstanding performance in the sensing field, especially in the detection of ultra-low concentration and ultra-low content objects, such as medical detection and gas detection. In view of this, the present application aims to utilize the high sensitivity sensing characteristics of carbon nanotube field effect transistor to realize its application in ultra-low concentration pesticide detection. SUMMARY

[0006] In view of the deficiencies of the prior art in detecting ultra-low concentration mixed organic pesticides, the present application aims to provide a method for detecting ultra-low concentration mixed organophosphorus pesticides based on CNT-FET and machine learning, which combines the high sensitivity sensing characteristics of carbon nanotube field effect transistor with the data processing advantages of machine learning algorithm to construct an analysis method for low concentration mixed organophosphorus pesticides, so as to realize rapid and accurate identification and quantitative analysis of low concentration mixed organophosphorus pesticide components, and provide strong technical support for food safety supervision, environmental monitoring and pesticide residue control in agricultural production process.

[0007] According to the first aspect of the present application, in the proposed method for detecting ultra-low concentration mixed organophosphorus pesticides based on CNT-FET and machine learning, the electrical parameters tested by a single sensing unit are combined with machine learning algorithm to distinguish and quantify the concentration of mixed samples with high precision, and a single sensor is used to realize accurate classification and identification of multiple target organophosphorus pesticides, thereby effectively reducing the number of sensors required for detection, effectively controlling the detection cost, and reducing the overall volume of the equipment, improving the practicability and convenience.

[0008] According to the second aspect of the present application, in the proposed method for detecting ultra-low concentration mixed organophosphorus pesticides based on CNT-FET and machine learning, the CNN convolutional neural network is combined with channel attention mechanism and spatial attention mechanism to distinguish and quantify the concentration of mixed samples with high precision, and high-accuracy and efficient identification of low concentration mixed organophosphorus pesticide components and their concentrations is realized.

[0009] The method for detecting ultra-low concentration mixed organophosphorus pesticides based on CNT-FET and machine learning according to the embodiments of the present application comprises the following steps:

[0010] Step 1, preparing a carbon-based field effect transistor sensor for detecting organophosphorus pesticides, using a reticular structure carbon nanotube film to form a channel layer to form a carbon nanotube field effect transistor type CNT-FET sensor;

[0011] Step 2, testing the transfer curve of the CNT-FET sensor in mixed organophosphorus pesticides at different concentrations, and recording the basic transfer curve as the device baseline and the PBS transfer curve without adding mixed pesticides as the liquid environment baseline;

[0012] Step 3, converting the collected CNT-FET sensor transfer characteristic data into image format, constructing an image data set, and dividing the data set into training set and test set according to 8:2;

[0013] Step 4, constructing a recognition model based on a CNN convolutional neural network, wherein a CBAM double attention mechanism is fused, the CBAM double attention mechanism including a channel attention module and a spatial attention module, respectively used for multi-scale feature extraction and weight adjustment processing of features of different channels to suppress channel interference and enhance feature selectivity;

[0014] Step 5, training the recognition model based on the CNN convolutional neural network using the training set until a predetermined training round is reached;

[0015] Step 6, verifying the trained recognition model based on the CNN convolutional neural network using the test set, evaluating the recognition accuracy of the model, and obtaining a recognition model meeting the preset standard; and

[0016] Step 7, using the trained recognition model to recognize and output the pesticide components and concentrations in the mixed organophosphorus pesticides.

[0017] As an optional implementation, in the step 2, for the transfer curve test of the CNT-FET sensor, the source-drain voltage is fixed at -0.1 V, and the gate voltage scanning range is +1 V to -1 V in different test environments. ds )。

[0018] As an optional implementation, in the step 2, the transfer characteristic test sequence of the CNT-FET sensor is:

[0019] First, test the transfer characteristics of the device without adding a mixed pesticide solution sample, which serves as a baseline, and then sequentially increase the concentration gradient to test the added mixed pesticide solution sample.

[0020] As an optional implementation, the concentration gradient is set to 100 ag / mL, 1 fg / mL, 10 fg / mL, 100 fg / mL, and 1 pg / mL, respectively.

[0021] As an optional implementation, the test process of the step 2 specifically includes the following operations:

[0022] Step 21, test preparation: spin coating photoresist S1813 on the prepared CNT-FET sensor substrate to form a test window with a size of 80 microns * 10 microns;

[0023] Step 22, transfer characteristic test: using a Keithley 4200A instrument to record the transfer characteristics of the CNT-FET sensor, setting the fixed source-drain voltage to -0.1 V, and the gate voltage scanning range from +1 V to -1 V. Measure the source-drain current (Ids), wherein the top gate is used as a reference electrode, and the buffer is selected as 0.1*PBS;

[0024] wherein, the sensor transfer curve is tested in a non-liquid environment as the device baseline of the sensor;

[0025] Step 23, preparation of mixed organophosphorus pesticide solution: mixed organophosphorus pesticide solution is prepared, including methyl parathion, dimethoate, dimethoate-oxon, all in 0.1*PBS environment;

[0026] Step 24, baseline test: test the baseline transfer curve from the CNT-FET sensor, use 0.1*PBS solution as a blank group to calculate the liquid environment baseline of the sensor response as the background signal;

[0027] Step 25, mixed organophosphorus pesticide test: test different concentrations of mixed organophosphorus pesticide solution, the concentration is increased in turn, respectively 100ag / mL, 1fg / mL, 10fg / mL, 100fg / mL, 1pg / mL, 20μL sample is taken by a pipette to incubate each CNT-FET sensor for 5 minutes to maintain the same ion strength for all CNT-FET sensors, then set the test parameters according to the same test process in step 22 and collect the sensor transfer characteristics.

[0028] As an optional implementation, in step 23, the mixed organophosphorus pesticide solution includes a mixture of three organophosphorus pesticides, all prepared in 0.1*PBS environment, including:

[0029] dimethoate & dimethoate-oxon, dimethoate & methyl parathion, dimethoate-oxon & methyl parathion, same concentration 1:1 mixture and different concentration 1:1 mixture.

[0030] As an optional implementation, in step 4, the recognition model based on CNN convolutional neural network is constructed, including:

[0031] The recognition model based on CNN convolutional neural network includes an input layer, two convolutional layers, two fully connected layers, a dropout layer and an output layer, and fuses a CBAM dual attention mechanism composed of a channel attention module and a spatial attention module, wherein:

[0032] The channel attention module learns the weight for each channel feature map according to the global average pooling technology and the full connection operation logic, and reversely feeds the learned weight to the original feature map to selectively focus on the features of different channels in the convolutional neural network;

[0033] The spatial attention module focuses on the information of different spatial positions in the feature map according to the dual operation of maximum pooling and average pooling in the spatial dimension, performs a weight learning process on the feature map, and accurately feeds back the weight so that the features of the key spatial positions are enhanced.

[0034] The channel attention module and the spatial attention module are fused in the CNN model to optimize the feature representation.

[0035] As an optional implementation, in the step 4, the input layer and two convolution layers in the CNN convolutional neural network-based recognition model are designed as follows:

[0036] The input layer receives image data converted from the transferred characteristic data, and the input layer serves as a data inlet of the model to provide data input for subsequent convolution operations.

[0037] The first convolution layer extracts features from the image data of the input layer, and the size of the first convolution layer is set to 50x50x4. The basic features of the image are extracted through convolution operation on the image data, and the local information of the image corresponding to different concentrations of mixed organophosphorus pesticides is included. Then, the output of the convolution is activated by using a ReLU function.

[0038] The first max-pooling layer performs downsampling, and a 3x3 max-pooling operation is performed on the output data of the first convolution layer after ReLU activation to reduce the dimension of the data from 50x50x4 to 25x25x8, thereby reducing the complexity of subsequent calculations.

[0039] The second convolution layer further extracts deep features from the data output by the first max-pooling layer, further extracts deep features from the data, captures higher-level image feature patterns, and provides representative feature expressions for subsequent classification tasks. Then, the output of the convolution is activated by using a ReLU function.

[0040] The second max-pooling layer performs secondary downsampling, and a 3x3 max-pooling operation is performed on the output data of the second convolution layer after ReLU activation to convert the data from 25x25x8 to 12x12x32, thereby reducing the data volume and enhancing the robustness and expression ability of the features, and providing data for subsequent attention mechanism processing.

[0041] As an optional implementation, the CBAM double attention mechanism is composed of a channel attention module and a spatial attention module, and the CBAM double attention mechanism includes the following steps:

[0042] The channel attention module extracts multi-scale features, and performs spatial attention processing on the data output by the second max-pooling layer. First, two 3x3 pooling operations and one 1x1 pooling operation are performed, multi-scale pooling operations are performed from multiple angles to extract spatial feature information of the data, and feature changes at different positions and scales in the image are captured.

[0043] The spatial attention module performs feature weighting processing, introduces twice channel attention mechanism, and performs full connection operation on the result of spatial attention mechanism processing, weights and adjusts the features of different spatial attention channels through the full connection layer, gives corresponding weight according to the importance of each channel feature, enhances the channel feature expression, suppresses the interference information and improves the feature selectivity;

[0044] Then, the results processed by the spatial attention module and the channel attention module are fused, activated by a Sigmoid function, feature weight normalization is realized, and the output value of the fused features is mapped to the interval [0, 1], so as to give each feature a normalized weight.

[0045] As an optional implementation, in the step 4, the two full connection layers, the dropout layer and the output layer in the recognition model based on the CNN convolutional neural network are designed as follows:

[0046] The first full connection layer performs feature integration, inputs the data set processed by the double attention mechanism into the first full connection layer, integrates and nonlinearly transforms the output features, fuses the information of different dimensions and feature spaces, and extracts more representative feature representations;

[0047] The second full connection layer performs feature refining, receives the output data after the first full connection layer, further refines and transforms the features, further excavates the complex relationship between the features through a nonlinear activation function, and provides feature input for the final classification task;

[0048] The output layer performs classification result determination, the output layer determines the classification of different concentration organic phosphorus mixed pesticides based on the output result of the second full connection layer, outputs the final classification result of the organic phosphorus mixed pesticides, and completes the entire classification and recognition task.

[0049] The above embodiment of the application provides a solution to the detection problem of low-concentration mixed organic phosphorus pesticide components and their concentrations. The detection method combines the carbon nanotube field effect transistor (CNT-FET) sensor and the machine learning algorithm, uses the high sensitivity and low detection limit characteristics of the CNT-FET sensor to accurately capture sample information, realizes ultra-low concentration detection, and the detection lower limit is as low as 84.58 ag / mL; at the same time, the CNN convolutional neural network is used to fuse the channel attention mechanism and the spatial attention mechanism, the mixed pesticide data measured by the sensor is subjected to feature extraction and deep learning training, and after multiple iterations, the recognition accuracy for three kinds of mixed pesticides is nearly 100%, and the concentration quantitative recognition is as high as 99.32%.

[0050] The method proposed in this invention can detect, identify, and analyze mixed samples with ultra-low concentrations, providing efficient, accurate, and reliable technical support for the detection of mixed organophosphorus pesticide samples in fields such as environmental monitoring, food safety, and agricultural production.

[0051] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below may be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Furthermore, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0052] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0053] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings.

[0054] Figure 1 This is a flowchart of a method for detecting ultra-low concentration mixed organophosphorus pesticides based on CNT-FET and machine learning according to an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the structure of a carbon nanotube field-effect transistor sensor according to an embodiment of the present invention.

[0056] Figure 3a , 3b These are the detection curve and calibration curve of the CNT-FET sensor according to an embodiment of the present invention.

[0057] Figure 4 This is a schematic diagram of the CNN model architecture fusion and attention mechanism according to an embodiment of the present invention.

[0058] Figure 5 This is a schematic diagram illustrating the iterative training process and accuracy recording results according to an embodiment of the present invention.

[0059] Figure 6 This is a heatmap of the classification results of the machine learning model according to an embodiment of the present invention. Detailed Implementation

[0060] For a more complete understanding of the technical content of the present application, specific embodiments are described below with reference to the accompanying drawings.

[0061] Aspects of the present application are described in the disclosure by reference to the accompanying drawings, which show many illustrative embodiments. The embodiments of the present disclosure are not necessarily intended to include all aspects of the present application. It should be understood that the various concepts and embodiments introduced above, and those described in more detail below, can be implemented in any of numerous ways, as the disclosed concepts and embodiments are not limited to any one implementation. Additionally, some aspects of the present application can be used independently of any other aspects of the present application, or can be used in any appropriate combination.

[0062] {Example 1}

[0063] In conjunction Figure 1 , Figure 2 As shown in the drawings, the CNT-FET and machine learning based ultra-low concentration mixed organophosphorus pesticide detection method according to the embodiments of the present application includes the following steps:

[0064] Step 1, preparing a carbon-based field effect transistor sensor for detecting organophosphorus pesticides, using a reticular structure carbon nanotube film to form a channel layer, forming a carbon nanotube field effect transistor type CNT-FET sensor;

[0065] Step 2, testing the transfer curve of the CNT-FET sensor in different concentrations of mixed organophosphorus pesticides, and recording the basic transfer curve as the device baseline, and the PBS transfer curve without adding mixed pesticides as the liquid environment baseline;

[0066] Step 3, converting the collected CNT-FET sensor transfer characteristic data into image format, constructing image data set, and dividing the data set into training set and test set according to 8:2;

[0067] Step 4, constructing a recognition model based on CNN convolutional neural network, wherein a CBAM dual attention mechanism is fused, the CBAM dual attention mechanism includes a channel attention module and a spatial attention module, respectively used for multi-scale feature extraction and weight adjustment processing of features of different channels, to suppress channel interference and enhance feature selectivity;

[0068] Step 5, training the recognition model based on CNN convolutional neural network using the training set until the predetermined training round is reached;

[0069] Step 6, verifying the trained recognition model based on CNN convolutional neural network using the test set, evaluating the recognition accuracy of the model, and obtaining a recognition model that meets the preset standard; and

[0070] Step 7, using the recognition model obtained by training to recognize and output the pesticide components and concentrations in the mixed organophosphorus pesticide.

[0071] As an optional implementation, in step 2, for the transfer curve test of the CNT-FET sensor, the fixed source-drain voltage is -0.1V, the gate voltage scanning range is +1V to -1V, and the source-drain current (Ids) is measured. ds ) in different test environments.

[0072] As an optional implementation, in step 2, the transfer characteristic test sequence of the CNT-FET sensor is as follows:

[0073] First, test the transfer characteristics of the device without adding mixed pesticide solution samples, which serves as the baseline, and then test the added mixed pesticide solution samples in order of increasing concentration gradient.

[0074] As an optional implementation, the concentration gradient is set to 100 ag / mL, 1 fg / mL, 10 fg / mL, 100 fg / mL, and 1 pg / mL, respectively.

[0075] As an optional implementation, the test process of step 2 specifically includes the following operations:

[0076] Step 21, test preparation: spin-coat photoresist S1813 on the prepared CNT-FET sensor substrate to form a test window with a size of 80 microns * 10 microns;

[0077] Step 22, transfer characteristic test: use Keithley 4200A instrument to record the transfer characteristics of the CNT-FET sensor, set the fixed source-drain voltage to -0.1V, and the gate voltage scanning range from +1V to -1V. Measure the source-drain current (Ids), where the top gate is used as the reference electrode, and the buffer solution is 0.1*PBS;

[0078] Wherein, the sensor transfer curve is tested in the liquid-free environment as the device baseline of the sensor;

[0079] Step 23, preparation of organophosphorus pesticide solution: mix organophosphorus pesticide solutions including methyl parathion, dimethoate, and dimethoate oxide, all prepared in a 0.1*PBS environment;

[0080] Step 24, baseline test: test the basic transfer curve of the CNT-FET sensor, use 0.1*PBS solution as the blank group to calculate the liquid environment baseline of the sensor response as the background signal;

[0081] Step 25, mixed organophosphorus pesticide test: test different concentrations of mixed organophosphorus pesticide solution, the concentration is increased in turn, respectively 100 ag / mL, 1 fg / mL, 10 fg / mL, 100 fg / mL, 1 pg / mL, 20 μL sample is taken by pipette, incubate each CNT-FET sensor for 5 minutes, to keep all CNT-FET sensors measure the same ion strength, then set the test parameters according to the same test process of test step 22 and collect sensor transfer characteristics.

[0082] As an optional implementation, in step 23, the mixed organophosphorus pesticide solution includes a mixture of three organophosphorus pesticides, each prepared in 0.1*PBS environment, including:

[0083] dimethoate & dimethoate, dimethoate & methyl parathion, dimethoate & methyl parathion, same concentration 1:1 mixture and different concentration 1:1 mixture.

[0084] As an optional implementation, in step 4, the recognition model based on CNN convolutional neural network is constructed, including:

[0085] The recognition model based on CNN convolutional neural network includes an input layer, two convolutional layers, two fully connected layers, a dropout layer and an output layer, and fuses a CBAM double attention mechanism composed of a channel attention module and a spatial attention module, wherein:

[0086] The channel attention module learns the weight for each channel feature map according to the global average pooling technology and the full connection operation logic, and feeds back the learned weight to the original feature map in reverse, so as to selectively pay attention to the features of different channels in the convolutional neural network;

[0087] The spatial attention module focuses on the information of different spatial positions in the feature map according to the dual operation of maximum pooling and average pooling in spatial dimension, performs weight learning process on the feature map, and accurately feeds back the weight, so that the features of key spatial positions are enhanced;

[0088] The channel attention module and the spatial attention module are fused in the CNN model to optimize the feature representation.

[0089] As an optional implementation, in step 4, in the recognition model based on CNN convolutional neural network, the input layer and two convolutional layers are designed as follows:

[0090] The data input of the input layer receives the image data converted from the transfer characteristic data, and the input layer serves as the data inlet of the model, providing data input for subsequent convolution operation;

[0091] The first convolutional layer extracts features, performs convolution operation on the image data of the input layer, and the size of the first convolutional layer is set to 50*50*4; the basic features of the image are extracted by performing convolution operation on the image data, which contains the local information of the image corresponding to the mixed organic phosphorus pesticide data of different concentrations; then the ReLU function is used to activate the convolution output;

[0092] The first max-pooling layer performs downsampling, and a 3*3 max-pooling operation is performed on the output data of the first convolutional layer after ReLU activation to perform downsampling processing, reducing the dimension of the data from 50*50*4 to 25*25*8, thereby reducing the complexity of subsequent calculations;

[0093] The second convolutional layer extracts deep features, and the data output by the first max-pooling layer is subjected to convolution operation again, so as to further extract deep features in the data through the second convolutional layer, capture higher-level image feature patterns, and provide representative feature expressions for subsequent classification tasks; then the ReLU function is used to activate the convolution output;

[0094] The second max-pooling layer performs secondary downsampling, and a 3*3 max-pooling operation is performed on the output data of the second convolutional layer after ReLU activation to perform downsampling, converting the data from 25*25*8 to 12*12*32, thereby reducing the data amount and enhancing the robustness and expression ability of the features, and providing data for subsequent attention mechanism processing.

[0095] As an optional implementation, the fusion is a CBAM double attention mechanism composed of a channel attention module and a spatial attention module, wherein:

[0096] The channel attention module extracts multi-scale features, and performs spatial attention processing on the data output by the second max-pooling layer. First, two 3*3 pooling operations and one 1*1 pooling operation are performed, multi-scale pooling operations are performed to extract spatial feature information of the data from multiple angles, and feature changes at different positions and scales in the image are captured;

[0097] The spatial attention module performs feature weighting processing, introduces two channel attention mechanisms, and performs full connection operation on the results of the spatial attention mechanism processing, weights the features of different spatial attention channels through the full connection layer, assigns appropriate weights to each channel feature according to the importance of the channel feature, enhances the channel feature expression, suppresses interference information, and improves the feature selectivity;

[0098] Then, the results processed by the spatial attention module and the channel attention module are fused, activated by a Sigmoid function, feature weight normalization is realized, the output value of the fused features is mapped to the interval [0, 1], and each feature is assigned a normalized weight.

[0099] As an optional embodiment, in the step 4, in the recognition model based on the CNN convolutional neural network, the two full connection layers, the dropout layer and the output layer are designed as follows:

[0100] The first full connection layer performs feature integration, inputs the data set processed by the double attention mechanism into the first full connection layer, integrates and nonlinearly transforms the output features, fuses the information in different dimensions and feature spaces, and extracts more representative feature representations;

[0101] The second full connection layer performs feature refining, receives the output data after the first full connection layer, further refines and transforms the features, further excavates the complex relationships between the features through the nonlinear activation function, and provides feature input for the final classification task;

[0102] The output layer performs classification result determination, determines the classification of the different concentration organic phosphorus mixed pesticides based on the output result of the second full connection layer, outputs the final classification result of the organic phosphorus mixed pesticides, and completes the entire classification and recognition task.

[0103] Therefore, the method for detecting and analyzing ultra-low concentration mixed organic phosphorus pesticides provided by the present application combines the advantages of high sensitivity and ultra-low detection limit of the carbon nanotube CNT-FET sensor and the nonlinear feature mining and data processing capability based on the CNN convolutional neural network, the CBAM double attention mechanism further improves the deep feature selectivity, the feature space is deeply optimized, the important channel feature expression is enhanced, the interference information of the unimportant channel is inhibited, the selectivity of the model to the features is improved, the overall performance of the model is significantly improved, the recognition accuracy of the three mixed pesticides (methyl parathion, dimethoate and oxime dimethoate) is nearly 100%, the concentration quantitative recognition is as high as 99.32%, the efficient and accurate detection of ultra-low concentration mixed organic phosphorus pesticides is realized, and the method is suitable for on-site rapid detection scene, and provides an efficient detection scheme for low concentration mixed organic phosphorus pesticides in the fields of environmental monitoring, food safety and agricultural production.

[0104] {Example 2}

[0105] In this example, the method for detecting and analyzing ultra-low concentration mixed organic phosphorus pesticides based on CNT-FET and machine learning of the foregoing embodiments is further described.

[0106] CNT-FET sensor

[0107] As an optional embodiment, the CNT-FET sensor can be prepared based on the existing semiconductor process technology.

[0108] As an example, the CNT-FET sensor has an insulating substrate 10 / 20, a semiconductor carbon nanotube layer 30, a source and drain electrode 40, a gate oxide layer 50, a top gate electrode layer 60 and a test passivation layer 70, a conductive channel 80 stacked layer by layer from bottom to top, and the structural schematic diagram is shown in the accompanying drawings Figure 2 .

[0109] The top gate electrode layer 60, the source and drain electrode 30 are each provided with an independent lead wire for signal lead-out. The lead wire is encapsulated and insulated.

[0110] As described above, the test process uses PBS buffer which does not contact the source electrode, the drain electrode and the CNT channel layer.

[0111] It should be understood that the preparation process of the CNT-FET sensor according to the existing mature process is as follows:

[0112] Step 1-1, depositing a reticular carbon nanotube film on the Si-SiO2 substrate by solution treatment method;

[0113] Step 1-2, uniformly gluing the surface of the carbon sheet, developing the required pattern by using a photolithography process and evaporating metal to form source and drain electrodes, and then removing the glue and peeling off;

[0114] Step 1-3, again uniformly gluing and photolithography, using a reactive ion etching process to etch the CNT outside the channel region;

[0115] Step 1-4, depositing yttrium in the channel region by electron beam evaporation, and heat-oxidizing to form a yttrium oxide layer;

[0116] Step 1-5, again using a photolithography process to expose the gate pattern, and then evaporating metal to form a gate electrode for direct contact with organophosphorus pesticides as a sensitive layer.

[0117] Step 1-6, spin-coating photoresist on the prepared sensor substrate to passivate and form a test window with a size of 80 microns*10 microns.

[0118] In the present application, the design of the CNT-FET device as an example is prepared as follows:

[0119] - substrate cleaning: the Si / SiO2 substrate is sequentially placed in acetone, anhydrous ethanol and ultrapure water, each is cleaned by ultrasonic cleaning and then taken out and dried with nitrogen to obtain the Si / SiO2 substrate;

[0120] - carbon tube deposition: the cleaned Si / SiO2 substrate is placed in a 30g / mL pure semiconductor carbon nanotube solution, the solution completely immerses the substrate and is placed still for at least 12 hours, then washed with toluene solution and finally dried with nitrogen to obtain the Si / SiO2 / CNT substrate;

[0121] - Photoresist spin coating and baking: First, the Si / SiO2 / CNT substrate is placed on the spin coater to spin coat the photoresist LOR evenly, and then baked on the hot plate at 165°C for 5 minutes. Then, the photoresist S1813 is spin coated again and baked at 115°C for 1 minute, which facilitates the subsequent removal and electrode pattern presentation using a double-layer photoresist process;

[0122] - Photoetching and development: The substrate is exposed to ultraviolet light in the photoetching machine and developed by the developing solution to obtain an accurate electrode pattern, and the developed wafer is obtained;

[0123] - Source-drain electrode formation: The wafer is placed in an electron beam coater, and Ti / Pd / Au layers are deposited in sequence using evaporation technology, with thicknesses of 0.6 nm, 20 nm, and 40 nm, respectively, to form a P-type contact electrode. The width of the CNT-FET channel is set to 80 microns, and the length is 20 microns;

[0124] - Photoresist removal: After the wafer completes metal evaporation, the substrate is immersed in Remover PG solution for at least 8 hours to completely remove the photoresist and form the source-drain electrode;

[0125] - Channel etching: As described in the previous steps, the wafer is first coated with photoresist and then subjected to photoetching and development. The corresponding etching layer is selected, and the size is 80 microns x 50 microns to protect the area between the source-drain electrode using photoresist. Excess carbon nanotubes are removed by reactive ion etching technology to form a conductive channel;

[0126] - Gate dielectric layer formation: The wafer is deposited twice with 3 nm of yttrium (Y) by electron beam evaporation technology, and then heated in air at 270°C for 30 minutes to form a yttrium oxide (Y2O3) gate dielectric layer;

[0127] - Gate electrode formation: The wafer is first coated with photoresist and then subjected to photoetching and development. The corresponding gate electrode layer is selected and 20 nm of palladium (Pd) and 20 nm of gold (Au) are evaporated by electron beam evaporation technology to complete the preparation of the CNT-FET sensor.

[0128] It should be understood that after completing the device preparation, a coating process is performed before testing the device characteristic curve, which includes: spin coating photoresist S1813 on the prepared sensor substrate to form a test window with a size of 80 microns * 10 microns.

[0129] Sensor characteristic curve test

[0130] - FET transfer characteristics were tested by Keithley 4200A instrument, with source-drain voltage of -0.1V, gate voltage scanning range of +1V to -1V, top gate as reference electrode, 0.1x PBS buffer, and source-drain current (Ids) was measured;

[0131] - Mixed organophosphorus pesticide solutions of methyl parathion, dimethoate, dimethoate oxide, etc. were prepared in 0.1x PBS respectively;

[0132] - The transfer curve of the sensor without liquid drop was tested as the device baseline;

[0133] - The 0.1x PBS solution was used as a blank group for testing, and the obtained device curve was the device baseline in liquid environment, i.e. the liquid environment baseline, which was used to calculate the response of the sensor to concentration change;

[0134] - Different concentrations of mixed organophosphorus pesticide solutions were tested, 20 μL of sample was taken by a pipette gun to incubate each sensing unit for 5 minutes, the ion strength measured by FET was kept the same, and then the FET transfer characteristics data were collected.

[0135] Specifically, different concentrations of mixed organophosphorus pesticide solutions were tested, the concentrations were sequentially increased, and were 100 ag / mL, 1 fg / mL, 10 fg / mL, 100 fg / mL, and 1 pg / mL respectively, 20 μL of sample was taken by a pipette gun to incubate each sensing unit for 5 minutes to keep the same ion strength measured by all FETs, and then the FET transfer characteristics were collected by the semiconductor analyzer according to the same test process and condition setting parameters (source-drain voltage of -0.1V, gate voltage scanning range of +1V to -1V, top gate as reference electrode, and source-drain current Ids was measured) as mentioned above.

[0136] In combination Figure 3a As shown in the figure, the transfer characteristics of the CNT-FET device showed obvious changes as the concentration of dimethoate increased. Specifically, the threshold voltage of the transfer curve showed a leftward movement, and at the same time, as the target concentration increased from 100 ag / mL to the pg / mL level, the source-drain current continued to rise, clearly indicating that the conductivity of the semiconductor channel was continuously increasing.

[0137] In order to observe that the current change was caused by dimethoate, rather than other interference factors, we carried out multiple 0.1x PBS blank tests before formally adding dimethoate for testing. The test results showed that the change of the device current was not linear with the use of the buffer and the accumulation of the test times, that is, the stable background signal provided by the 0.1x PBS buffer had little interference on the detection signal, which could be almost ignored, which created good conditions for accurate detection of dimethoate.

[0138] To further verify the detection capability of the present application for dimethoate, the present application specifically constructs a calibration graph, which uses the absolute value of the relative current change (i.e. I DS After careful analysis of the calibration curve, it is found that the FET sensor plays a key role, and its linear detection range can cover the interval from 100 ag / mL to 1 pg / mL, and the calibration sensitivity of the sensor reaches 0.135 (see Figure 3b , under the condition that the number of devices is n = 10, which fully demonstrates the performance of the method of the present application in the field of dimethoate detection.

[0139] The FET sensor response is calculated by the equation at Vg = -1 V, where I0is the current I DS at -1 V in the blank sample (baseline), and ΔI / I0is the difference between I DS of the sample and the blank.

[0140] The calibration curve is constructed by plotting the average relative conductance of all devices versus the concentration of the analyte on a logarithmic scale. The vertical bars reported in the calibration curve represent the standard error of the relative response of all test FET devices.

[0141] To determine the linear range of the sensor, a logistic model is used to fit the calibration curve, and the linear region of the fitted curve is determined as the linear range. The calibration sensitivity is defined as the slope of the linear region in the calibration curve. The detection limit is calculated according to the definition of the International Union of Pure and Applied Chemistry (IUPAC):

[0142] First, the standard deviation of the blank sample test is determined using the equation which can also be called the background signal σ, then the equation where s is the slope of the calibration curve, k is 3, and the confidence level is 99.6%. From the calibration curve shown in Figure 3b , it can be found that s = 0.09885 and X L = 0.8458. Then, the concentration of organophosphorus corresponding to X LOD (LOD) is obtained from the fitted calibration curve.

[0143] Model building, training and testing

[0144] 1. CNN model architecture

[0145] In this example, the Figure 4As shown, the model includes an input layer, two convolutional layers for feature extraction, two fully connected layers responsible for information integration and decision-making, a dropout layer to prevent overfitting, and an output layer.

[0146] The convolutional operation of the CNN model can deeply analyze and mine the image spatial architecture, accurately lock the key position information, and autonomously extract high-level and strongly representative features from the original pixel data, greatly improving the efficiency and accuracy.

[0147] To improve the performance of the model, the CBAM (Convolutional Block Attention Module) dual attention mechanism is fused in the neural network model based on CNN designed in the application.

[0148] In an embodiment of the application, the CBAM attention mechanism is composed of a channel attention module and a spatial attention module.

[0149] 1) Channel attention module: first, use global average pooling technology to compress the feature map of each channel into a value with global representation, to obtain the overall feature information of the channel, and to compress the feature of the data picture. Then, based on the global information obtained above, a fine weight learning process is carried out relying on the fully connected operation logic. In the fully connected layer, the adaptive weight value for each channel is calculated by adjusting the connection weight of the neuron, and these weight values reflect the importance of the corresponding channel feature to the model decision. Then, the learned weight information is fed back to the original feature map in a reverse direction, and the features of the key channels are deepened and strengthened by element-wise multiplication, so that the channel information that plays a key supporting role in the model judgment is highlighted, thereby guiding the model to focus more on the extraction and processing of important feature information.

[0150] 2) Spatial attention module: first, perform maximum pooling and average pooling operations on the feature map in the spatial dimension. The maximum pooling can capture the local maximum value in the feature map and highlight the significant area in the feature map; the average pooling can obtain the average feature of the feature map, which reflects the feature distribution from the overall picture pixel position. The results of the two kinds of pooling operations are processed through a series of subsequent processes such as splicing and convolution, to complete the weight learning process of different spatial positions of the feature map, and to accurately calculate the weight information of each spatial position. Similarly, these weight information is fed back to the original feature map, and the features in the key spatial position are significantly enhanced by multiplying the corresponding position elements of the original feature map, just like giving the important spatial area a "focus mark", prompting the model to pay more attention to these key areas in the subsequent processing process, thereby improving the model's ability to accurately recognize complex features.

[0151] Through the synergistic operation of the two attention mechanisms, the neural network's ability to interpret features is significantly improved, the feature space is deeply optimized, and the overall performance of the model is also significantly improved, laying a solid foundation for subsequent application scenarios.

[0152] At the same time, the fusion of CNN network and CBAM attention mechanism not only improves the recognition accuracy, but also shortens the training time, making the model more efficient and accurate in processing low-concentration mixed pesticide data.

[0153] 2. Data set construction

[0154] Based on the obtained sensor characteristic curves, including three mixed pesticide test curves, device baseline, liquid environment baseline, and each single concentration mixed pesticide curve.

[0155] For example, three mixed pesticides are configured as: dimethoate & dimethoate, dimethoate & methyl parathion, dimethoate & methyl parathion with the same concentration 1:1 mixed and different concentration 1:1 mixed.

[0156] The transfer characteristic curve of mixed pesticides is digitally converted to become an image format suitable for model training. The image data will be used as the input of the subsequent convolutional neural network (CNN) model to adapt to the characteristics of the CNN model in processing image data, laying a foundation for the effective operation of the model.

[0157] Thus, a device transfer characteristic image data set covering three mixed pesticides is constructed.

[0158] Then, using the random sampling method, the entire data set is accurately divided into training set and test set according to the ratio of 8:2, which lays a solid data foundation for rigorous training and accurate verification of the subsequent model.

[0159] 3. Model training and verification

[0160] The model undergoes rigorous iterative training for multiple rounds, and the results of each round of training are recorded in detail in Figure 5 After 40 rounds of training, the model achieves 100% accurate recognition accuracy on the test set. To further rule out the possibility of accidental interference, the invention further expands the training rounds to 60 rounds, and the final results show that within the expanded training interval of 40 to 60 rounds, the average accuracy of the test set is stably maintained at a high level of 99.32%.

[0161] At the same time, in order to fully test the generalization ability of the model, we also select samples that do not participate in the previous training to rigorously test the model, and the obtained confusion matrix result shows that the average accuracy is still as high as 95.83%, fully demonstrating the excellent performance and high reliability of the model.

[0162] It can be seen that the detection method for mixed organophosphorus pesticides with ultra-low concentration designed by the application has excellent performance in sensitivity and detection limit. The carbon nanotube field effect transistor sensor can sensitively perceive pesticide components at nanomolar concentration, and the detection lower limit is as low as 84.58 ag / mL, which enables it to accurately capture low-concentration pesticide residues in the environment and food fields, and builds a strong defense line for safety protection. After 60 iterations of training, the average accuracy of the test set of the neural network model constructed by combining the machine learning algorithm increases to 99.32%. Even if it faces samples that do not participate in the early training, the average accuracy of the confusion matrix is 95.83%, which further demonstrates the high accuracy and high precision of the model. Whether it is the identification of mixed pesticides or the quantitative analysis of concentration, it can achieve high-accuracy identification.

[0163] At the same time, we see that by reducing the number of detection sensors in the design of the application, the preparation cost can be effectively controlled, the device volume is significantly reduced, the convenience and practicality are greatly increased, especially for on-site rapid detection, and it creates efficient, accurate and reliable technical support for the fields of environmental monitoring, food safety, etc.

[0164] {Example 3}

[0165] In this example, we more specifically elaborate on the construction of the identification model.

[0166] Combined Figure 4 As shown in the figure, the identification model based on the CNN convolutional neural network includes an input layer, two convolutional layers, two fully connected layers, a dropout layer and an output layer, and fuses the CBAM double attention mechanism composed of a channel attention module and a spatial attention module.

[0167] The data input of the input layer receives the image data converted from the transferred characteristic data. The input layer serves as the data inlet of the model and provides data input to be processed for subsequent convolution operations.

[0168] The first convolutional layer extracts features, and performs convolution operation on the image data of the input layer. The size of the first convolutional layer is set to 50x50x4. By performing convolution operation on the image data, the basic features of the image are extracted, which contains the local information of the image corresponding to different concentrations of mixed organophosphorus pesticides. Then the convolution output is activated using the ReLU function. The ReLU function is expressed as f(x) = max(0, x), which can introduce a nonlinear factor, so that the network can learn more complex patterns, and at the same time alleviate the problem of gradient disappearance and improve the training efficiency.

[0169] The first max-pooling layer performs downsampling, and a 3*3 max-pooling operation is performed on the output data of the first convolution layer after ReLU activation to perform downsampling processing, so that the data is reduced from 50*50*4 to 25*25*8, thereby effectively retaining key feature information in the image and reducing the complexity of subsequent calculation.

[0170] The second convolution layer further excavates deep features in the data, and captures higher-level image feature patterns, thereby providing representative feature expression for subsequent classification tasks.

[0171] The second max-pooling layer performs secondary downsampling, and a 3*3 max-pooling operation is performed on the output data of the second convolution layer after ReLU activation to perform downsampling, so that the data is converted from 25*25*8 to 12*12*32, thereby further reducing the data amount and enhancing the robustness and expression ability of the features, and providing data for subsequent attention mechanism processing.

[0172] The CBAM dual attention mechanism is composed of a channel attention module and a spatial attention module.

[0173] The channel attention module performs weight learning for each channel feature map according to the global average pooling technology and full connection operation logic, reversely feeds the learned weight to the original feature map, and selectively pays attention to the features of different channels in the convolutional neural network. The spatial attention module focuses on the information of different spatial positions in the feature map according to the dual operation of spatial dimension max-pooling and average-pooling, performs weight learning process on the feature map, and accurately feeds back the weight, so that the features of the key spatial positions are enhanced. Therefore, the fusion of the channel attention module and the spatial attention module in the CNN model optimizes the feature representation.

[0174] As an optional embodiment, the channel attention module performs multi-scale feature extraction on the output data of the second max-pooling layer, first performs two 3*3 pooling operations and one 1*1 pooling operation, extracts spatial feature information of the data from multiple angles through different scale pooling operations, and captures feature changes in different positions and scales of the image.

[0175] As an optional embodiment, the spatial attention module performs feature weighting processing, introduces twice channel attention mechanism, and performs full connection operation on the result of the spatial attention mechanism processing, weights and adjusts the features of different spatial attention channels through the full connection layer, gives corresponding weights according to the importance of each channel feature, enhances the feature expression of important channels, suppresses the interference information of unimportant channels, and improves the selectivity of the model to the features.

[0176] Then, the results processed by the spatial attention module and the channel attention module are fused, and a Sigmoid function is used for activation operation to realize feature weight normalization.

[0177] The mathematical expression of the Sigmoid function is:

[0178] Where x is the input value, and e is a natural constant, approximately equal to 2.71828.

[0179] The fused feature output value is mapped to the [0, 1] interval through the Sigmoid function, and each feature is given a normalized weight, highlighting the important features in model decision-making, thereby improving the attention and expression ability of the model to the features.

[0180] The first full connection layer integrates the features, inputs the data set processed by the double attention mechanism into the first full connection layer, integrates and nonlinearly transforms the output features, fuses the information of different dimensions and feature spaces, and extracts more representative feature representations;

[0181] The second full connection layer refines the features, receives the output data after the first full connection layer, and further refines and transforms the features through a nonlinear activation function to further mine the complex relationships between the features and provide feature input for the final classification task;

[0182] The output layer determines the classification result, the output layer determines the classification result of different concentration organic phosphorus mixed pesticides based on the output result of the second full connection layer, outputs the final organic phosphorus mixed pesticide classification result, and completes the entire classification and recognition task.

[0183] Although the present application has been disclosed as above with reference to the preferred embodiments, it is not intended to limit the present application. Those skilled in the art can make various modifications and improvements without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application shall be subject to the scope defined by the claims.

Claims

1. A method for detecting ultra-low concentration mixed organophosphorus pesticides based on CNT-FET and machine learning, characterized in that, The application relates to a method for identifying mixed organophosphorus pesticides by using a carbon-based field effect transistor sensor. The method comprises the following steps: Step 1: preparing a carbon-based field effect transistor sensor for detecting organophosphorus pesticides, using a reticular carbon nanotube film to form a channel layer, and forming a carbon nanotube field effect transistor type CNT-FET sensor; Step 2: testing the transfer curve of the CNT-FET sensor in mixed organophosphorus pesticides with different concentrations, and recording the basic transfer curve as the device baseline and the PBS transfer curve without adding mixed pesticides as the liquid environment baseline; Step 3: converting the collected CNT-FET sensor transfer characteristic data into an image format, constructing an image data set, and dividing the data set into a training set and a test set according to an 8:2 ratio; Step 4: constructing a recognition model based on a CNN convolutional neural network, wherein a CBAM double attention mechanism is fused, the CBAM double attention mechanism comprises a channel attention module and a spatial attention module, and is respectively used for multi-scale feature extraction and weight adjustment processing of features of different channels, so that channel interference is inhibited and feature selectivity is enhanced; Step 5: training the recognition model based on the CNN convolutional neural network by using the training set until a predetermined training round is reached; Step 6: verifying the trained recognition model based on the CNN convolutional neural network by using the test set, evaluating the recognition accuracy of the model, and obtaining a recognition model meeting a preset standard; And 2. The CNT-FET and machine learning based ultra-low concentration mixed organophosphorus pesticide detection method according to claim 1, characterized in that, The step 2, for the transfer curve test of the CNT-FET sensor, the fixed source-drain voltage is -0.1 V, the gate voltage scanning range is +1 V to -1 V, and the source-drain current (I ds ) is measured in different test environments.

3. The CNT-FET and machine learning based ultra-low concentration mixed organophosphorus pesticide detection method according to claim 1, characterized in that, Step 7: using the trained recognition model to recognize and output pesticide components and concentrations in mixed organophosphorus pesticides. In step 2, the transfer characteristic test sequence of the CNT-FET sensor is as follows:

4. The CNT-FET and machine learning based ultra-low concentration mixed organophosphorus pesticide detection method according to claim 3, characterized in that, First, the device transfer characteristic of a sample without adding mixed pesticide solution is tested, which is used as the baseline, and then the sample with mixed pesticide solution is tested in a concentration gradient. 5.The ultra-low concentration mixed organophosphorus pesticide detection method based on CNT-FET and machine learning according to any one of claims 1-4, characterized in that, The concentration gradient is set to 100 ag / mL, 1 fg / mL, 10 fg / mL, 100 fg / mL and 1 pg / mL. The test process of step 2 specifically comprises the following operations: Step 22, Transfer characteristics test: CNT-FET sensor transfer characteristics were recorded using Keithley 4200A instrument, with a fixed source-drain voltage of -0.1 V, a gate voltage scan range from +1 V to -1 V, and the source-drain current (I ds ) was measured. where the top gate was used as the reference electrode, and the buffer was selected as 0.1*PBS; Step 21: test preparation: spin coating photoresist S1813 on the prepared CNT-FET sensor substrate to form a test window with a size of 80 microns*10 microns; In the process of testing the sensor transfer curve without adding a liquid environment, the sensor is used as the device baseline; Step 23: preparing an organophosphorus pesticide solution: mixing organophosphorus pesticide solutions, including methyl parathion, dimethoate and dimethoate, all prepared in a 0.1*PBS environment; Step 24: baseline test: testing the basic transfer curve of the CNT-FET sensor, using 0.1*PBS solution as a blank group for calculating the liquid environment baseline of the sensor response as a background signal; Step 25, mixed organophosphorus pesticide test: test different concentrations of mixed organophosphorus pesticide solution, the concentration is increased in turn, respectively 100ag / mL, 1fg / mL, 10fg / mL, 100fg / mL, 1pg / mL, 20μL sample is taken by pipette, incubate each CNT-FET sensor for 5 minutes, to keep all CNT-FET sensor measurement same ion intensity, then set test parameters and collect sensor transfer characteristics according to the same test process of test step 22. 6.The ultra-low concentration mixed organophosphorus pesticide detection method based on CNT-FET and machine learning according to claim 5, wherein, In step 23, the mixed organophosphorus pesticide solution includes a mixture of three organophosphorus pesticides, all prepared in 0.1*PBS environment, including: dimethoate & dimethoate, dimethoate & methyl parathion, dimethoate & methyl parathion with same concentration 1:1 mixture and different concentration 1:1 mixture. 7.The ultra-low concentration mixed organophosphorus pesticide detection method based on CNT-FET and machine learning according to claim 1, wherein, In step 4, the recognition model based on CNN convolutional neural network is constructed, including: The recognition model based on CNN convolutional neural network includes input layer, two convolutional layers, two fully connected layers, dropout layer and output layer, and CBAM double attention mechanism composed of channel attention module and spatial attention module is fused, wherein: The channel attention module learns the weight for each channel feature map according to the global average pooling technology and fully connected operation logic, and the learned weight is fed back to the original feature map in reverse, so as to selectively focus on the features of different channels in the convolutional neural network; The spatial attention module focuses on the information of different spatial positions in the feature map according to the dual operation of maximum and average pooling in spatial dimension, performs weight learning process on the feature map, and accurately feeds back the weight, so that the features of key spatial positions are enhanced; Through the fusion of channel attention module and spatial attention module in CNN model, the feature representation is optimized.

8. The CNT-FET and machine learning based ultra-low concentration mixed organophosphorus pesticide detection method according to claim 7, characterized in that, In step 4, in the recognition model based on CNN convolutional neural network, the input layer and two convolutional layers are designed as follows: The data input of the input layer receives the image data converted from the transfer characteristic data, and the input layer serves as the data entrance of the model, providing data input for subsequent convolution operation; The first convolutional layer extracts features from the image data of the input layer, and the size of the first convolutional layer is set to 50×50×4; by performing convolution operation on the image data, the basic features of the image are extracted, including the local information of the image corresponding to different concentrations of mixed organophosphorus pesticides; then the convolution output is activated using ReLU function; The first maximum pooling layer performs downsampling, and the maximum pooling operation with size of 3×3 is performed on the output data of the first convolutional layer after ReLU activation, which reduces the dimension from 50×50×4 to 25×25×8, reducing the complexity of subsequent calculation; The second convolutional layer further excavates deep features in the data and captures higher-level image feature patterns by performing a convolution operation on the data output by the first max-pooling layer, thereby providing representative feature expression for subsequent classification tasks; and then the convolution output is activated using a ReLU function. The second max-pooling layer performs secondary downsampling, and a 3×3 max-pooling operation is performed on the output data of the second convolutional layer after ReLU activation to downsample the data from 25×25×8 to 12×12×32, thereby reducing the data volume and enhancing the robustness and expression ability of the features, and providing data for subsequent attention mechanism processing. 9.The ultra-low concentration mixed organophosphorus pesticide detection method based on CNT-FET and machine learning according to claim 8, wherein, The CBAM dual attention mechanism is composed of a channel attention module and a spatial attention module. The channel attention module performs multi-scale feature extraction and spatial attention processing on the data output by the second max-pooling layer. First, two 3×3 pooling operations and one 1×1 pooling operation are performed to extract spatial feature information of the data from multiple angles and capture feature changes at different positions and scales in the image. The spatial attention module performs feature weighting processing. Two channel attention mechanisms are introduced, and full connection operation is performed on the results of the spatial attention mechanism processing. The features of different spatial attention channels are weighted and adjusted through the full connection layer, and corresponding weights are assigned to each channel feature according to its importance, thereby enhancing channel feature expression, suppressing interference information, and improving feature selectivity. Then, the results processed by the spatial attention module and the channel attention module are fused, activated by a Sigmoid function, and normalized to map the fused feature output value to the [0, 1] interval, thereby assigning a normalized weight to each feature.

10. The CNT-FET and machine learning based ultra-low concentration mixed organophosphorus pesticide detection method according to any one of claims 7-9, characterized in that, In the CNN convolutional neural network-based recognition model in step 4, the two full connection layers, the dropout layer, and the output layer are designed as follows: The first full connection layer integrates features. The data set processed by the dual attention mechanism is input into the first full connection layer, which integrates and nonlinearly transforms the output features, thereby fusing information in different dimensions and feature spaces and extracting more representative feature representations. The second full connection layer refines features. The output data after the first full connection layer is received, and the second full connection layer further refines and transforms the features, thereby further excavating complex relationships between the features and providing feature input for the final classification task. The output layer determines the classification result. The output layer determines the classification result of the different concentrations of mixed organophosphorus pesticides based on the output result of the second full connection layer, outputs the final classification result of the mixed organophosphorus pesticides, and completes the entire classification and recognition task.

Images