Method for detecting tomato root rot based on volatile group information

Abstract

The present application relates to the technical field of tomato disease prevention and treatment, in particular to a method for detecting tomato root rot based on volatile group information; it is divided into four steps of volatile gas collection, data analysis, intelligent diagnosis and result display, supplemented by cloud database and model updating; first, the volatile compounds of tomato roots are obtained through high-sensitivity sampling equipment, the volatile components are analyzed by GC-IMS technology, the data of <component, concentration> are obtained, and the characteristic vector is generated, then the generation and verification of the audit identifier are combined to ensure the compliance and integrity of the extracted data, then based on the classification model, the health score and the classification result are combined to accurately judge the disease grade and provide prevention and treatment suggestions, and the diagnosis result is presented in a visual way; the cloud database continuously accumulates historical data, optimizes the classification model, and improves the system adaptability and detection effect. The present application realizes early and accurate diagnosis and intelligent management of tomato root rot, and improves the agricultural disease detection efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tomato disease prevention and treatment, in particular to a method for detecting tomato root rot based on volatile group information. BACKGROUND

[0002] Tomato is an important economic crop worldwide, widely planted in various agricultural production areas, and its yield and quality are directly related to agricultural economic benefits. However, tomato is easily attacked by diseases during growth, among which root rot is a common and serious soil-borne disease caused by fungi, bacteria or other pathogens, which can cause plant root rot, growth weakness, and ultimately affect fruit yield or even plant death. This disease has the characteristics of strong hidden transmission and fast outbreak, and once it occurs, it is difficult to control in time and can cause serious economic losses.

[0003] A tomato plant disease monitoring method is disclosed in Chinese invention patent No. CN117969771B, which specifically comprises: initializing a tomato planting greenhouse scene, arranging a gas detector in the scene, using the gas detector to determine the concentration, combining the concentration characteristic values obtained from each determination result, and obtaining the nitrogen abnormality characteristic through continuous measurement of the concentration characteristic values. Finally, the disease monitoring and early warning are carried out according to the obtained nitrogen abnormality characteristic; the regularity of concentration change degree in the greenhouse is effectively identified, and the occurrence degree of nitrogen dioxide discharge abnormal behavior in the tomato planting greenhouse scene is quantified, thereby providing reliable mathematical basis for analyzing and distinguishing non-infectious diseases caused by harmful gases and infectious diseases caused by pathogens, thereby more scientifically and efficiently distinguishing diseases such as white spots, water-stained lesions, wilting, and flower and fruit drop of tomato plants, and further enhancing the prevention and control efficiency and quality.

[0004] However, the current disease detection research based on volatile group information is still in the exploratory stage and has not been widely applied in agricultural actual production. The main bottlenecks include: how to efficiently extract characteristic volatiles related to diseases from complex gas mixtures; how to efficiently process and diagnose volatile group data through intelligent analysis models; and how to transform research results into convenient and cost-controllable agricultural application technology; therefore, it is of great significance to develop an efficient system suitable for tomato root rot detection by combining advanced volatile group detection technology and intelligent data analysis methods to improve disease management efficiency and realize precision agriculture. SUMMARY

[0005] The present application aims to solve the problems in the background art and proposes a method for detecting tomato root rot based on volatile group information.

[0006] The technical solution of the present application is a method for detecting tomato root rot based on volatile group information, which comprises

[0007] S1, construct a tomato root rot detection system based on volatile group information;

[0008] S2, use a high-sensitivity volatile gas sampling device and arrange it in the root area of the tomato plant to ensure that the device can capture the volatile compounds released by the roots, determine the sampling time and environmental conditions, and conduct periodic sampling to collect volatile gases from the tomato roots;

[0009] S3, filter particulate matter or water vapor that may be mixed into the volatile gases from the tomato roots to obtain a detection sample, use GC-IMS technology to classify and detect the gas components in the detection sample, generate <component, concentration> summary data, generate a feature vector, and generate an audit identifier for the <component, concentration> summary data and the feature vector;

[0010] S4, audit the compliance and integrity of the extracted feature vector and <component, concentration> summary data, input the feature vector into a pre-trained classification model to obtain a diagnosis result and calculate a health score, and based on the health score and the diagnosis result, perform double detection on the state of the tomato plant, based on a root rot grade division rule table and a correction rule, output the state grade of the tomato plant and the corresponding strategy;

[0011] S5, display the received tomato plant state grade and corresponding strategy based on a visualization device, and the user manages according to the tomato plant state grade and corresponding strategy;

[0012] S6, upload the volatile group data and diagnosis results of each detection to the cloud to form historical data accumulation, and periodically retrain the classification model based on new data.

[0013] Preferably, a tomato root rot detection system based on volatile group information comprises:

[0014] A volatile group information collection module for collecting volatile gases from the roots of tomatoes;

[0015] A data analysis module for analyzing volatile group components and generating volatile group spectrum data;

[0016] An intelligent diagnosis module that uses a pre-trained deep learning model to analyze volatile group data and output root rot diagnosis results;

[0017] A user interaction module for displaying root rot detection results and prevention suggestions;

[0018] A cloud database and model update module for storing tomato root rot volatile group feature libraries and updating detection models.

[0019] Preferably, the generation process of the feature vector is as follows:

[0020] S31. Standardize the concentration data of each volatile substance in the <component, concentration> summary data, normalizing the concentration to the [0, 1] interval:

[0021]

[0022] In the formula, C norm (i) represents the standardized concentration of the i-th volatile compound; C VOCs (i) represents the detection concentration of the i-th volatile compound; C max and C min These represent the maximum and minimum concentrations of all volatile compounds in the sample, respectively.

[0023] S32. Calculate the relative concentrations of each component in the sample:

[0024]

[0025] In the formula, R i Represents the relative concentration ratio of the i-th volatile compound; n represents the number of all volatile compound species in the volatile group characteristic spectral library; C norm (j) represents the standardized concentration of the j-th volatile compound;

[0026] S33, All relative concentration ratios R i Combined into an eigenvector R = [R1, R2, ..., R n ].

[0027] Preferably, the process for generating the audit identifier is as follows:

[0028] S4. Choose a large prime number p = 2281, and the corresponding irreducible primitive polynomial of degree p is q = x. 2281 +x 715 +1, randomly select an element Satisfy g n mod q = 1;

[0029] Where n is a finite field multiplication group The level And p is a prime number, and satisfies n=2 p -1 is a Mersenne prime; x is a polynomial indeterminate variable; g≠1;

[0030] S42. Select a random number a∈(0,n), and calculate the audit code CA=g. a mod q, and expose the parameter set {p, g, n, q, CA};

[0031] S43. Convert the feature vector R and the <component, concentration> summary data Data into binary string data, and concatenate them together to obtain the encoded data Bidata;

[0032] S44, calculate identifier generation parameter Idg: Idg = g Bidata mod p;

[0033] S45, randomly select an integer k∈(0, n), calculate first-order audit identifier IA1 = g k mod q;

[0034] Calculate second-order audit identifier IA2 = Bidata x k + a x IA1 mod n;

[0035] S46, generate audit identifier IA = [IA1, IA2].

[0036] Preferably, the audit process of the compliance and integrity of the extracted feature vector and the <component, concentration> summary data is as follows:

[0037] S51, convert the extracted feature vector R and the <component, concentration> summary data Data into binary string data, and splice them together to obtain the encoding data Bidata';

[0038] S52, calculate the first-order audit parameter PI1:

[0039] S53, calculate the second-order audit parameter PI2:

[0040] Wherein, CA is the audit code;

[0041] S54, if PI1 = PI2, the extracted {[IA1, IA2], Data, R} audit passes, that is, the compliance and integrity of the extracted feature vector R and the <component, concentration> summary data Data passes.

[0042] Preferably, the generation method of the diagnosis result is as follows:

[0043] S61, select a deep fully connected neural network as a classification model, and the network structure is as follows:

[0044] S6101, input layer: obtain n-dimensional volatile group feature vector R = [R1, R2,..., R n ];

[0045] S6102, first hidden layer: extract high-order nonlinear features, including 2n neurons and ReLU activation function:

[0046] Wherein, n represents the input dimension;

[0047] Second hidden layer: further compress the feature representation, preserve key information, including n neurons and ReLU activation function:

[0048] Regularization: add Dropout between the second hidden layer and the third hidden layer;

[0049] Third hidden layer: provide low-dimensional feature representation for classification layer, including n neurons and ReLU activation function:

[0050] where, represents half of the input dimension, rounded up;

[0051] S6103, output layer: contains 4 neurons, uses Softmax activation function, outputs normalized probability of each state:

[0052]

[0053] where, P class,i represents the probability of the i-th category; z class,i represents the i-th category score output by the neural network; C represents the total number of root rot state categories;

[0054] S62, based on the normalized probability of each state, predict the state probability, and output the diagnosis result according to the maximum probability principle:

[0055] P current = [P healthy , P early , P mid , P severe ];

[0056]

[0057] where, P current represents the predicted state probability output by the classification model; P healthy , P early , P mid and P severe represent the probability of the plant being in a healthy state, the probability of early root rot, the probability of medium root rot, and the probability of severe root rot, respectively; represents the output diagnosis result.

[0058] Preferably, the construction process of the classification model is as follows:

[0059] S71, obtain historical feature data of tomato root rot, and use sliding mean to remove sensor noise:

[0060]

[0061] where, Ri denotes the feature value of the i-th sample after denoising; k denotes the size of the sliding window; R j denotes the j-th dimension of the original feature vector;

[0062] S72, divide the data set, 70% as the training set, 20% as the validation set, and 10% as the test set;

[0063] S73, define the loss function and accuracy:

[0064]

[0065]

[0066] In the formula, L represents the loss function; m represents the number of batch samples; C represents the number of categories; y i,c denotes the true label of sample i belonging to category c; P i,c denotes the probability of sample i belonging to category c predicted by the model; TP denotes true positive; TN denotes true negative; FP denotes false positive; FN denotes false negative;

[0067] S74, define the optimization algorithm, combined with momentum and adaptive learning rate adjustment, to quickly find the global optimal solution:

[0068] The parameter update formula is:

[0069]

[0070] In the formula, θ t denotes the parameter after the t-th update; θ t-1 denotes the parameter after the t-1-th update; α denotes the learning rate; denotes the first-order momentum correction term, which is the exponential weighted average of the gradient; denotes the second-order momentum correction term, which is the exponential weighted average of the gradient square; ε denotes the smoothing factor, used to prevent the denominator from being 0;

[0071] The momentum update formula is:

[0072]

[0073] In the formula, m t denotes the first-order momentum, which is the moving average of the gradient; v t denotes the second-order momentum, which is the average of the gradient square; β1 and β2 respectively denote the momentum hyperparameters; g t denotes the current gradient;

[0074] Learning rate adjustment: reduce the learning rate by 10% every 10 iterations to stabilize convergence;

[0075] S75, train to get the classification model.

[0076] Preferably, the training process of the classification model is as follows:

[0077] S81, initialize the weights and biases of the neural network, and use the Xavier initialization method to ensure gradient stability;

[0078] The initialization formula is:

[0079]

[0080] In the formula, W represents the weight matrix; n in and n out respectively represent the input and output neuron numbers of the current layer; u represents the upper and lower limits of uniform distribution, which is used to initialize the initial value of the neural network weight W,

[0081] S82, divide the training set into multiple small batches, and randomly select a batch of small batch samples from the training set each time, and perform forward and backward propagation;

[0082] S83, input the feature vector through each layer of the network, calculate the output probability distribution P, and calculate the loss value L;

[0083] S84, calculate the gradient of the loss function to each parameter: and adjust the parameters using an optimization algorithm;

[0084] S85, after each iteration, evaluate the loss and accuracy using the validation set to observe whether the model performance improves;

[0085] S86, if the validation set loss does not decrease for 10 consecutive iterations or reaches the set maximum number of training, stop training, otherwise repeat S81-S85;

[0086] S87, test the model with test set data, and if the model meets the expectations, save the trained model.

[0087] Preferably, the correction rule is: the range of health score S is used as the basic standard for root rot grade, and the classification result P~P current provides type information of root rot, which is used to correct boundary conditions: if P~P current = early, but S≥0.6, directly upgrade the state to "severe disease"; if P~P current = severe, but S<0.3, downgrade the state to "moderate disease";

[0088] Where, early represents early disease; severe represents severe disease.

[0089] Compared with the prior art, the above technical scheme of the present application has the following beneficial technical effects:

[0090] The above scheme of tomato root rot detection based on volatile group information has the following beneficial technical effects:

[0091] The present application designs a method for tomato root rot detection based on volatile group information. This scheme integrates multiple advanced technologies and methods, not only improving the detection efficiency and accuracy of tomato root rot, but also having the advantages of real-time monitoring, user-friendly, data security, strong adaptability, and positive significance for promoting the development of modern agricultural technology:

[0092] (1) Improve diagnostic accuracy: By using GC-IMS technology to analyze the volatile gases released by the roots of tomatoes, combined with advanced deep learning models, different root rot states can be more accurately identified and distinguished, and combined with the generation and verification of audit identifiers, the compliance and integrity of the extracted data are guaranteed. The analysis method based on chemical composition is more objective and accurate than traditional visual inspection or experience-based judgment;

[0093] (2) Early diagnosis and prevention: The present application can realize early diagnosis of tomato root rot, so that farmers can take appropriate prevention measures at the early stage of disease development, effectively avoiding further deterioration and spread of the disease;

[0094] (3) Data-driven decision support: Data collected and analyzed by cloud database can help farmers better understand the development patterns and influencing factors of diseases, and provide scientific basis for formulating more effective management strategies;

[0095] (4) User-friendly interface: The mobile application or PC platform provided makes it easy for users to obtain detection results and prevention recommendations, improving the usability and ease of use of the system;

[0096] (5) Continuous learning and optimization: With the accumulation of more data and continuous training of classification models, the diagnostic ability of the system will be further improved, and it can adapt to the changes of different varieties of tomatoes and multiple disease characteristics;

[0097] (6) Promote agricultural modernization: The implementation of this scheme helps to promote the development of traditional agriculture towards intelligent and precise direction, which conforms to the trend of sustainable development of modern agriculture. BRIEF DESCRIPTION OF DRAWINGS

[0098] Figure 1 A method flowchart of a method for tomato root rot detection based on volatile group information is provided for the present application;

[0099] Figure 2 A system architecture diagram of a tomato root rot detection system based on volatile group information is provided for the present application. DETAILED DESCRIPTION

[0100] In one embodiment, as shown in Figure 1 and Figure 2 The present application proposes a method for detecting tomato root rot based on volatile group information, as shown in Figure 1 The method comprises the following specific implementation steps:

[0101] S1. Constructing a tomato root rot detection system based on volatile group information, as shown in Figure 2 The system comprises a volatile group information acquisition module, a data analysis module, an intelligent diagnosis module, a user interaction module, and a cloud database and model updating module.

[0102] The volatile group information acquisition module is used to collect volatile gases from the roots of tomatoes.

[0103] The data analysis module is integrated with a GC-IMS (Gas Chromatography-Ion Mobility Spectrometry) instrument for analyzing volatile components and generating volatile group spectrum data.

[0104] The intelligent diagnosis module uses a pre-trained deep learning model to analyze volatile group data and output root rot diagnosis results.

[0105] The user interaction module provides a mobile application or PC platform to display root rot detection results and prevention suggestions.

[0106] The cloud database and model updating module is used to store the volatile group characteristics library of tomato root rot and update the detection model.

[0107] S2. The volatile group information acquisition module uses high-sensitivity volatile gas sampling equipment (including but not limited to a multi-hole adsorption tube or an electronic nose) and arranges it in the root area of the tomato plant to ensure that the equipment can capture volatile compounds released by the roots, determines the sampling time and environmental conditions (such as sampling when the light is stable during the day), avoids interference from soil humidity or external contaminated gases, and performs periodic sampling to collect volatile gases from the roots of tomatoes.

[0108] S3. The data analysis module analyzes the collected volatile gases from the roots of tomatoes, and the specific implementation process is as follows:

[0109] S31. Use a physical filtration device to remove particulate matter or water vapor that may be mixed into the volatile gases from the roots of tomatoes obtained by sampling, to ensure sample purity and obtain a detection sample.

[0110] S32, using GC-IMS technology, classifying and detecting the gas components in the detection sample, separating the volatile components in the sample, detecting and identifying the components according to the mobility of the compounds in the electric field after ionization, recording the gas chromatography retention time (t GC ), IMS migration time (t IMS ) and corresponding concentration of each volatile, forming a volatile group characteristic spectrum library, recording each volatile component and its concentration data, and generating <component, concentration> summary data;

[0111] S33, standardizing the volatile concentration data, normalizing the concentration to the [0, 1] interval, and eliminating the influence of different sample concentration ranges:

[0112]

[0113] In the formula, C norm (i) represents the standardized concentration of the i-th volatile; C VOCs (i) represents the detection concentration of the i-th volatile; C max and C min represent the maximum and minimum concentrations of all volatiles in the sample, respectively;

[0114] S34, calculating the relative concentration of each component in the sample:

[0115]

[0116] In the formula, R i represents the relative concentration ratio of the i-th volatile; n represents the number of all volatile species in the volatile group characteristic spectrum library; C norm (j) represents the standardized concentration of the j-th volatile;

[0117] S35, combining all the relative concentration ratios R i into a feature vector R = [R1, R2,..., R n ];

[0118] S36, generating an audit identifier for the feature vector and <component, concentration> summary data, and the generation process is as follows:

[0119] S3601, select a large prime number p = 2281, and the corresponding p-th irreducible polynomial q = x 2281 +x 715 +1, randomly select an element satisfying g n mod q = 1;

[0120] where n is the order of the multiplicative group 2p of the finite field F ​and p is a prime number, and satisfies n = 2 p -1 is a Mersenne prime; x is a polynomial indeterminate; g≠1;

[0121] S3602, select a random number a∈(0, n), calculate the audit code CA = g a mod q, and transmit {p, g, n, q, CA} to the intelligent diagnosis module;

[0122] S3603, convert the feature vector R and <component, concentration> summary data Data into binary string data, and splice them together to obtain encoding data Bidata;

[0123] S3604, calculate the identifier generation parameter Idg: Idg = g Bidata modp;

[0124] S3605, randomly select an integer k∈(0, n), calculate the first-order audit identifier IA1 = g k mod q;

[0125] Calculate the second-order audit identifier IA2 = Bidata x k + a x IA1 mod n;

[0126] S3606, generate the audit identifier IA = [IA1, IA2];

[0127] S37, transmit {[IA1, IA2], Data, R} to the user interaction module, and the cloud database and model updating module.

[0128] S4, the intelligent diagnosis module extracts {[IA1, IA2], Data, R} from the cloud database and model updating module, constructs a classification model suitable for the volatile group characteristics based on the feature vector R, and distinguishes the root rot state of the tomato plant, to accurately diagnose the tomato root rot. The specific implementation process is as follows:

[0129] S41, audit the extracted {[IA1, IA2], Data, R}, to ensure the compliance and integrity of the extracted feature vector R and <component, concentration> summary data Data. The auditing process is as follows:

[0130] S4101, convert the extracted feature vector R and <component, concentration> summary data Data into binary string data, and splice them together to obtain encoding data Bidata';

[0131] S4102, calculate the first-order audit parameter PI1: PI1 = g IA2 ;

[0132] S4103, calculate the second-order audit parameter PI2: PI2 = (IA1)Bidata CA IA1 modq;

[0133] wherein CA is an audit code;

[0134] S4104, if PI1=PI2, the extracted {[IA1, IA2], Data, R} from the cloud database and the model update module is audited, that is, the compliance and integrity of the extracted feature vector R and the <component, concentration> summary data Data can be guaranteed;

[0135] S42, a deep fully connected neural network (FCNN) is selected as the core model, and the network structure is as follows:

[0136] S4201, input layer: obtain n-dimensional volatile group feature vector R=[R1, R2,..., R n ];

[0137] S4202, hidden layer:

[0138] (1) first hidden layer: extract high-order nonlinear features, including neurons and activation functions:

[0139] Number of neurons: 2n, i.e. twice the input dimension;

[0140] Activation function: ReLU, rectified linear unit;

[0141] (2) second hidden layer: further compress the feature representation and retain key information, including neurons and activation functions:

[0142] Number of neurons: n;

[0143] Activation function: ReLU;

[0144] (3) regularization: add Dropout between hidden layers to avoid overfitting, and the Dropout probability is set to 0.3 in this embodiment, i.e. 30% of the neuron connections are randomly discarded;

[0145] (4) third hidden layer: provide low-dimensional feature representation for the classification layer, including neurons and activation functions:

[0146] Number of neurons: i.e. half of the input dimension, rounded up;

[0147] Activation function: ReLU;

[0148] S4203, output layer: contains 4 neurons, uses Softmax activation function, and outputs the normalized probability of each state:

[0149]

[0150] wherein P class,i represents the probability of the i-th category; z class,i represents the i-th category score of the neural network output; C represents the total number of root rot state categories, and in the embodiment, 4 categories are adopted, i.e. healthy, early, medium and severe;

[0151] S43, input the feature vector R = [R1, R2,..., R n ] into the pre-trained classification model, predict the state probability, and output the diagnosis result according to the maximum probability principle:

[0152] P current = [P healthy , P early , P mid , P severe ];

[0153]

[0154] wherein P current represents the predicted state probability output by the classification model; P healthy , P early , P mid and P severe represent the probability of the plant being in a healthy state, the probability of early root rot, the probability of medium root rot and the probability of severe root rot, respectively; represents the output diagnosis result;

[0155] S44, calculate the confidence Confidence = max (P current ), if the confidence Confidence ≥ 0.8, directly adopt the classification result Otherwise, perform the following S45;

[0156] S45, extract each volatile component and its concentration data recorded in the <component, concentration> summary data, combine the health score quantitative evaluation, analyze the deviation degree, perform secondary verification on the classification result, and further refine the root rot diagnosis and degree, and the specific implementation steps are as follows:

[0157] Calculate the deviation degree of the current plant from the healthy state, i.e. the health score:

[0158]

[0159] wherein S represents the health score, and the larger the S, the more severe the root rot of the plant; R current (i) represents the i-th item of the current feature; R healthy(i) represents the health characteristic corresponding to the i-th item of the current characteristic; n represents the number of all volatile species in the <component, concentration> summary data;

[0160] S46. Based on classification results The health score S, the root rot disease classification rule table (as shown in Table 1), and the correction rules are used to classify the status of tomato plants and output the corresponding strategies.

[0161] Table 1. Rules for Classifying the Levels of Root Rot

[0162]

[0163]

[0164] The revised rules are as follows: The range of the health score S is used as the basic standard for the classification of root rot, and the classification results are... Provides information on the type of root rot to correct certain boundary cases:

[0165] if However, if S≥0.6, the status is directly upgraded to "severe disease"; if However, if S < 0.3, the condition will be downgraded to "moderate disease";

[0166] S47. Transmit the root status level of tomato plants and the corresponding strategy to the user interaction module and the cloud database and model update module.

[0167] S5. The user interaction module displays the received root status level of tomato plants and corresponding strategies based on a visualization device. It supports access via mobile application or PC. Users can manage tomato plants according to the root status level and corresponding strategies.

[0168] The S6 cloud database and model update module uploads the volatile group data and diagnostic results of each test to the cloud, forming a historical data accumulation. It regularly retrains the classification model based on the new data to optimize the classification effect, continuously enriches the volatile group feature library, covers different varieties of tomatoes and various disease characteristics, and enhances the applicability of the system.

[0169] Example 2: This invention proposes a method for detecting tomato root rot based on volatile group information. Compared to Example 1, this example also includes a method for constructing and training a classification model, the specific implementation steps of which are as follows:

[0170] S1. Obtain historical characteristic data of tomato root rot, and use moving average or low-pass filtering to remove sensor noise to ensure accurate characteristic values:

[0171]

[0172] In the formula, Ri denotes the feature value of the i-th sample after denoising; k denotes the size of the sliding window; R j denotes the j-th dimension of the original feature vector.

[0173] S2, divide the data set, 70% as the training set, 20% as the validation set, and 10% as the test set.

[0174] S3, define the loss function and accuracy:

[0175]

[0176]

[0177] In the formula, L represents the loss function; m represents the number of batch samples; C represents the number of categories, which is 4 in this embodiment; y i,c denotes the true label (one-hot encoding) of sample i belonging to category c; P i,c denotes the probability of sample i belonging to category c predicted by the model; TP denotes true positive; TN denotes true negative; FP denotes false positive; and FN denotes false negative.

[0178] S4, define the optimization algorithm, combined with momentum and adaptive learning rate adjustment, to quickly find the global optimal solution:

[0179] The parameter update formula is:

[0180]

[0181] In the formula, θ t denotes the parameter after the t-th update; θ t-1 denotes the parameter after the t-1-th update; α denotes the learning rate, which is set to 0.001 in this embodiment; denotes the first-order momentum correction term, i.e., the exponential weighted average of the gradient; denotes the second-order momentum correction term, i.e., the exponential weighted average of the gradient square; ε denotes the smoothing factor, used to prevent the denominator from being 0;

[0182] The momentum update formula is:

[0183]

[0184] In the formula, m t denotes the first-order momentum, i.e., the moving average of the gradient; v t denotes the second-order momentum, i.e., the average of the gradient square; β1 and β2 respectively denote the momentum hyperparameters; g t denotes the current gradient;

[0185] Learning rate adjustment: the learning rate is decayed by 10% every 10 epochs (iterations) to stabilize convergence.

[0186] S5, training the classification model, the training steps are as follows:

[0187] S51, initializing the weights and biases of the neural network, using the Xavier initialization method to ensure gradient stability;

[0188] The initialization formula is:

[0189]

[0190] In the formula, W represents the weight matrix; n in and n out respectively represent the input and output neuron numbers of the current layer; u represents the upper and lower limits of uniform distribution, which is used to initialize the initial value of the neural network weight W,

[0191] S52, divide the training set into multiple mini-batches (batch), in this embodiment, the batch size is 32, randomly extract a batch of samples from the training set each time, and perform forward and backward propagation;

[0192] S53, input the feature vector through each layer of the network, calculate the output probability distribution P, and calculate the loss value L;

[0193] S54, calculate the gradient of the loss function to each parameter: and adjust the parameters using the optimization algorithm;

[0194] S55, after each epoch, evaluate the loss and accuracy using the validation set, and observe whether the model performance is improved;

[0195] S56, if the validation set loss does not decrease for 10 consecutive epochs or reaches the set maximum training number (the maximum training number set in this embodiment is 50 epochs), stop training, otherwise repeat S51-S55;

[0196] S57, test the model with test set data, if the model meets the expectation, save the trained model as the classification model.

[0197] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A method for detecting tomato root rot based on volatile group information, characterized in that, include: S1. Construct a tomato root rot detection system based on volatile group information; S2. Use a high-sensitivity volatile gas sampling device and place it in the root area of ​​the tomato plant to ensure that the device can capture volatile compounds released from the roots. Determine the sampling time and environmental conditions, conduct periodic sampling, and collect volatile gases from the tomato roots. S3. Filter out particulate matter or water vapor that may be mixed in with the volatile gases from the tomato roots to obtain a test sample. Use GC-IMS technology to classify and detect the gas components in the test sample, generate <component, concentration> summary data, generate feature vectors, and generate audit identifiers for the <component, concentration> summary data and feature vectors. The process of generating the audit identifier is as follows: A1. Choose a large prime number p = 2281, the corresponding irreducible primitive polynomial of degree p is q = x. 2281 +x 715 +1, randomly select an element Satisfy g n mod q = 1; Where n is the finite field F 2p multiplication group The level And p is a prime number, and satisfies n=2 p -1 is a Mersenne prime; x is a polynomial indeterminate variable; g≠1; A2. Select a random number a∈(0,n), and calculate the audit code CA=g. a mod q, and expose the parameter set {p, g, n, q, CA}; A3. Convert the feature vector R and the <component, concentration> summary data Data into binary string data, and concatenate them together to obtain the encoded data Bidata; A4. Calculate the identifier generation parameter Idg: Idg=g Bidata modp; A5. Randomly select an integer k∈(0,n), and calculate the first-order audit identifier IA1=g. k mod q; Calculate the second-order audit identifier IA2 = Bidata × k + a × IA1 mod n; A6. Generate audit identifier IA = [IA1, IA2]; S4. Audit the compliance and completeness of the extracted feature vectors and <component, concentration> summary data. Input the feature vectors into the pre-trained classification model to obtain the diagnosis results and calculate the health score. Based on the health score and diagnosis results, perform dual detection on the tomato plant status. Based on the root rot disease level classification rule table and correction rules, output the tomato plant status level and corresponding strategy. The audit process for verifying the compliance and completeness of the extracted feature vectors and <component, concentration> summary data is as follows: B1. Convert the extracted feature vector R and the <component, concentration> summary data Data into binary string data, and concatenate them together to obtain the encoded data Bidata'. B2. Calculate the first-order audit parameter PI1: PI1 = g I A2; B3. Calculate the second-order audit parameter PI2: Where CA stands for audit code; B4. If PI1 = PI2, then the extracted {[IA1, IA2], Data, R} passes the audit, meaning that the compliance and completeness of the extracted feature vector R and the <component, concentration> summary data Data pass the audit. S5. The received tomato plant status level and corresponding strategy are displayed on the visualization device, and the user manages the tomato plant according to the tomato plant status level and corresponding strategy. S6. Upload the volatile group data and diagnostic results of each test to the cloud to form historical data accumulation, and retrain the classification model regularly based on the new data.

2. The method for detecting tomato root rot based on volatile group information according to claim 1, characterized in that, A system for detecting tomato root rot based on volatile group information includes: The volatile matter information acquisition module is used to collect volatile gases from the roots of tomatoes; The data analysis module is used to analyze the volatile components and generate volatile group spectrum data; The intelligent diagnostic module uses a pre-trained deep learning model to analyze volatile group data and output root rot diagnosis results; The user interaction module is used to display root rot detection results and prevention and control suggestions. The cloud-based database and model update module is used to store the volatile group feature library of tomato root rot and update the detection model.

3. The method for detecting tomato root rot based on volatile group information according to claim 1, characterized in that, The process of generating eigenvectors is as follows: S31. Standardize the concentration data of each volatile substance in the <component, concentration> summary data, normalizing the concentration to the [0, 1] interval: In the formula, C norm (i) represents the standardized concentration of the i-th volatile compound; C VOCs (i) represents the detection concentration of the i-th volatile compound; C max and C min These represent the maximum and minimum concentrations of all volatile compounds in the sample, respectively. S32. Calculate the relative concentrations of each component in the sample: In the formula, R i Represents the relative concentration ratio of the i-th volatile compound; n represents the number of all volatile compound species in the volatile group characteristic spectral library; C norm (j) represents the standardized concentration of the j-th volatile compound; S33, All relative concentration ratios R i Combined into an eigenvector R = [R1, R2, ..., R n ].

4. The method for detecting tomato root rot based on volatile group information according to claim 1, characterized in that, The diagnostic results are generated as follows: S61. A deep fully connected neural network is selected as the classification model. The network structure is as follows: S6101, Input Layer: Obtain the n-dimensional volatile group feature vector R = [R1, R2, ..., R...]. n ]; S6102, First Hidden Layer: Extracts high-order nonlinear features, including 2n neurons and the ReLU activation function: Where n represents the input dimension; Second hidden layer: Further compresses the feature representation while retaining key information, including n neurons and the ReLU activation function: Regularization: Add Dropout between the second and third hidden layers; The third hidden layer: provides low-dimensional feature representations for the classification layer, including One neuron and ReLU activation function: in, This represents half of the input dimension, rounded up. S6103, Output Layer: Contains 4 neurons, uses the Softmax activation function, and outputs the normalized probability of each state: In the formula, P class,i z represents the probability of the i-th category; class,i C represents the score of the i-th category output by the neural network; C represents the total number of root rot status categories. S62. Based on the normalized probability of each state, predict the state probability and output the diagnostic result according to the maximum probability principle: P current =[P healthy ,P early ,P mid ,P severe ]; Among them, P current P represents the predicted state probability output by the classification model. healthy P early P mid and P severe These represent the probabilities of the plant being in a healthy state, the early stage of root rot, the mid-stage of root rot, and the severe stage of root rot, respectively. This indicates the output diagnostic results.

5. The method for detecting tomato root rot based on volatile group information according to claim 4, characterized in that, The process of building a classification model is as follows: S71. Obtain historical characteristic data of tomato root rot and use moving average to remove sensor noise: In the formula, R i ' represents the denoised feature value of the i-th sampling point; k represents the sliding window size; R j Represents the j-th dimension of the original eigenvector; S72. Divide the dataset into three parts: 70% as the training set, 20% as the validation set, and 10% as the test set. S73. Define the loss function and accuracy: In the formula, L represents the loss function; m represents the batch size; C represents the number of classes; y i,c This indicates the true label of sample i belonging to category c; P i,c TP represents the probability that sample i belongs to class c as predicted by the model; TN represents a true positive; FP represents a false positive; FN represents a false negative. S74. Define an optimization algorithm that combines momentum and adaptive learning rate adjustment to quickly find the global optimum: The parameter update formula is: In the formula, θ t θ represents the parameter after the t-th update; t-1 Represents the parameters after the (t-1)th update; α represents the learning rate; This represents the first-order momentum correction term, which is the exponentially weighted average of the gradient. ε represents the second-order momentum correction term, which is the exponentially weighted average of the squared gradient; ε represents the smoothing factor, used to prevent the denominator from being 0. The momentum update formula is: m t =β1m t-1 +(1-β1)g t In the formula, m t This represents the first-order momentum, i.e., the gradient moving average; v t The second-order momentum is represented by the squared gradient; β1 and β2 represent the momentum hyperparameters, respectively; g t Indicates the current gradient; Learning rate adjustment: The learning rate is decreased by 10% every 10 iterations to stabilize convergence; S75. The classification model is obtained through training.

6. The method for detecting tomato root rot based on volatile group information according to claim 5, characterized in that, The training process for the classification model is as follows: S81. Initialize the weights and biases of the neural network, using the Xavier initialization method to ensure gradient stability; The initialization formula is: In the formula, W represents the weight matrix; n in and n out These represent the number of input and output neurons in the current layer, respectively; u represents the upper and lower limits of the uniform distribution, used to initialize the initial values ​​of the neural network weights W. S82. Divide the training set into multiple mini-batches, and randomly select a batch of mini-batches from the training set each time for forward and backward propagation. S83. The input feature vector passes through each layer of the network, the output probability distribution P is calculated, and the loss value L is calculated. S84. Calculate the gradient of the loss function with respect to each parameter: And the parameters are adjusted using optimization algorithms; S85. After each iteration, use the validation set to evaluate the loss and accuracy, and observe whether the model performance has improved. S86. If the validation set loss no longer decreases after 10 consecutive iterations or reaches the set maximum number of training iterations, stop training; otherwise, repeat S81 to S85. S87. Test the model using the test set data. If the model meets expectations, save the trained model.

7. The method for detecting tomato root rot based on volatile group information according to claim 1, characterized in that, The revised rule is: the range of the health score S is used as the basic standard for the classification of root rot, and the classification results are... Provide information on the type of root rot to correct for boundary conditions: if However, if S≥0.6, the status is directly upgraded to "severe disease"; if However, if S < 0.3, the condition will be downgraded to "moderate disease"; Among them, "early" indicates early-stage disease, and "severe" indicates severe disease.

Images