Granary pest prediction system and method based on smell sensor

Abstract

The invention discloses a granary pest prediction system and method based on a smell sensor, and the granary pest prediction system based on the smell sensor provides important technical support and guarantee for grain storage management during monitoring and early warning. The variational mode decomposition module is used for performing vibration mode decomposition on the internal environment data of the granary, and effectively capturing and extracting pest behaviors by extracting potential pest activity signals in the granary; the depth extreme learning machine module optimized by the sparrow search algorithm receives the feature data extracted by the variational mode decomposition module, and performs parameter optimization on a depth extreme learning machine by using the sparrow search algorithm so as to improve the prediction performance and generalization ability of the model; and the prediction model evaluation module is used for comprehensively evaluating the prediction model obtained by training, including verifying the accuracy, robustness and generalization ability of the model, and ensuring the reliability and effectiveness of the model in an actual granary environment.

Description

Technical Field

[0001] The present invention relates to the technical field of granary pest detection in deep learning, and in particular to a granary pest prediction system and method based on an odor sensor. Background Art

[0002] The high frequency outbreak of storage pests poses a serious threat to food and material security, causing significant economic losses. Therefore, it is of great significance to accurately and timely detect pests and reduce the impact of pests on storage, guide scientific storage operations, and promote the continuous development of the storage industry towards high quality and high yield. Traditional storage pest detection tasks require warehouse managers to work on-site, but the storage area is large and there are many types of pests. There are problems such as time-consuming and labor-intensive work and insufficient pest control experts. With the development of machine learning and deep learning, some new technologies have gradually replaced the original manual detection methods. However, machine learning relies on manual extraction of pest feature information, which is easily affected by subjective factors and the extracted features have poor robustness, resulting in low accuracy of detection results. For example, Chinese patent publication number CN106919204A discloses a system and method for regulating the habitat of stored pests; Chinese patent authorization announcement number CN101701906B discloses a method and device for detecting stored pests based on near-infrared hyperspectral imaging technology; Chinese patent publication number CN102706870A discloses a method and system for real-time monitoring of stored pest occurrence information.

[0003] Therefore, it is a problem worth studying to provide a granary pest prediction system and method based on odor sensors with greatly improved feature extraction capability and detection accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a granary pest prediction system and method based on odor sensors with greatly improved feature extraction capabilities and detection accuracy. The granary pest prediction system based on odor sensors of the present invention monitors and warns, providing important technical support and guarantee for grain storage management.

[0005] The object of the present invention is achieved in that: A granary pest prediction system based on an odor sensor includes a variational mode decomposition module for decomposing data in the granary into multiple vibration modes, and effectively capturing and extracting pest behaviors by extracting potential pest activity signals inside the granary; The deep extreme learning machine module is used to receive the vibration mode data output by the variational mode decomposition module and optimize it through the sparrow search algorithm to train the prediction model; The prediction model evaluation module is used to conduct a comprehensive evaluation of the trained prediction model, including verifying the accuracy, robustness and generalization ability of the prediction model, and ensuring the reliability and effectiveness of the prediction model in the actual granary environment.

[0006] The variational mode decomposition module receives and processes multi-dimensional data of the granary, wherein the multi-dimensional data includes temperature, humidity, air pressure and gas composition.

[0007] The deep extreme learning machine module includes multiple deep extreme learning machine networks, each of which is used to process a vibration mode output by the variational modal decomposition module; a sparrow search algorithm is used to optimize the parameters of the deep extreme learning machine network to improve the accuracy and generalization ability of the prediction model.

[0008] In the prediction model evaluation module, a cross-validation method is used to evaluate the trained prediction model to verify its prediction performance on different data sets; model performance indicators are provided; and the quality of the prediction model is objectively evaluated.

[0009] A method for predicting pests in a granary based on an odor sensor comprises the following steps: Step 1: Setting of signal collection distribution points: Arrange multiple sensor nodes in the granary to achieve full coverage of the granary to ensure that the collected signals are not missed and to achieve the purpose of economic rationality based on meeting the detection needs; Step 2: Data collection: The sensor network covers all areas in the granary to collect real-time environmental data inside the granary; the data collection frequency is set according to the actual situation to ensure the timeliness and accuracy of the data; Step 3: Acquisition signal processing: preprocessing the data collected in step 2 to obtain processed data, and preprocessing the collected data includes removing outliers and reducing noise; Step 4: Calculation formula for processed signal; for the data preprocessed in step 3, a variational mode decomposition algorithm is used to decompose and extract multiple modes of the data to separate the source signal samples into multiple characteristic signal combinations. Step 5: Model training; divide the data samples after the variational mode decomposition in step 4 into a training set and a test set, and put the training set into the improved sparrow algorithm to optimize the deep extreme learning machine module for training; Step 6: Actual warehouse inspection: use the trained deep extreme learning machine module in step 5 to predict pests on the grain warehouse odor samples; input the data collected in the grain warehouse into the pest prediction model to obtain the prediction results of storage pests.

[0010] In step 4, the iteration formula for each mode is Among them, u k(n) represents the kth mode in the nth iteration, δ(t) is the Dirac function, λ represents the penalty term coefficient, which is used to control the constraint strength of frequency separation, ω k represents the center frequency of the kth module; Through the above iterative update, the algorithm can adaptively adjust the center frequency and bandwidth of each mode, so as to accurately extract each modal component in the signal; the following formula describes the frequency adjustment process: In the formula, represents the center frequency of the nth iteration of module k, μ represents the step size parameter, represents the state or value of the kth mode in the nth iteration; Each mode of the signal is updated according to the iterative formula during the iteration process to ensure the frequency separation between the modes and maintain the bandwidth smoothness of each mode. The center frequency of the mode is adjusted according to formula (2) to find the best frequency representation; the algorithm convergence during the iteration process depends on the formula: Where μ k (n) represents the state or value of the kth mode in the nth iteration, ε represents the minimum threshold set, which is used to judge the convergence condition of the algorithm; max k ||u k (n+1) -u k (n) ||<ε means that when the changes of all modes are less than the threshold, the algorithm converges; n>N max Indicates that when the number of iterations is greater than or equal to the set maximum number of iterations, the algorithm stops iterating.

[0011] In step 5, optimizing the hyperparameters of the deep extreme learning machine module using the improved sparrow algorithm improved by the sparrow search algorithm includes the following steps: Initialize the sparrow population, including the population size, number of discoverers, number of sparrows warned, dimension of the objective function, upper and lower bounds of parameters, maximum number of iterations, and solution accuracy; Calculate the fitness value for each sparrow, and record the current optimal fitness value and its corresponding position, as well as the worst fitness value and its corresponding position; randomly select sparrows from the sparrow population as warning sparrows; among them, the explorer, which has a larger foraging range and guides the foraging direction of the population, its position is updated as follows: Where M is the maximum number of iterations; X a i，jis the location information of the i-th sparrow in the a-th generation in the population in the j-th dimension; α is a random number between (0,1]; Q is a random number that satisfies the normal distribution; L is a matrix of size i×d with all elements being 1; R 2 ∈[0, 1], ST∈[0.5, 1], are the warning value and safety value respectively; Perform one iteration, update the position information of all sparrows, and recalculate the fitness value of each sparrow; when R 2 ≥ST, indicating that some sparrows have discovered the threat, the discoverer value converges and moves to a safe position; R 2 <ST means that the discoverer can expand the search range, and the value of each dimension of the sparrow individual will decrease; when the sparrow is in a hungry state, its value is controlled to conform to the normal distribution; otherwise, the value is taken near the optimal position, and the value is relatively stable. The position update method of the joiner is shown in the following formula: Where: x ij a+1 is the optimal position occupied by the discoverer in the a+1th iteration, is the global worst position in the a-th iteration, A is a multidimensional matrix with one row and each element is 1 or -1; When the proportion of alerters is 10% to 20% of the total population, the position of the alerters is updated as follows: In the formula, is the current global optimal position; β is the step length parameter; K is a random number between [-1,1], which is the step length control parameter of the sparrow's moving direction; f g is the fitness value of the current individual; f g and f w are the global optimal fitness and the global worst fitness respectively; ε is a constant to avoid the denominator being 0; when f i >f g When , it means that the sparrow is at the edge of the population and is vulnerable to threats from natural enemies, so it needs to move to the middle; when f i =f g , which means that the sparrow in the middle of the population is threatened, and it will randomly move closer to the sparrows in other directions; According to the current state of the sparrow population, update the optimal and worst positions and fitness values ​​of the entire population; determine whether the maximum number of iterations or solution conditions have been reached, if so, output the optimal value, otherwise return to recalculate the fitness of each sparrow; Perform iterative operations until the maximum number of iterations is reached to obtain the global optimal solution and the best fitness value; The global optimal value and the best fitness value are obtained as the input of the DELM algorithm, and the DELM network is continuously trained using the training set until the total error is less than the expected error; The DELM network consists of multiple ELM-AE stacks. In the training process, at the initial stage, the training data X is used as the target output X1=X of the first ELM-AE, and then the output weight β of the first layer is calculated. 1 ; Then, the output matrix H1 of the hidden layer of the first layer of ELM-AE is used as the input and target output of the next ELM-AE, and so on; the last layer is trained using ELM, where H i+1 is the output matrix of the last hidden layer, T is the sample label; the input weight matrix of each hidden layer is W i+1 =β T i+1 ; Solve the output weight β of the last hidden layer of DELM i+1 The formula is: Where: C is the regularization parameter; I is the unit matrix with the same dimension as HTH; H is the hidden layer output matrix, which contains the output of all training samples after the activation function from the input layer to the hidden layer, and each row corresponds to the hidden layer output of a training sample; X is the target value matrix or vector of the training sample, and each row corresponds to the target value of a training sample.

[0012] The beneficial effects of the present invention are as follows: the granary prediction system based on the odor sensor of the present invention can be applied to the prediction and monitoring of pests in various granary environments, realizing real-time monitoring and early warning of pest population dynamics in the granary, and providing important technical support and guarantee for grain storage management; the variational mode decomposition module is used to perform vibration mode decomposition on the internal environmental data of the granary, and effectively capture and extract the pest behavior by extracting potential pest activity signals inside the granary; the deep extreme learning machine module optimized by the sparrow search algorithm receives the feature data extracted by the variational mode decomposition module, and optimizes the parameters of the deep extreme learning machine using the sparrow search algorithm to improve the prediction performance and generalization ability of the model; the prediction model evaluation module is used to comprehensively evaluate the prediction model obtained through training, including verifying the accuracy, robustness and generalization ability of the model, to ensure the reliability and effectiveness of the model in the actual granary environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the present invention; Figure 2 Schematic diagram of the sparrow search algorithm of the present invention. DETAILED DESCRIPTION

[0014] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0015] Embodiment 1: A granary pest prediction system based on an odor sensor includes a variational mode decomposition module, which is used to decompose the data in the granary into multiple vibration modes, and to effectively capture and extract the pest behavior by extracting potential pest activity signals inside the granary; the variational mode decomposition module receives and processes the multidimensional data of the granary, and the multidimensional data includes temperature, humidity, air pressure and gas composition. The variational mode decomposition module can adjust the vibration mode decomposition strategy in real time according to the changes in the internal environment of the granary by adaptively adjusting parameters; combined with machine learning technology, it realizes automatic optimization of the vibration mode decomposition process and improves the robustness and adaptability of the prediction model.

[0016] The deep extreme learning machine module is used to receive the vibration mode data output by the variational mode decomposition module and optimize it through the sparrow search algorithm to train the prediction model; the deep extreme learning machine module includes multiple deep extreme learning machine networks, each of which is used to process a vibration mode output by the variational mode decomposition module; the sparrow search algorithm is used to optimize the parameters of the deep extreme learning machine network to improve the accuracy and generalization ability of the prediction model. The deep extreme learning machine module uses a convolutional neural network (CNN) to extract features from the vibration mode data to improve the abstraction and expression ability of the data representation; the residual network (ResNet) structure is used to construct a deep extreme learning machine network to accelerate the training process and improve the convergence speed and efficiency of the model.

[0017] The prediction model evaluation module is used to comprehensively evaluate the trained prediction model, including verifying the accuracy, robustness and generalization ability of the prediction model, and ensuring the reliability and effectiveness of the prediction model in the actual granary environment; in the prediction model evaluation module, the cross-validation method is used to evaluate the trained prediction model to verify its prediction performance on different data sets; model performance indicators such as accuracy, recall rate, F1 value, etc. are provided to objectively evaluate the quality of the prediction model. The prediction model evaluation module provides customized model evaluation indicators and analysis reports for different types of pests and different grain varieties to help users better understand the scope and limitations of the prediction model; by visually displaying the prediction results and error distribution of the prediction model, users can intuitively evaluate the prediction ability and stability of the model. Embodiment 2:

[0018] A method for predicting pests in granaries based on odor sensors, taking a tall flat warehouse as an example, Figure 1 As shown, the following steps are included: Step 1: Setting up signal collection distribution points: Arrange multiple sensor nodes in the granary; set up 32 gas (odor) collection points in the granary.

[0019] Step 2: Data collection: The sensor network covers all areas in the granary to collect real-time internal environmental data of the granary; the data collection frequency is set according to actual conditions to ensure the timeliness and accuracy of the data.

[0020] Step 3: Acquisition signal processing: preprocessing the data collected in step 2 to obtain processed data, and the preprocessing of the collected data includes removing outliers and noise reduction.

[0021] Step 4: Calculation formula for processed signal: For the data preprocessed in step 3, the variational mode decomposition algorithm is used to decompose the multiple modes of the extracted data, and the source signal samples are separated into multiple characteristic signal combinations. In step 4, the iterative formula for each mode is Among them, u k (n) represents the kth mode in the nth iteration, δ(t) is the Dirac function, λ represents the penalty term coefficient, which is used to control the constraint strength of frequency separation, ω k represents the center frequency of the kth module; Through the above iterative update, the algorithm can adaptively adjust the center frequency and bandwidth of each mode, so as to accurately extract each modal component in the signal; the following formula describes the frequency adjustment process: In the formula, ω k (n) The center frequency of the nth iteration of module k, μ represents the step size parameter, μ k (n) , represents the state or value of the kth mode in the nth iteration; Each mode of the signal is updated according to the iterative formula during the iteration process to ensure the frequency separation between the modes and maintain the bandwidth smoothness of each mode. The center frequency of the mode is adjusted according to formula (2) to find the best frequency representation; the algorithm convergence during the iteration process depends on the formula: Where μ k (n) represents the state or value of the kth mode in the nth iteration, ε represents the minimum threshold set, which is used to judge the convergence condition of the algorithm; max k ||u k (n+1) -u k (n) ||<ε means that when the changes of all modes are less than the threshold, the algorithm converges; n>N maxIndicates that when the number of iterations is greater than or equal to the set maximum number of iterations, the algorithm stops iterating.

[0022] Step 5: model training; the data samples after the variational mode decomposition in step 4 are divided into a training set and a test set, and the training set is put into the improved sparrow algorithm to optimize the deep extreme learning machine module for training; in step 5, the sparrow algorithm improved by the sparrow search algorithm is used to optimize the hyper parameters of the initial pest prediction model, including the following steps, such as Figure 2 As shown: Initialize the sparrow population, including the population size, number of discoverers, number of sparrows warned, dimension of the objective function, upper and lower bounds of parameters, maximum number of iterations, and solution accuracy; Calculate the fitness value for each sparrow, and record the current optimal fitness value and its corresponding position, as well as the worst fitness value and its corresponding position; randomly select sparrows from the sparrow population as warning sparrows; among them, the explorer, which has a larger foraging range and guides the foraging direction of the population, its position is updated as follows: Where M is the maximum number of iterations; X a i，j is the location information of the i-th sparrow in the a-th generation in the population in the j-th dimension; α is a random number between (0,1]; Q is a random number that satisfies the normal distribution; L is a matrix of size i×d with all elements being 1; R 2 ∈[0, 1], ST∈[0.5, 1], are the warning value and safety value respectively; Perform one iteration, update the position information of all sparrows, and recalculate the fitness value of each sparrow. 2 ≥ST, indicating that some sparrows have discovered the threat, the discoverer value converges and moves to a safe position; R 2 <ST means that the discoverer can expand the search range, and the value of each dimension of the sparrow individual will decrease. When the sparrow is in a hungry state, its value is controlled to conform to the normal distribution; otherwise, it is taken near the optimal position, and the value is relatively stable. The position update method of the joiner is shown in the following formula: Where: x ij a+1 is the optimal position occupied by the discoverer in the a+1th iteration, is the global worst position in the a-th iteration, A is a multidimensional matrix with one row whose elements are 1 or -1, and others refer to values ​​other than i>n / 2 above. When the proportion of alerters is 10% to 20% of the total population, the position of the alerters is updated as follows: In the formula, is the current global optimal position; β is the step length parameter; K is a random number between [-1,1], which is the step length control parameter of the sparrow's moving direction; f g is the fitness value of the current individual; f g and f w are the global optimal fitness and the global worst fitness respectively; ε is a constant to avoid the denominator being 0. i >f g When , it means that the sparrow is at the edge of the population and is vulnerable to threats from natural enemies, so it needs to move to the middle; when f i =f g , which means that the sparrow in the middle of the population is threatened, and it will randomly move closer to the sparrows in other directions; According to the current state of the sparrow population, update the optimal and worst positions and fitness values ​​of the entire population; determine whether the maximum number of iterations or solution conditions have been reached, if so, output the optimal value, otherwise return to recalculate the fitness of each sparrow; Perform iterative operations until the maximum number of iterations is reached to obtain the global optimal solution and the best fitness value; The global optimal value and the best fitness value are obtained as the input of the DELM algorithm, and the DELM network is continuously trained using the training set until the total error is less than the expected error; The DELM network consists of multiple ELM-AE stacks. In the training process, at the initial stage, the training data X is used as the target output of the first ELM-AE (X1 = X), and then the output weight β of the first layer is calculated. 1 ; Then, the output matrix H1 of the hidden layer of the first layer of ELM-AE is used as the input and target output of the next ELM-AE, and so on; the last layer is trained using ELM, where H i+1 is the output matrix of the last hidden layer, T is the sample label; the input weight matrix of each hidden layer is W i+1 =β T i+1 Solve the output weight β of the last hidden layer of DELM i+1 The formula is: Where: C is the regularization parameter; I is the unit matrix with the same dimension as HTH; H is the hidden layer output matrix, which contains the output of all training samples after the activation function from the input layer to the hidden layer, and each row corresponds to the hidden layer output of a training sample; X is the target value matrix or vector of the training sample, and each row corresponds to the target value of a training sample.

[0023] Step 6: Actual warehouse inspection: use the trained model in step 5 to predict pests on the grain warehouse odor samples; input the odor information data collected in the grain warehouse into the pest prediction model to obtain the prediction results of storage pests.

[0024] Phosphine fumigation is required for pest control in warehouses every summer and autumn. The cost of pest control is about 1 yuan per ton of grain. The present invention can provide early warning and take early protection, thus saving the cost of pest control.

Claims

1. A granary pest prediction system based on odor sensor, characterized by: It includes a variational mode decomposition module, which is used to decompose the data in the granary into multiple vibration modes, and effectively capture and extract the pest behavior by extracting potential pest activity signals inside the granary; The deep extreme learning machine module is used to receive the vibration mode data output by the variational mode decomposition module and optimize it through the sparrow search algorithm to train the prediction model; The prediction model evaluation module is used to conduct a comprehensive evaluation of the trained prediction model, including verifying the accuracy, robustness and generalization ability of the prediction model, and ensuring the reliability and effectiveness of the prediction model in the actual granary environment.

2. The granary pest prediction system based on odor sensor according to claim 1 is characterized by: The variational mode decomposition module receives and processes multi-dimensional data of the granary, wherein the multi-dimensional data includes temperature, humidity, air pressure and gas composition.

3. The granary pest prediction system based on odor sensor according to claim 1 is characterized by: The deep extreme learning machine module includes multiple deep extreme learning machine networks, each of which is used to process a vibration mode output by the variational modal decomposition module; a sparrow search algorithm is used to optimize the parameters of the deep extreme learning machine network to improve the accuracy and generalization ability of the prediction model.

4. The granary pest prediction system based on odor sensor according to claim 1 is characterized by: In the prediction model evaluation module, a cross-validation method is used to evaluate the trained prediction model to verify its prediction performance on different data sets; model performance indicators are provided; and the quality of the prediction model is objectively evaluated.

5. A method for predicting pests in granaries based on odor sensors, characterized in that: The following steps are involved: Step 1: Setting up signal collection distribution points: Arrange multiple sensor nodes in the granary; To achieve full warehouse coverage to ensure that no collected signals are missed, and to achieve the purpose of economic rationality based on meeting the detection needs; Step 2: Data collection; The sensor network covers every area in the granary to collect data about the internal environment of the granary in real time; the frequency of data collection is set according to the actual situation to ensure the timeliness and accuracy of the data; Step 3: Acquisition signal processing; Preprocessing the data collected in step 2 to obtain processed data, wherein the preprocessing of the collected data includes removing outliers and performing noise reduction processing; Step 4: Calculation formula of processed signal; for the data preprocessed in step 3, a variational mode decomposition algorithm is used to decompose multiple modes of the extracted data to separate the source signal samples into multiple characteristic signal combinations; Step 5: Model training; divide the data samples after the variational mode decomposition in step 4 into a training set and a test set, and put the training set into the improved sparrow algorithm to optimize the deep extreme learning machine module for training; Step 6: Actual warehouse inspection: use the trained deep extreme learning machine module in step 5 to predict pests on the grain warehouse odor samples; input the data collected in the grain warehouse into the pest prediction model to obtain the prediction results of storage pests.

6. The method for predicting pests in granaries based on odor sensors according to claim 5 is characterized by: In step 4, the iteration formula for each mode is Among them, u k (n) represents the kth mode in the nth iteration, δ(t) is the Dirac function, λ represents the penalty term coefficient, which is used to control the constraint strength of frequency separation, ω k represents the center frequency of the kth module; Through the above iterative update, the algorithm can adaptively adjust the center frequency and bandwidth of each mode, so as to accurately extract each modal component in the signal; the following formula describes the frequency adjustment process: In the formula, represents the center frequency of the nth iteration of module k, μ represents the step size parameter, represents the state or value of the kth mode in the nth iteration; Each mode of the signal is updated according to the iterative formula during the iteration process to ensure the frequency separation between the modes and maintain the bandwidth smoothness of each mode. The center frequency of the mode is adjusted according to formula (2) to find the best frequency representation; the algorithm convergence during the iteration process depends on the formula: Where μ k (n) represents the state or value of the kth mode in the nth iteration, ε represents the minimum threshold set, which is used to judge the convergence condition of the algorithm; max k ||u k (n+1) -u k (n) ||<ε means that when the changes of all modes are less than the threshold, the algorithm converges; n>N max Indicates that when the number of iterations is greater than or equal to the set maximum number of iterations, the algorithm stops iterating.

7. The method for predicting pests in granaries based on odor sensors according to claim 5, characterized in that: In step 5, optimizing the hyperparameters of the deep extreme learning machine module using the improved sparrow algorithm improved by the sparrow search algorithm includes the following steps: Initialize the sparrow population, including the population size, number of discoverers, number of sparrows warned, dimension of the objective function, upper and lower bounds of parameters, maximum number of iterations, and solution accuracy; Calculate the fitness value for each sparrow, and record the current optimal fitness value and its corresponding position, as well as the worst fitness value and its corresponding position; randomly select sparrows from the sparrow population as warning sparrows; among them, the explorer, which has a larger foraging range and guides the foraging direction of the population, its position is updated as follows: Where M is the maximum number of iterations; X a i，j is the location information of the i-th sparrow in the a-th generation in the population in the j-th dimension; α is a random number between (0,1]; Q is a random number that satisfies the normal distribution; L is a matrix of size i×d with all elements being 1; R2∈[0,1], ST∈[0.5,1] are the warning value and safety value, respectively; Perform an iteration, update the position information of all sparrows, and recalculate the fitness value of each sparrow; when R2≥ST, it means that some sparrows have found a threat, and the discoverer value converges and moves to a safe position; R2<ST means that the discoverer can expand the search range, and the value of each dimension of the sparrow individual will decrease; when the sparrow is in a hungry state, control its value to conform to the normal distribution; otherwise, take a value near the optimal position, and the value is relatively stable. The position update method of the joiner is shown in the following formula: Where: x ij a+1 is the optimal position occupied by the discoverer in the a+1th iteration, is the global worst position in the a-th iteration, A is a multidimensional matrix with one row and each element is 1 or -1; When the proportion of alerters is 10% to 20% of the total population, the position of the alerters is updated as follows: In the formula, is the current global optimal position; β is the step length parameter; K is a random number between [-1,1], which is the step length control parameter of the sparrow's moving direction; f g is the fitness value of the current individual; f g and f w They are the global optimal fitness and the global worst fitness respectively; ε is a constant, which prevents the denominator from being zero. i >f g When , it means that the sparrow is at the edge of the population and is vulnerable to threats from natural enemies, so it needs to move to the middle; when f i =f g , which means that the sparrow in the middle of the population is threatened, and it will randomly move closer to the sparrows in other directions; According to the current state of the sparrow population, update the optimal and worst positions and fitness values ​​of the entire population; determine whether the maximum number of iterations or solution conditions have been reached, if so, output the optimal value, otherwise return to recalculate the fitness of each sparrow; Perform iterative operations until the maximum number of iterations is reached to obtain the global optimal solution and the best fitness value; The global optimal value and the best fitness value are obtained as the input of the DELM algorithm, and the DELM network is continuously trained using the training set until the total error is less than the expected error; The DELM network consists of multiple ELM-AE stacks. In the training process, at the initial stage, the training data X is used as the target output X1=X of the first ELM-AE, and then the output weight β1 of the first layer is calculated. Then, the output matrix H1 of the hidden layer of the first ELM-AE is used as the input and target output of the next ELM-AE, and so on. The last layer is trained using ELM, where H i+1 is the output matrix of the last hidden layer, T is the sample label; the input weight matrix of each hidden layer is W i+1 =β T i+1 ; Solve the output weight β of the last hidden layer of DELM i+1 The formula is: Where: C is the regularization parameter; I is the unit matrix with the same dimension as HTH; H is the hidden layer output matrix, which contains the output of all training samples after the activation function from the input layer to the hidden layer, and each row corresponds to the hidden layer output of a training sample; X is the target value matrix or vector of the training sample, and each row corresponds to the target value of a training sample.

Images