Agricultural pest ecological regulation AI auxiliary decision-making system and application thereof

Abstract

The invention relates to the technical field of pest prevention and control, and discloses an agricultural pest ecological regulation AI auxiliary decision-making system and application thereof.The agricultural pest ecological regulation AI auxiliary decision-making system comprises a farmland area division module used for dividing farmland areas; the growth data collection module is used for collecting growth data of the brassicaceous vegetables; the pest prediction module is used for predicting the pest density within seven days and evaluating whether pests burst or not within seven days; the comprehensive prevention and control scheme generation module is used for providing a comprehensive prevention and control scheme integrating mixed pesticide application, physical prevention and control measures and biological prevention and control measures; by predicting whether the pests of the cruciferous vegetables burst or not and fully considering the pest killing rate, the pesticide residue and the required cost value of each prevention and control measure, compared with single chemical prevention and control, the system can balance multiple prevention and control targets and improve the prevention and control effect.

Description

Technical Field

[0001] The present invention relates to the technical field of pest control, and more specifically, to an AI-assisted decision-making system for ecological regulation of agricultural pests and its application. Background Art

[0002] Agricultural pest control of cruciferous vegetables has always been an important part of ensuring crop yield and quality. Although chemical pesticides can effectively reduce pest density in a short period of time, long-term use will bring about problems such as pesticide residues, environmental pollution, and increased pest resistance. It is necessary to use a comprehensive ecological control method to control pests, including biological control, physical control, and chemical control, so as to effectively control the pest population and achieve long-term sustainable control effects. However, the implementation of ecological control requires comprehensive consideration of the effectiveness of pest control, the residual status of drugs, and the amount of money required. Its complexity makes the decision-making process difficult. Summary of the invention

[0003] The purpose of the present invention is to provide an AI-assisted decision-making method, system and storage medium for agricultural pest ecological regulation in order to solve the above-mentioned problems.

[0004] The present invention provides an AI-assisted decision-making system for ecological regulation of agricultural pests, comprising:

[0005] The farmland area division module is used to divide the farmland area, including:

[0006] Determine farmland boundaries and draw farmland boundary maps;

[0007] Evenly distribute measurement points in the farmland and measure the sunshine hours at each point;

[0008] According to the sunshine hours, the area with sunshine hours difference within one hour is preliminarily divided into several blocks with regular shapes;

[0009] Adjust the initially divided blocks to ensure the integrity and continuity of each block;

[0010] Number each block, the Nth block is numbered N;

[0011] The growth data collection module is used to collect the growth data of cruciferous vegetables, including:

[0012] Starting from the first day of the planting period, data are collected once a day in all blocks. The growth parameters of cruciferous vegetables collected in each block each time include: the growth period of cruciferous vegetables, the average air temperature of the day, the average air humidity of the day, the average soil humidity of the day, the average light intensity of the day, the average rainfall of the day, the average density of various pests on the day, and the average density of natural enemies of various pests on the day. The growth parameters of cruciferous vegetables collected in each block each time are encoded to obtain a growth parameter vector, and the growth parameter vectors of all blocks on the same day are spliced ​​to obtain a growth data vector;

[0013] The pest prediction module is used to predict the pest density within seven days and evaluate whether there will be a pest outbreak within seven days. The data vector set is input into the prediction neural network, and the prediction neural network outputs the result indicating whether there will be a pest outbreak in the farmland within the next seven days. If the result is yes, it is determined whether the time interval between the current date and the previous pest control date is greater than the set time threshold. If the judgment result is yes, the integrated control plan generation module is started;

[0014] The data vector set X is represented as:

[0015] X={x t-29, x t-28, …x t}

[0016] Among them, x t-29 、x t-28 、x t Represent the growth data vectors of the t-29th, t-28th and tth days respectively; the tth day is defined as the current day;

[0017] The integrated control program generation module is used to provide an integrated control program that integrates mixed pesticide application, physical and biological control measures, including:

[0018] The data vector set is input into the trained decision neural network, and the decision neural network outputs the action vector;

[0019] Decode the action vectors and output an integrated control plan to implement pest control on cruciferous vegetables;

[0020] An action vector is represented as:

[0021]

[0022] in, They represent the dosage of the first and h types of chemical control drugs in blocks 1 and N, respectively;

[0023] Respectively, they represent the usage of the first and nth physical control methods in blocks 1 and N;

[0024] They represent the usage of the 1st and mth biological control methods in blocks 1 and N, respectively.

[0025] Furthermore, if the number of days between the tth day and the date of starting planting is less than 30, an interpolation operation is performed on the missing growth data vector, and the interpolated growth data vector is a 0 vector.

[0026] Furthermore, the default value of the time threshold is 10 days.

[0027] Furthermore, the decision neural network uses a reinforcement learning model based on a deep Q network. The deep Q network includes an input layer, a hidden layer, and an output layer. The hidden layer uses a neural network to fit the Q value function, where the neural network includes a policy network and a target network.

[0028] Furthermore, the Q-value function Q(s,a) represents the reward that can be obtained after selecting action a in state s. The reward value is calculated by the reward function, where state s is the growth data vector x at the current moment t , action a is the action vector.

[0029] Furthermore, a reward function is set. The reward function needs to comprehensively consider the insecticide effect, drug residue and cost. The insecticide effect is the reduction ratio of pest density, the drug residue is the residual amount after the drug is used, and the cost is the total cost of implementing the prevention and control measures. It includes the immediate reward function and the final reward function. The immediate reward function can be expressed as:

[0030] R t =-αP t -βB t

[0031] Among them, R t represents the immediate reward at time t, P t represents the amount of pesticide used at time t, B t represents the sum of the costs of the biological and physical control measures selected at time t, α and β are weight parameters, and their default values ​​are 0.01 and 0.005 respectively;

[0032] The final reward needs to be calculated after the prevention and control measures have been implemented for a period of time and the prevention and control effects are evaluated, which can be expressed as:

[0033] R T =δE T -εD T -ζC T

[0034] Among them, δ, ε, ζ are weight parameters, and their default values ​​are 100, 200, and 0.002, respectively. T represents the period from the implementation of control measures to the evaluation of control effects. R Tis the final reward of period T, E T , D T , C T They represent the percentage of pest density reduction, drug residue and total cost in period T respectively.

[0035] Furthermore, the specific steps for updating the strategy network and the target network are:

[0036] Step 101, define the target Q value y of the Q value, y can be expressed as:

[0037]

[0038] Among them, r is the reward value obtained after executing action a in the current state s, γ is the discount factor, the default value is 0.9, Q target is the Q value of the target network, s' is the next state transferred to after executing action a, and a' is the action selected in state s';

[0039] Step 102, use the policy network to calculate the current Q value:

[0040] Q current =Q policy (s,a)

[0041] Among them, Q current is the Q value of the policy network, s is the current state, a is the current action, Q policy represents the policy network;

[0042] Step 103, using the mean square error as the loss function, calculate the error between the current Q value and the target Q value:

[0043]

[0044] Where N is the number of mini-batch samples, the default value is 64, i is the target Q value of the i-th sample, Q current,i is the current Q value of the i-th sample;

[0045] Step 104, minimizing the loss function through the back propagation algorithm and the Adam optimizer, and updating the parameters of the policy network;

[0046] Step 105, define the target network update interval H, the default value of H is 100, and the target network parameters are updated every H training steps; updating the target network parameters is to copy the parameters of the current policy network to the target network.

[0047] Furthermore, the ε-greedy strategy is used to select actions, including:

[0048] Step 201, ε represents the probability of selecting a random action, and its initial value is 0.9. 1-ε represents the probability of selecting the optimal action. During the training process, ε will decay linearly, specifically:

[0049]

[0050] Among them: 0 is 0.9, min is 0.1, T is the total number of training steps, the default value is 1000, tg is the value at the tg step;

[0051] Step 202, generating a random number r between 0 and 1;

[0052] Step 203, if r< tg , then choose a random action if r ≥ tg Then select the action with the highest current Q value.

[0053] The present invention provides an AI-assisted decision-making method for ecological regulation of agricultural pests. The following steps are performed based on the aforementioned AI-assisted decision-making system for ecological regulation of agricultural pests:

[0054] Step 301, starting from the first day of the planting period, collect growth data once a day in different blocks.

[0055] Step 302, encoding the daily growth data of different regions into an overall growth data vector;

[0056] Step 303, combining the growth data vectors of a total of 30 days into a growth data vector set, inputting the set into the prediction neural network, and outputting the result indicating whether an insect pest will break out in the farmland within the next seven days. If the result is yes, determining whether the time interval between the current date and the previous pest control date is greater than a set time threshold, if so, executing step 304, if less than, executing step 301; if no, executing step 301;

[0057] Step 304, inputting the data vector set into the trained decision neural network, and the decision neural network outputs the action vector;

[0058] Step 305 , decode the action vector, output a comprehensive pest control plan for implementation, and perform pest control on cruciferous vegetables, and then return to step 301 .

[0059] The present invention provides an application device that runs the aforementioned agricultural pest ecological regulation AI-assisted decision-making system through a processor.

[0060] The beneficial effect of the present invention is that the present invention predicts whether there will be an outbreak of pests in cruciferous vegetables, fully considers the insecticide rate, drug residue, and required cost of each prevention and control measure, and makes a comprehensive prevention and control plan including chemical, physical, and biological control. Compared with single chemical prevention and control, the system can balance multiple prevention and control objectives and improve the prevention and control effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a system module diagram of the present invention;

[0062] Figure 2 It is a flow chart of an AI-assisted decision-making method for ecological regulation of agricultural pests of the present invention.

[0063] In the figure: farmland area division module 101, growth data collection module 102, pest prediction module 103, integrated prevention and control plan generation module 104. DETAILED DESCRIPTION

[0064] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0065] In an agricultural pest ecological regulation AI-assisted decision-making system of the present invention, Figure 1 As shown, including:

[0066] The farmland area division module is used to divide the farmland area, including:

[0067] Determine farmland boundaries and draw farmland boundary maps;

[0068] Evenly distribute measurement points in the farmland and measure the sunshine hours at each point;

[0069] According to the sunshine hours, the area with sunshine hours difference within one hour is preliminarily divided into several blocks with regular shapes (such as rectangles and squares);

[0070] Adjust the initially divided blocks to ensure that there are no breaks within each block, the boundaries are clear and there is no overlap, and the blocks are adjacent to each other;

[0071] Number each block, the Nth block is numbered N;

[0072] The growth data collection module is used to collect the growth data of cruciferous vegetables, including:

[0073] Starting from the first day of the planting period, data are collected once a day in all blocks. The growth parameters of cruciferous vegetables collected in each block each time include: the growth period of cruciferous vegetables, the average air temperature of the day, the average air humidity of the day, the average soil humidity of the day, the average light intensity of the day, the average rainfall of the day, the average density of various pests on the day, and the average density of natural enemies of various pests on the day. The growth parameters of cruciferous vegetables collected in each block each time are encoded to obtain a growth parameter vector, and the growth parameter vectors of all blocks on the same day are spliced ​​to obtain a growth data vector;

[0074] In one embodiment of the present invention, the growth parameter vectors of all blocks on the same day are concatenated in the order of the block IDs.

[0075] Data management module, which is used to store and process various types of data;

[0076] The pest prediction module is used to predict the pest density within seven days and evaluate whether the pests will break out within seven days. The data vector set is input into the prediction neural network, and the prediction neural network outputs the result indicating whether the pests will break out in the farmland within the next seven days. If the result is yes, it is determined whether the time interval between the current date and the previous pest control date is greater than the set time threshold. The default value of the time threshold is 10 days. The data vector set X is expressed as:

[0077] X={x t-29 ,x t-28 ,…x t}

[0078] Among them, x t-29 ,,x t-28 、x t Represent the growth data vectors of the t-29th, t-28th and tth days respectively; the tth day is defined as the current day; if the number of days between the tth day and the date of planting is less than 30, the missing growth data vectors are interpolated, and the interpolated growth data vectors are 0 vectors;

[0079] The prediction neural network consists of a hidden layer and an output layer, and the hidden layer consists of multiple layers of fully connected neurons.

[0080] Assume there are L hidden layers, each containing n l neurons, l = 1, 2, ..., L, then the calculation formula of the lth hidden layer is:

[0081] h l =f(W l h l-1 +b l )

[0082] Among them, h lrepresents the output of the lth layer, L ≥ l ≥ 1, h 0 =X,W l is the weight matrix of the lth layer, b l is the bias vector of the lth layer, and f is the ReLU activation function;

[0083] The output layer outputs the last layer of the hidden layer h L Converted into prediction results, the formula for calculating the probability of pest outbreak in the output layer is as follows:

[0084]

[0085] in, It represents the probability of pest outbreak in farmland on the dth day in the future. Its value is between 0 and 1, d = 1, 2, ..., 7. is the d-th weight matrix of the output layer, h L represents the output of the Lth hidden layer, is the dth bias vector of the output layer, σ is the Sigmoid activation function;

[0086] The probability value Convert to binary output i,d :

[0087]

[0088] Among them, d =1 indicates that an insect pest outbreak will occur on the dth day in the future, o d =0 means that there will be no insect pest outbreak on the future day d.

[0089] The integrated control program generation module is used to provide an integrated control program that integrates mixed pesticide application, physical and biological control measures, including:

[0090] The data vector set is input into the trained decision neural network, and the decision neural network outputs the action vector, which is expressed as:

[0091]

[0092] in, They represent the dosage of the first and h types of chemical control drugs in blocks 1 and N, respectively;

[0093] Respectively, they represent the usage of the first and nth physical control methods in blocks 1 and N; They represent the usage of the 1st and mth biological control methods in blocks 1 and N, respectively;

[0094] For example, farmland A is divided into three blocks. Whitefly outbreaks occurred in blocks 2 and 3 at the seedling stage. The order of chemical control drugs in the action vector is thiamethoxam, chlorfenapyr·imidacloprid, pyrethrin, and cypermethrin. The chemical control action vector part output by the decision neural network is {0, 1.6, 1.5, 0, 1.2, 1.2, 0, 0, 0, 0, 0}, indicating that thiamethoxam and chlorfenapyr·imidacloprid are used in block 2 at a dosage of 1.6 and 1.2 grams per mu, respectively, and thiamethoxam and chlorfenapyr·imidacloprid are used in block 3 at a dosage of 1.5 and 1.2 grams per mu, respectively. The order of physical control methods is covering with insect-proof nets, using yellow sticky insect boards, and using shade nets. The physical control action vector part output by the decision neural network is {0, 0 , 0, 0, 20, 20, 0, 0, 0}, indicating that yellow sticky insect boards are used in blocks 2 and 3 at a dosage of 20 per mu; the order of biological control methods is the use of nuclear polyhedrosis virus, the release of parasitic wasps, and the release of Spodoptera alfalfa. The biological control action vector part output by the decision neural network is {0, 0, 0, 0, 300, 300, 0, 0}, indicating that parasitic wasps are released in blocks 2 and 3 at a dosage of 300 per mu; based on the above three parts, the action vector output by the decision neural network is {0, 1.6, 1.5, 0, 1.2, 1.2, 0, 0, 0, 0, 0, 0, 0, 0, 20, 20, 0, 0, 0, 0, 0, 0, 300, 300, 0, 0, 0}.

[0095] Decode the action vectors and output an integrated control plan to implement pest control on cruciferous vegetables.

[0096] In one embodiment of the present invention, the calculation formula of the decision neural network is as follows:

[0097] H (0) =x (1)

[0098] u (t) =σ(W u X (t) +W u H (t-1) +b u )

[0099] r (t) =σ(W r X (t) +W r H (t-1) +b r )

[0100] C (t) =tanh(W c X (t) +W c r (t) ⊙H(t-1) +b c )

[0101] H (t) =(1-u (t) )⊙C (t) +u (t) ⊙H (t-1)

[0102] OUT=σ(W o *H (t) +b o )

[0103] Among them, x (1) and X (t) Respectively represent the first and tth growth data vectors of the data vector set X, W u , W r , W c , W o represents the trainable weight parameter, b u 、b r 、b c 、b o represents the trainable bias parameter, ⊙ represents the dot product, u (t) represents the tth first intermediate feature, r (t) represents the t-th second intermediate feature, C (t) represents the tth third intermediate feature, H (t-1) and H (t) They represent the t-1th and tth output features respectively, OUT represents the output vector, the value of the i-th component of the output vector represents the Q value of the i-th action, tanh represents the tanh function, and σ represents the Sigmoid function.

[0104] The action output by the decision neural network is the action with the largest Q value.

[0105] The decision neural network uses a reinforcement learning model based on a deep Q network. The deep Q network includes an input layer, a hidden layer, and an output layer. The hidden layer uses a neural network to fit the Q value function. The neural network includes a policy network and a target network.

[0106] The Q-value function Q(s,a) represents the reward that can be obtained after selecting action a in state s. The reward value is calculated by the reward function, where state s is the growth data vector x at the current moment. t , action a is the action vector.

[0107] Set up a reward function. The reward function needs to comprehensively consider the insecticide effect, drug residue and cost. The insecticide effect is the reduction ratio of pest density, the drug residue is the residual amount of the drug after use (mg / kg), and the cost is the total cost of implementing prevention and control measures (yuan). It includes the immediate reward function and the final reward function. The immediate reward function can be expressed as:

[0108] R t =-αP t -βB t

[0109] Among them, R t represents the immediate reward at time t, P t represents the amount of pesticide used at time t (g / mu), B t represents the sum of the costs of the biological and physical control measures selected at time t, α and β are weight parameters, and their default values ​​are 0.01 and 0.005 respectively;

[0110] The final reward needs to be calculated after the prevention and control measures have been implemented for a period of time and the prevention and control effects are evaluated, which can be expressed as:

[0111] R T =δE T -εD T -ζC T

[0112] Among them, δ, ε, ζ are weight parameters, and their default values ​​are 100, 200, and 0.002, respectively. T represents the period from the implementation of control measures to the evaluation of control effects. R T is the final reward of period T, E T , D T , C T They represent the percentage of pest density reduction, drug residue and total cost in period T respectively.

[0113] The specific steps to update the policy network and target network are:

[0114] Step 101, define the target Q value y of the Q value, y can be expressed as:

[0115]

[0116] Among them, r is the reward value obtained after executing action a in the current state s, γ is the discount factor, the default value is 0.9, and q target is the Q value of the target network, s′ is the next state transferred to after executing action a, and a′ is the action selected in state s′;

[0117] Step 102, use the policy network to calculate the current Q value:

[0118] Q current=Q policy (s,a)

[0119] Among them, Q current is the Q value of the policy network, s is the current state, a is the current action, Q policy represents the policy network;

[0120] Step 103, using mean square error (MSE) as the loss function, calculate the error between the current Q value and the target Q value:

[0121]

[0122] Where N is the number of mini-batch samples, the default value is 64, i is the target Q value of the i-th sample, Q current,i is the current Q value of the i-th sample;

[0123] Step 104, minimize the loss function through the back propagation algorithm and Adam optimizer, and update the parameters of the policy network, specifically,

[0124]

[0125] Among them, θ policy are the parameters of the policy network before updating, represents the parameters of the updated policy network, η is the learning rate, the default value is 0.01, Loss is the gradient of the loss function with respect to the policy network parameters;

[0126] Step 105, define the target network update interval H, the default value of H is 100, and the target network parameters are updated every H training steps; updating the target network parameters is to copy the parameters of the current policy network to the target network.

[0127] When selecting actions, the ε-greedy strategy is used to select actions, including:

[0128] Step 201, ε represents the probability of selecting a random action, and its initial value is 0.9. 1-ε represents the probability of selecting the optimal action. During the training process, ε will decay linearly, specifically:

[0129]

[0130] Among them: ∈0 is 0.9, ∈ min is 0.1, T is the total number of training steps, the default value is 1000, ∈ tg is the value of ∈ at step tg;

[0131] Step 202, generating a random number r between 0 and 1;

[0132] Step 203, if r<∈ tg , then choose a random action if r ≥ ∈ tg Then select the action with the highest current Q value.

[0133] In the actual planting process, since the environmental conditions (such as the duration of light) in different blocks of farmland are different, these conditions will affect the distribution and density of pests. At this time, when the integrated pest control plan is implemented based on the average density of pests in the entire farmland, the best resource utilization effect and pest control effect cannot be achieved. For example, the natural enemies of some pests are small in size and have limited mobility. Uniform release will result in a large capital investment and cannot achieve the best control effect. In the aforementioned embodiment, the farmland is divided into different blocks according to different environmental conditions, and the best integrated control plan is generated by block, which can maximize the effect of biological control.

[0134] Regular pest control according to the growth cycle can only be applied to planting cycles without sudden situations. In the aforementioned embodiment of the present invention, a method of predicting pest outbreaks is adopted to generate a prevention and control strategy for pest control before the pest outbreak, which can be applied to planting cycles with sudden pest situations.

[0135] In one embodiment of the present invention, an AI-assisted decision-making method for ecological regulation of agricultural pests is provided, such as Figure 2 As shown, the following steps are included:

[0136] Step 301, starting from the first day of the planting period, collect growth data once a day in different blocks.

[0137] Step 302, encoding the daily growth data of different regions into an overall growth data vector;

[0138] Step 303, combining the growth data vectors of a total of 30 days into a growth data vector set, inputting the set into the prediction neural network, and outputting the result indicating whether an insect pest will break out in the farmland within the next seven days. If the result is yes, determining whether the time interval between the current date and the previous pest control date is greater than a set time threshold, if so, executing step 304, if less than, executing step 301; if no, executing step 301;

[0139] Step 304, inputting the data vector set into the trained decision neural network, and the decision neural network outputs the action vector;

[0140] Step 305 , decode the action vector, output a comprehensive pest control plan for implementation, and perform pest control on cruciferous vegetables, and then return to step 301 .

[0141] At least one embodiment of the present invention provides a storage medium storing non-temporary computer-readable instructions. When the non-temporary computer-readable instructions are executed by a computer, the steps of the aforementioned AI-assisted decision-making method for ecological regulation of agricultural pests can be executed.

[0142] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. An AI-assisted decision-making system for ecological regulation of agricultural pests, characterized in that: include: The farmland area division module is used to divide the farmland area, including: Determine farmland boundaries and draw farmland boundary maps; Evenly distribute measurement points in the farmland and measure the sunshine hours at each point; According to the sunshine hours, the area with sunshine hours difference within one hour is preliminarily divided into several blocks with regular shapes; The growth data collection module is used to collect the growth data of cruciferous vegetables, including: Starting from the first day of the planting period, data are collected once a day in all blocks. The growth parameters of cruciferous vegetables collected in each block each time include: the growth period of cruciferous vegetables, the average air temperature of the day, the average air humidity of the day, the average soil humidity of the day, the average light intensity of the day, the average rainfall of the day, the average density of various pests on the day, and the average density of natural enemies of various pests on the day. The growth parameters of cruciferous vegetables collected in each block each time are encoded to obtain a growth parameter vector, and the growth parameter vectors of all blocks on the same day are spliced ​​to obtain a growth data vector; The pest prediction module is used to predict the pest density within seven days and evaluate whether there will be a pest outbreak within seven days. The data vector set is input into the prediction neural network, and the prediction neural network outputs the result indicating whether there will be a pest outbreak in the farmland within the next seven days. If the result is yes, it is determined whether the time interval between the current date and the previous pest control date is greater than the set time threshold. If the judgment result is yes, the integrated control plan generation module is started; The data vector set X is represented as: X={x t-29 ,x t-28 ,…x t } Among them, x t-29 、x t-28 、x t Represent the growth data vectors of the t-29th, t-28th and tth days respectively; the tth day is defined as the current day; The integrated control program generation module is used to provide an integrated control program that integrates mixed pesticide application, physical and biological control measures, including: The data vector set is input into the trained decision neural network, and the decision neural network outputs the action vector; Decode the action vectors and output an integrated control plan to implement pest control on cruciferous vegetables; An action vector is represented as: in, They represent the dosage of the first and h types of chemical control drugs in blocks 1 and N, respectively; Respectively, they represent the usage of the first and nth physical control methods in blocks 1 and N; They represent the usage of the 1st and mth biological control methods in blocks 1 and N, respectively.

2. The AI-assisted decision-making system for ecological regulation of agricultural pests according to claim 1 is characterized in that: If the number of days between the tth day and the date of starting planting is less than 30, the missing growth data vector is interpolated, and the interpolated growth data vector is a 0 vector.

3. The AI-assisted decision-making system for ecological regulation of agricultural pests according to claim 1 is characterized in that: The default value of the time threshold is 10 days.

4. The AI-assisted decision-making system for ecological regulation of agricultural pests according to claim 1 is characterized in that: The decision neural network uses a reinforcement learning model based on a deep Q network. The deep Q network includes an input layer, a hidden layer, and an output layer. The hidden layer uses a neural network to fit the Q value function, where the neural network includes a policy network and a target network.

5. The AI-assisted decision-making system for ecological regulation of agricultural pests according to claim 4 is characterized in that: The Q-value function Q(s,a) represents the reward that can be obtained after selecting action a in state s. The reward value is calculated by the reward function, where state s is the growth data vector x at the current moment. t , action a is the action vector.

6. The AI-assisted decision-making system for ecological regulation of agricultural pests according to claim 5 is characterized in that: Set up a reward function. The reward function needs to comprehensively consider the insecticide effect, drug residue and cost. The insecticide effect is the reduction ratio of pest density, the drug residue is the residual amount after the drug is used, and the cost is the total cost of implementing prevention and control measures. It includes an immediate reward function and a final reward function. The immediate reward function is expressed as: R t =-αP t -βB t Among them, R t represents the immediate reward at time t, P t represents the amount of pesticide used at time t, B t represents the sum of the costs of the biological and physical control measures selected at time t, α and β are weight parameters, and their default values ​​are 0.01 and 0.005 respectively; The final reward needs to be calculated after the prevention and control measures have been implemented for a period of time and the prevention and control effects are evaluated, which can be expressed as: R T =δE T -eD T -ζC T Among them, δ, ε, ζ are weight parameters, and their default values ​​are 100, 200, and 0.002, respectively. T represents the period from the implementation of control measures to the evaluation of control effects. R T is the final reward of period T, E T , D T , C T They represent the percentage of pest density reduction, drug residue and total cost in period T respectively.

7. The AI-assisted decision-making system for ecological regulation of agricultural pests according to claim 4 is characterized in that: The specific steps to update the policy network and target network are: Step 101, define the target Q value y of the Q value, y is expressed as: Among them, r is the reward value obtained after executing action a in the current state s, γ is the discount factor, the default value is 0.9, Q target is the Q value of the target network, s ′ is the next state after executing action a, a ′ For state s ′ Next, select the action; Step 102, use the policy network to calculate the current Q value: Q current =Q p olic y (s,a) Among them, Q current is the Q value of the policy network, s is the current state, a is the current action, Q policy represents the policy network; Step 103, using the mean square error as the loss function, calculate the error between the current Q value and the target Q value: Where N is the number of mini-batch samples, the default value is 64, i is the target Q value of the i-th sample, Q current,i is the current Q value of the i-th sample; Step 104, minimizing the loss function through the back propagation algorithm and the Adam optimizer, and updating the parameters of the policy network; Step 105, define the target network update interval H, the default value of H is 100, and the target network parameters are updated every H training steps; updating the target network parameters is to copy the parameters of the current policy network to the target network.

8. The AI-assisted decision-making system for ecological regulation of agricultural pests according to claim 5 is characterized in that: When selecting actions, the ε-greedy strategy is used to select actions, including: Step 201, ε represents the probability of selecting a random action, and its initial value is 0.

9. 1-ε represents the probability of selecting the optimal action. During the training process, ε will decay linearly, specifically: Among them: ∈0 is 0.9, ∈ min is 0.1, T is the total number of training steps, the default value is 1000, ∈ tg is the value of ∈ at step tg; Step 202, generating a random number r between 0 and 1; Step 203, if r<∈ tg , then choose a random action if r ≥ ∈ tg Then select the action with the highest current Q value.

9. An AI-assisted decision-making method for ecological regulation of agricultural pests, characterized in that: Based on the AI-assisted decision-making system for ecological regulation of agricultural pests as described in any one of claims 1 to 8, the following steps are performed: Step 301, starting from the first day of the planting period, collect growth data once a day in different blocks. Step 302, encoding the daily growth data of different regions into an overall growth data vector; Step 303, combining the growth data vectors of a total of 30 days into a growth data vector set, inputting the set into the prediction neural network, and outputting the result indicating whether an insect pest will break out in the farmland within the next seven days. If the result is yes, determining whether the time interval between the current date and the previous pest control date is greater than a set time threshold, if so, executing step 304, if less than, executing step 301; if no, executing step 301; Step 304, inputting the data vector set into the trained decision neural network, and the decision neural network outputs the action vector; Step 305 , decode the action vector, output a comprehensive pest control plan for implementation, and perform pest control on cruciferous vegetables, and then return to step 301 .

10. An application device, characterized in that: The AI-assisted decision-making system for ecological regulation of agricultural pests as described in any one of claims 1 to 8 is run by a processor.

Images