A pest situation forecasting lamp control method and system based on a prediction model

Abstract

The application discloses a kind of pest situation forecasting lamp control method and system based on prediction model, which comprises: fusion liquid state neural network and multi-head attention mechanism, construct time series prediction model.Data samples are divided, respectively trained to obtain long-term pest situation prediction model and short-term pest situation prediction model.System calls long-term prediction model, generates preliminary work plan according to macro pest situation trend, and based on this plan, pest situation forecasting lamp is started.In the preset time after the start of the forecasting lamp, the system calls the short-term prediction model at a fixed frequency, analyzes and outputs the short-term future pest situation trend label in real time.According to the short-term prediction label, dynamically adjust the end time of the current work period to form a real-time closed loop of prediction-adjustment-execution.The system includes a model building unit, a model division unit and a scheduling control unit.Using the application improves the dynamic response capability of pest situation forecasting and other changes in pest situation.The application can be widely used in the field of intelligent control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and in particular to a method and system for controlling insect monitoring lamps based on a predictive model. Background Technology

[0002] In recent years, with the advancement of agricultural modernization, agricultural production methods are transforming towards green, intelligent, and sustainable development. Against this backdrop, pest control concepts are also changing accordingly. Traditional control models centered on chemical pesticides, due to long-term use leading to increased pesticide resistance in pests, ecological degradation, and rising risks of pesticide residues in agricultural products, are no longer sufficient to meet the demands of high-quality agricultural development.

[0003] To address these challenges, a green plant protection system guided by the principle of "prevention first, integrated pest management" is gradually becoming mainstream. This system emphasizes the synergistic application of non-chemical methods such as physical control, biological control, and intelligent monitoring to build an environmentally friendly, resource-saving, precise, and efficient integrated pest management mechanism.

[0004] Among them, the widely used green and intelligent plant protection equipment—the insect monitoring lamp—integrates the dual functions of physical insect trapping and pest monitoring. Its basic working principle is as follows: it uses a light source of a specific wavelength (such as 365nm ultraviolet light) to attract phototactic agricultural pests, which are then captured by electric shock from a high-voltage grid or by suction from a negative pressure fan. Simultaneously, the equipment has a built-in high-definition image acquisition module, an automatic counting unit, and multi-parameter environmental sensors (such as temperature, humidity, light intensity, and rainfall). It can use image recognition technology to identify the species and count the number of trapped insects. Some models also support remote data upload and platform-based visual management.

[0005] Currently, insect monitoring lamps generally employ relatively simple operating logic, with typical modes including "light-controlled start / stop" (automatically turning on when ambient illuminance falls below a threshold) and "timed start / stop" (e.g., operating from 7:00 PM to 6:00 AM the following day). Simultaneously, the trapping light source typically operates continuously at a fixed wavelength (e.g., 365nm ultraviolet light), lacking specificity for targeting pests. This static operating mode is insufficiently adaptable to the dynamic changes in insect populations within the complex and ever-changing agricultural production environment. Agricultural pest occurrence exhibits significant seasonality, diurnal rhythms, and microclimate dependence. For example, the peak activity of the rice leaf roller in early rice season is concentrated between 8:00 PM and 10:00 PM, while in late rice season it extends to 10:00 PM to midnight; the fall armyworm is more active at night under high temperature and humidity conditions. Existing fixed-time operation modes cannot detect these dynamic changes, often resulting in the monitoring lamps running idle during pest low periods or missing pests due to premature shutdown during peak periods. This not only results in wasted energy consumption, further exacerbating the battery life problem, but also weakens the timeliness and accuracy of pest control. Summary of the Invention

[0006] In view of this, in order to solve the technical problem that most existing insect monitoring lamp control methods are based on simple light-controlled start / stop and timed start / stop logic, resulting in insufficient adaptability to dynamic changes in insect populations, this invention proposes an insect monitoring lamp control method based on a prediction model, which includes the following steps: Model Construction and Training: First, a time series prediction model is constructed by integrating a liquid neural network and a multi-head attention mechanism. Training sample sets with different time spans are built, and the time series prediction models are trained separately to obtain long-term and short-term insect infestation prediction models. Long-Term Planning and Activation: The system calls the long-term prediction model, generates a preliminary work plan based on macro-level insect infestation trends, and automatically activates the insect monitoring lights based on this plan. Short-Term Dynamic Feedback: After the monitoring lights are activated for a preset time, the system calls the short-term prediction model at a fixed frequency, analyzes in real time, and outputs "short-term future insect infestation trend labels." Real-Time Closed-Loop Control: Based on the short-term prediction labels, the system dynamically adjusts the end time of the current work period, forming a real-time closed loop of "prediction-adjustment-execution" until the end of the day's work. Model Iterative Optimization: The system periodically performs incremental learning on the long-term and short-term models, fine-tuning model parameters using new data. When the prediction accuracy of the fine-tuned model exceeds a preset threshold, the original model is automatically replaced, achieving continuous system evolution.

[0007] In addition to the above method, the present invention also proposes a pest monitoring lamp control system based on a prediction model, including a model building unit, a model partitioning unit, and a scheduling control unit.

[0008] Based on the above scheme, this invention provides a method and system for controlling insect pest monitoring lamps based on a predictive model. By integrating a liquid neural network and a multi-head attention mechanism into an insect pest prediction model and intelligent scheduling control, it collaboratively addresses two core challenges: energy efficiency management and the spatiotemporal dynamic adaptability to insect pests. Based on a structured operation plan generated from long-term insect pest prediction, the system precisely triggers equipment start-up and shutdown before the target active period, effectively avoiding ineffective operation. Simultaneously, relying on high-frequency short-term insect pest prediction, the operating time is dynamically adjusted (extended when pest activity increases, and shut down earlier when activity decreases), ensuring that the monitoring lamps respond in real-time to changes in pest diurnal rhythms and microclimate dependence. This mechanism improves equipment endurance stability and energy utilization efficiency while also enhancing the dynamic response capability to changes in insect activity. Attached Figure Description

[0009] Figure 1 This is a flowchart of the steps of a method for controlling insect monitoring lamps based on a prediction model according to the present invention. Figure 2 This is a schematic diagram of the LNN-Attention architecture in a specific embodiment of the present invention; Figure 3This is a schematic diagram of the scheduling and control process in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the working timeline of a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the hardware platform of the system of the present invention. Detailed Implementation

[0010] In addition to the insufficient adaptability to dynamic changes in insect populations mentioned in the background technology, existing methods also suffer from the following problems: weak energy efficiency management and insufficient endurance stability. Most insect monitoring lamps are deployed in hilly and mountainous areas, orchards, or scattered small-scale farmland where power grid coverage is difficult, mainly relying on a solar-battery hybrid power supply system. However, unfavorable weather conditions or tree canopy shading will lead to a decrease in photovoltaic charging efficiency and insufficient battery energy storage. At this time, if the insect monitoring lamps are still forced to operate according to the preset time period, they are very likely to be interrupted due to power depletion, seriously affecting continuous monitoring and control capabilities; the trapping strategy lacks targeting and has poor ecological compatibility. Different insects have significant differences in their sensitivity to the spectrum: Polycoleoptera beetles are sensitive to 365nm ultraviolet light, while most moths prefer blue-violet light around 420nm. Current insect monitoring lamps mostly use fixed wavelength trapping light sources, which cannot dynamically adjust spectral parameters according to the active pest species, resulting in the accidental killing of non-target insects (especially beneficial natural enemies). This will disrupt the ecological balance of farmland, weaken the natural ability to control pests, and make it impossible to simultaneously achieve both "pest control" and "benefit protection".

[0011] In summary, while existing insect monitoring lamps have achieved automation at the data acquisition level, they still have significant shortcomings in intelligent scheduling in the time dimension and precise control in the spectral dimension. There is an urgent need to construct an energy-efficient dynamic operational decision-making mechanism with insect prediction and targeted trapping capabilities to improve control effectiveness, extend equipment lifespan, protect the ecological balance of farmland, and better serve the development needs of modern green agriculture.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0014] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0015] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0016] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0017] To address the problems of low energy efficiency, poor adaptability, and insufficient ecological compatibility in the traditional operation mode of insect monitoring lamps, this invention proposes an intelligent and dynamic pest control solution—a time-spectrum adaptive scheduling system and method for monitoring lamps based on LNN-Attention insect perception. This solution constructs a closed-loop dynamic operation mode of "prediction-scheduling-execution-feedback-update" by deeply integrating edge computing, lightweight neural networks, multi-source heterogeneous data driving, and online incremental learning strategies. It aims to break through the limitations of traditional static control modes, achieving better pest control results with shorter working time and energy consumption, and realizing an efficient, precise, and sustainable pest control mode.

[0018] Reference Figure 1 This is a flowchart illustrating an optional example of the insect monitoring lamp control method based on a prediction model proposed in this invention. This method can be applied to computer equipment, and the control method proposed in this embodiment may include, but is not limited to, the following steps: Step S1: Construct a time series prediction model based on liquid neural network and multi-head attention mechanism; Step S2: Construct training sample sets with different time spans and train the time series prediction model respectively to obtain the long-term insect infestation prediction model and the short-term insect infestation prediction model. Step S3: Call the long-term insect infestation prediction model to obtain a preliminary work plan; Step S4: Based on the preliminary work plan, turn on the insect monitoring lamp. After the insect monitoring lamp has been turned on for a preset time, call the short-term insect infestation prediction model at a preset frequency and output short-term future insect infestation trend labels. Step S5: Adjust the end time of the current working period based on the short-term future insect infestation trend label, and repeat the process until the end time is reached. Step S6: Incremental learning is performed on the long-term insect infestation prediction model and the short-term insect infestation prediction model, and the model parameters are adjusted.

[0019] In some feasible embodiments, step S1 specifically includes: This embodiment proposes a time series prediction model architecture—LNN-Attention—that integrates Liquid Neural Network (LNN) and Multi-head Attention to predict future insect infestation information. This architecture combines the continuous modeling capability of LNN for time-series dynamic systems with the association extraction capability of attention mechanisms for multi-source heterogeneous information, offering significant technical advantages. The architecture comprises the following three core components: LNN dynamic encoding module: LNN is a continuous-time neural network inspired by biological neural systems, which describes the hidden state using continuous-time differential equations. Dynamic evolution: in, For a moment The input feature vector, For learnable parameters The defined nonlinear function. It can be seen that LNN incorporates the input into the rate of change of the state (i.e., the hidden state). In the derivative of the equation, the continuous state trajectory is obtained by solving the differential equation (the present invention uses the ODE solver approximation), so that the model treats the input as a continuous time signal, and finally models non-stationary, asynchronous, real-time dynamic information more naturally.

[0020] At discrete time points At this point, the continuous-time differential equation is discretized and solved using numerical integration methods (such as the Euler approximation) to obtain the state update: This structure can efficiently capture nonlinear and non-stationary time-series insect infestation features with extremely low parameter count. Compared to traditional static recurrent neural networks (such as LSTM and GRU), LNN has stronger generalization ability and lower computational overhead, making it particularly suitable for deployment in resource-constrained edge devices to meet the requirements of real-time response and low-power operation.

[0021] Multi-head attention context enhancement module: While LNNs excel at capturing local dynamics, they still face information integration bottlenecks when handling long sequences or multivariate coupling relationships. To address this, this invention embeds a multi-head attention mechanism after the LNN output layer to explicitly model global dependencies between different time steps and multiple source variables. This mechanism can effectively identify potential associations between key environmental factors (such as sudden temperature changes and rapid humidity increases) and pest activity, thereby improving the accuracy of modeling the periodic patterns of pest infestations.

[0022] Specifically, LNN uses continuous-time differential equations to process the input time series. Dynamic encoding is performed, and the hidden state sequence is obtained by numerically solving ordinary differential equations (ODEs). Each of them It is up to the deadline A compressed representation of the dynamic coupling of insect infestation and environmental data.

[0023] Stack the state sequence into a matrix form: Subsequently, in this way As input to the multi-head self-attention mechanism, a query (Q), key (K), and value (V) matrix is ​​generated through linear projection: in For learnable weight parameters, is the subspace dimension, used to map the LNN output to the attention computation space.

[0024] To capture multi-scale temporal dependencies, the following approach is adopted: Several parallel attention heads, each using independent projection weights. Perform differential transformation on the input: Ultimately The output of the head is spliced ​​together and projected through the output matrix. Enhanced representation: Fusion prediction module: To simultaneously utilize local dynamic evolution and global contextual dependence, the nearest neighbor from the LNN is selected. Hidden state at each time step and its corresponding attention enhancement representation (in For matrix The (Lines), spliced ​​together to construct fused features: This fusion feature simultaneously includes: local dynamic information, derived from the continuous differential evolution trajectory of the input signal within the recent time window by the LNN; and global contextual information, derived from the multi-head attention mechanism's processing of the entire historical sequence. The projection of the weighted integration result at recent points in time.

[0025] at last, Input a fully connected prediction head to generate outputs for downstream multi-step prediction tasks (e.g., future predictions). (Insect population density sequence or activity probability distribution at each time step).

[0026] The overall architecture of this model is referenced Figure 2 .

[0027] In some feasible embodiments, the long-term worm prediction model specifically includes: This model aims to achieve macroscopic prediction of daily pest infestation status, providing a priori decision-making basis for the daily operation plan of monitoring lamps. Its core task is to predict the active time periods of each major pest and their corresponding optimal trapping spectral parameters based on historical environmental-pestation time series data.

[0028] The input window is the time-series data (time step 1 hour) collected from the past 7 work sessions, including: time, environmental parameters (average temperature, average humidity, etc.); insect species image recognition results; number of electric shocks from the power grid.

[0029] If the monitoring light only works for 1 hour on the first day and 3 hours on the second day, the generated time-series data is shown in the table below: Table 1. Time series data of the long-term insect infestation prediction model Output format: The model uses hourly granularity to make multi-step predictions for key monitoring periods (18:00 of the current day to 16:00 of the next day), and calculates and outputs the results for each time step. Pest activity probability and match the recommended wavelength .

[0030] Pest activity probability Calculated using a multi-label classification head (fully connected layer + Sigmoid activation function): in It is a collection of pest categories (such as brown planthoppers, aphids, etc.). It is the Sigmoid activation function. For LNN-Attention fusion features, These are learnable parameters.

[0031] Subsequently retained (like The consecutive time points are combined to target the dominant insect species. Working time window Simultaneously, the optimal wavelength is matched by searching existing databases. .

[0032] Output example: This result directly drives the scheduling and control system to generate the daily work plan.

[0033] In some feasible embodiments, the short-term insect infestation prediction model specifically includes: This model focuses on real-time prediction of short-term, high-frequency insect infestation fluctuations, used to dynamically adjust the start / stop status and operating duration of monitoring lights to improve response sensitivity. Its core task is to predict insect infestation trends in the near future based on short-term environmental-insect infestation time-series data.

[0034] Input variables: The input window uses time-series data from the past hour (5-minute time step), including: time, environmental parameters (average temperature, average humidity, etc.); insect species image recognition results; number of electric shocks from the power grid; and some preprocessed features. See the table below: Table 2 Time series data of the short-term insect infestation prediction model Output format: This model uses a 5-minute time step to predict the sequence of electric shock events in the power grid within the next hour: , in Indicates the first The predicted number of shocks is calculated within a 5-minute interval. Summing the sequences yields an estimate of the total number of shocks expected in the next hour. , Used to characterize the intensity of short-term insect activity.

[0035] Let the reference value be the total number of actual electric shocks in the most recent full hour. The relative rate of change is defined as: Adding 1 to the denominator is used to avoid division by zero and to maintain numerical stability under low insect population base.

[0036] To scientifically determine the short-term trend of insect infestation, this plan adopts the quantile method, combining historical insect infestation data statistical distribution with actual business needs, and sets three types of relative change rate thresholds.

[0037] First, based on the effective hourly electric shock event sequence recorded by the target area monitoring equipment within a preset historical period (preferably 1–2 years). and its sum Calculate each valid time point Relative rate of change: All Sort by numerical value, construct an empirical distribution, and extract the following quantiles: 85th percentile As an upward threshold ; 15th percentile Take its absolute value as the descent threshold. .

[0038] Finally, the insect infestation trend label is determined according to the following rules: In some feasible embodiments, steps S3-S5 specifically include: Based on the outputs of long-term and short-term pest infestation prediction models, the system dynamically starts and stops the pest monitoring lamps and configures their spectrum. Its core objective is to achieve maximum pest control effectiveness with minimal operating time and energy consumption. Specifically, the module forms a closed-loop control system through a two-layer decision-making mechanism: Macro-planning layer: Based on the structured mapping table (such as "peach borer, 19:00, 22:00, 430nm") output by the long-term pest situation prediction model, the module analyzes the core active time period and the main wavelength, and triggers the start of the monitoring lamp and wavelength configuration and other wake-up work 5 minutes before the start of the time period, reducing the invalid running time and improving the spatiotemporal accuracy of pest control.

[0039] Micro-control layer: Based on the pest trend labels (rising / stable / falling) output by the short-term pest dynamic response model, the module dynamically adjusts the working period (extending by 30 minutes when rising, shortening by 10 minutes when falling, and maintaining when stable), continuously optimizing operating efficiency while ensuring the accuracy of pest monitoring.

[0040] Workflow diagram as follows Figure 3 As shown. Before each pest control operation begins, a long-term pest prediction model is invoked to obtain a preliminary work plan (work time period). and the wavelength of the trapping lamp And store it in the main control system. At startup time... In the first 5 minutes, the main controller sends commands to the monitoring lights to complete the initial configuration before operation, and then at the start time... Activate the monitoring light. After activation, if the current working period is no longer than one hour, the system will run until the end time. Otherwise, one hour after activation, the main control system will call the short-term insect infestation prediction model every 10 minutes. Based on the short-term future insect infestation trend label output by the model, the system will adjust the end time accordingly. Make adjustments and repeat the process until the end time is reached. The main control system then turns off the monitoring lights and enters sleep mode, repeating the above process 5 minutes before the next working period.

[0041] For example, working time axis Figure 4 As shown.

[0042] In some feasible embodiments, step S6 specifically includes: This embodiment deploys an online lightweight incremental learning mechanism at the edge, achieving continuous adaptive optimization of the prediction model through periodic data collection, lightweight model fine-tuning, and closed-loop verification, while meeting the computing power constraints of farmland edge computing devices. Its specific workflow is as follows: The system periodically (e.g., daily) retrieves insect monitoring data from the local storage module, including environmental parameters (temperature, humidity, light intensity), insect trend tags (rising / stable / falling), and the operating status of the monitoring lamps. The data quality assessment module filters noise from the collected data and removes sensor outliers (such as sudden changes in temperature and humidity exceeding thresholds) to ensure the reliability of the input data and provide high-quality training samples for incremental learning.

[0043] The prediction model based on the LNN-Attention architecture only incrementally updates the parameters of the multi-head attention mechanism (preserving the hidden layer structure of the liquid neural network), avoiding full model retraining. Gradient compression algorithms (such as Top-K sparsity) reduce computational complexity (typically by more than 65%), making the model update overhead suitable for the computing power limitations of edge devices. This fine-tuning process consumes less than 5% of system resources, ensuring real-time performance and low power consumption.

[0044] The fine-tuned model is validated for prediction accuracy on a small local test set (covering historical pest infestation samples, ≤100 samples). If the accuracy improves by ≥3%, the main model is automatically replaced; otherwise, it is rolled back to a historical stable version. This mechanism minimizes model update overhead (no cloud interaction required), enabling the system to dynamically adapt to regional ecological changes (such as sudden changes in pest patterns and climate anomalies), ensuring the accuracy of pest infestation prediction while avoiding the risk of model drift.

[0045] Based on the overall process described above, this invention also provides a system hardware platform, which consists of an embedded main controller, an insect monitoring lamp, and related sensors. The embedded main controller can be a Linux embedded device (such as a SOC device from brands like Allwinner or Rockchip), the insect monitoring lamp is a product with adjustable wavelength for trapping insects, and the sensors include environmental monitoring sensors such as those for temperature, humidity, and rainfall. The overall framework is as follows: Figure 5 As shown, this system collects environmental data in real time using temperature, humidity, rainfall, and light sensors. It integrates long-term and short-term insect infestation prediction models to achieve accurate insect infestation trend analysis, dynamically driving the insect monitoring lamps to perform on-demand control. The system relies on storage and 4G modules to achieve local-cloud synchronous storage of environmental and insect monitoring data, providing data support for the online incremental learning mechanism. During operation, the system periodically samples sensor data and the operating status of the monitoring lamps, saving the data locally and in the cloud to support subsequent prediction model optimization and continuous evolution of the incremental learning mechanism.

[0046] Based on all the above, this invention uses an edge computing platform as the core and constructs a fusion architecture of liquid neural network (LNN) and multi-head attention mechanism. The LNN is used to extract the periodic and sudden characteristics of insect population density, and the multi-head attention mechanism is used to weight and fuse the correlation between meteorological factors and insect species characteristics to achieve future insect population prediction. At the specific model design level, the system adopts a hierarchical prediction strategy: (1) Long-term insect population prediction model. The historical environment-insect population time series data is used as input, and the LNN-Attention predicts and outputs the structured mapping table of "insect species - active time window - main wavelength" for the day. This model provides a macro decision basis for the scheduling module and generates the preliminary work plan for the day, including the core working period (such as 19:00-22:00) and the wavelength configuration of the trapping lamp required for the corresponding time period (such as 430nm for peach borer). (2) Short-term insect population prediction model. The environmental-insect population time series data of the past 1 hour is used as input, and the LNN-Attention predicts and outputs the insect population density change trend for the next 1 hour. This model provides micro-decision support for the scheduling module, enabling it to determine in real time whether the current work period needs to be extended or shortened, thus achieving dynamic start-stop control.

[0047] Ultimately, the long-term and short-term pest infestation prediction models, deeply coupled with the intelligent scheduling and control module, achieve refined control of the monitoring lamp operation: the structured mapping table of "pest species – active time window – primary wavelength" output by the long-term pest infestation prediction model is parsed by the scheduling module into a macro-level operation plan; the short-term pest infestation prediction model outputs pest trend labels (rising / stable / declining), driving the scheduling module to dynamically adjust the end time—extending operation when the pest infestation is rising, terminating early when it is declining, and maintaining the original plan when it is stable. This forms a two-level scheduling mechanism of "long-term planning + short-term dynamic correction," effectively avoiding ineffective lighting periods while ensuring the integrity of pest monitoring, and achieving on-demand start / stop and targeted trapping. This closed-loop control process schedules the working time and wavelength of the pest monitoring lamps when performing pest control tasks, not only reducing the system's ineffective energy consumption but also improving the trapping specificity and ecological adaptability for different insect species.

[0048] Furthermore, the system incorporates an online incremental learning mechanism to enhance its environmental adaptability and long-term stability. The system can periodically and automatically update its model, using newly collected insect infestation data to correct parameters in both long-term and short-term insect infestation prediction models, thus preventing decision failures due to environmental drift.

[0049] In summary, the insect pest monitoring lamp time-spectrum adaptive scheduling system and method proposed in this invention overcomes the shortcomings of traditional models in terms of energy efficiency, adaptability, and eco-friendliness through algorithmic innovation. Its core objective is to achieve better pest control results with shorter working time and energy consumption. It provides a feasible technical path for green plant protection and precision agriculture, which can effectively reduce pesticide dependence, protect farmland ecosystems, and improve the level of agricultural intelligence.

[0050] A predictive model-based insect monitoring and control system includes: The model building unit is used to execute step S1; The model is divided into units for executing step S2; The scheduling and control unit is used to execute steps S3-S5.

[0051] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0052] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a predictive model-based insect monitoring lamp control method as described above.

[0053] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0054] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for controlling insect pest monitoring lights based on a predictive model, characterized in that, Includes the following steps: A time series prediction model is constructed based on liquid neural networks and multi-head attention mechanisms; Training sample sets with different time spans are constructed and the time series prediction models are trained respectively to obtain long-term insect infestation prediction models and short-term insect infestation prediction models. A preliminary work plan was obtained by calling the long-term insect infestation prediction model; Based on the preliminary work plan, the insect monitoring lamp is turned on. After the insect monitoring lamp is turned on for a preset time, the short-term insect infestation prediction model is called at a preset frequency to output short-term future insect infestation trend labels. The end time of the current working period is adjusted based on the short-term future insect infestation trend label, and this process is repeated until the end time is reached.

2. The insect pest monitoring lamp control method based on a prediction model according to claim 1, characterized in that, The time series prediction model includes an LNN dynamic encoding module, a multi-head attention context enhancement module, and a fusion prediction module, wherein: The input is encoded based on the LNN dynamic encoding module, and the dynamic evolution of the hidden state is described by a continuous-time differential equation to obtain local dynamic features. Based on the multi-head attention context enhancement module, and combined with the local dynamic features, the global dependencies between different time steps and multi-source variables are explicitly modeled to obtain the attention enhancement representation; Based on the fusion prediction module, the local dynamic features and the attention enhancement representation are spliced ​​and fused, and the output is performed according to the fused features.

3. The insect pest monitoring lamp control method based on a prediction model according to claim 2, characterized in that, The step of constructing training sample sets with different time spans and training the time series prediction model separately to obtain the long-term insect infestation prediction model and the short-term insect infestation prediction model specifically includes: A first training set is constructed using time-series data at the first time step, and a second training set is constructed using time-series data at the second time step. The time-series data for the first time step and the time-series data for the second time step both include: time nodes, environmental parameters, insect species image recognition results, and the number of electric shocks from the power grid. The time series prediction model is trained based on the first training set to obtain a long-term insect infestation prediction model. The time series prediction model is trained based on the second training set to obtain a short-term insect infestation prediction model.

4. The insect pest monitoring lamp control method based on a prediction model according to claim 3, characterized in that, The output of the long-term insect infestation prediction model shown includes insect species, key monitoring periods, and recommended wavelengths.

5. The insect monitoring lamp control method based on a prediction model according to claim 3, characterized in that, The output calculation of the short-term insect infestation prediction model specifically includes: Based on a short preset time, predict the sequence of electric shocks to the power grid from the insect monitoring lamps in the future time period; The total number of electric shocks is calculated based on the power grid click sequence; The relative rate of change is calculated based on the total number of electric shocks and the reference baseline value; A trend threshold was set based on quantiles and historical electric shock time series. By combining the relative rate of change and the trend threshold, a short-term future insect infestation trend label is generated.

6. The insect pest monitoring lamp control method based on a prediction model according to claim 5, characterized in that, The short-term future insect infestation trend label is represented as follows: in, Indicates the rising threshold; This indicates the threshold for the decrease.

7. The insect pest monitoring lamp control method based on a prediction model according to claim 1, characterized in that, Also includes: Incremental learning is performed on the long-term insect infestation prediction model and the short-term insect infestation prediction model, and the model parameters are adjusted to obtain the fine-tuned prediction model. If the prediction accuracy of the fine-tuned prediction model is greater than a preset threshold, the original prediction model is replaced.

8. A predictive model-based insect pest monitoring lamp control system, characterized in that, include: The model building unit constructs a time series prediction model based on liquid neural networks and multi-head attention mechanisms; The model is divided into units, training sample sets with different time spans are constructed, and the time series prediction model is trained separately to obtain a long-term insect infestation prediction model and a short-term insect infestation prediction model. The scheduling and control unit is used to call the long-term insect infestation prediction model to obtain a preliminary work plan; based on the preliminary work plan, the insect infestation monitoring lamp is turned on; after the insect infestation monitoring lamp is turned on for a preset time, the short-term insect infestation prediction model is called at a preset frequency to output a short-term future insect infestation trend label; based on the short-term future insect infestation trend label, the end time of the current work period is adjusted, and the process is repeated until the end time is reached.

9. A control device for an insect pest monitoring lamp based on a predictive model, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a method as described in any one of claims 1-7.

Images