Self-cleaning drip irrigation system based on neural network prediction and intelligent agent and control method

Abstract

The application relates to a self-cleaning drip irrigation system based on neural network prediction and an intelligent agent and a control method, and belongs to the technical field of automatic irrigation. The application is used to solve the problems that an irrigation system lacks a prediction ability and a self-adaptive regulation and control mechanism and is difficult to cope with complex and changeable actual environments in a mode of relying on manual inspection and regular flushing. The system comprises a data acquisition and processing module, a neural network model training module and an intelligent agent decision module which are connected in sequence. The method comprises the following steps: performing multi-modal semantic state representation; carrying out deep reasoning based on the LLM cognitive ability optimized based on SFT; performing semantic retrieval and matching of an intelligent tool set; implementing parameterized action generation driven by the LLM to observe the execution result in text; starting strategy correction based on the LLM introspection mechanism; and performing long-term memory coupling and knowledge sedimentation.

Description

Technical Field

[0001] This invention relates to a self-cleaning drip irrigation system and control method based on neural network prediction and intelligent agent, belonging to the field of automatic irrigation technology. Background Technology

[0002] As modern agriculture develops towards precision and intelligence, drip irrigation systems, while possessing the advantages of high efficiency and water saving, generally suffer from pipe blockage and uneven water pressure during long-term operation, leading to decreased irrigation efficiency, resource waste, and uneven crop growth.

[0003] Despite the presence of filtration systems in existing drip irrigation systems, tiny particles of sediment, organic matter, or chemical deposits in the water can easily accumulate in the dead corners of the labyrinthine channels, causing the channels to gradually narrow or even become completely blocked. Around the moist environment, plant roots are more likely to grow towards the water source, resulting in "root invasion." Simultaneously, biofilms formed by rhizosphere microorganisms and algae can adhere to the inner walls of the channels, significantly reducing water uniformity. In underground drip irrigation systems, the instantaneous negative pressure after the system stops operating can easily draw soil particles and slurry from around the drippers back into the emitters, causing physical blockages.

[0004] Existing drip irrigation networks suffer from significant uneven hydraulic distribution when laid in complex terrain or over large areas: due to frictional losses and local resistance losses along the pipeline, the pressure difference between the beginning and end of the drip irrigation pipe is large, resulting in excessive outflow at the beginning and insufficient outflow at the end. Although some existing drippers have pressure compensation functions, their compensation threshold range is usually narrow, and in the intermittent water supply mode with high-frequency start and stop, the response hysteresis of the compensation diaphragm is severe, making it difficult to maintain constant outflow accuracy.

[0005] In summary, traditional irrigation systems, which rely on manual inspections and timed flushing, lack predictive capabilities and adaptive control mechanisms, making them ill-suited to cope with complex and ever-changing real-world environments. Summary of the Invention

[0006] To address the problem that irrigation systems relying on manual inspection and timed flushing lack predictive capabilities and adaptive control mechanisms, making it difficult to cope with complex and ever-changing real-world environments, this invention proposes a self-cleaning drip irrigation system and control method based on neural network prediction and intelligent agents.

[0007] The technical solution adopted by this invention to solve the above problems is: the self-cleaning drip irrigation system based on neural network prediction and intelligent agent proposed in this invention includes: The data acquisition and processing module is used to acquire multivariate time series data and preprocess the multivariate time series data to obtain preprocessed time series data. The neural network model training module is used to perform feature extraction and temporal dependency modeling based on the time series modeling method based on the TransformerEncoder structure according to the preprocessed time series data. It also uses the binary classification cross-entropy loss function to train the neural network model and optimize the parameters through backpropagation to obtain the congestion prediction results. The intelligent agent decision-making module takes the congestion prediction results, the current system state, and the neural network prediction labels as inputs, determines the type of action to be executed based on the decomposition and thinking results of the congestion problem, and makes decisions based on the observation results after the action is executed. When the predicted congestion probability exceeds the set threshold, the corresponding operation is triggered to maintain the normal operation of the system under low-risk conditions.

[0008] Furthermore, the data acquisition and processing module selects at least one of the following methods for preprocessing based on the acquired multivariate time series data: missing value imputation, outlier removal, normalization, and time alignment, to obtain preprocessed time series data. The intelligent agent decision-making module reconstructs continuous time series data into model input samples using a sliding window approach, and uses multivariate time series data of a predetermined number of time steps to predict the congestion status at future moments, thus obtaining neural network prediction labels.

[0009] Furthermore, this invention also proposes a control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agents, including: Step 1: The preprocessed multivariate time series data is fused with the congestion probability predicted by the neural network. The LLM model then transforms the fusion result into semantic description text containing the current operating context. Step 2: Based on the semantic description text, the LLM model trained under SFT supervision and combined with the drip irrigation operating condition domain knowledge base is used to perform logical analysis of the context, generate explanatory reasoning text to determine the blockage attribute, and deduce the potential impact of the blockage attribute on the water supply capacity of the pipeline network. Step 3: Based on the projection results of potential impacts, the LLM model selects the optimal combination of processing tools from the intelligent tool library through semantic label matching; Step 4: Based on the combination of processing tools, the LLM model translates the high-dimensional decision into execution instructions and control parameters that conform to the tool calling specifications and sends them to the execution mechanism; Step 5: Based on the execution results of the actuator, capture the dynamic response data of the pipeline network and convert it into observation results in natural language form, then feed it back to the LLM model; Step 6: Based on the comparison between the returned observation results and the expected goals, the LLM model autonomously analyzes the decision-making bias and corrects the internal reasoning logic. If the expected goals are not achieved, the iteration is restarted based on the corrected reasoning logic. Step 7: Based on the inference logic after iterative convergence, store the complete logic chain into the vector database memory module until the system reaches the preset stable operating state and then the control loop ends.

[0010] Furthermore, the training of the LLM model in step 2 includes: Step 2.1: Based on the sensor time-series data, equipment maintenance logs and expert decision records from the historical operation of the drip irrigation system, the pressure-flow-water quality numerical indicators and pipeline status are converted into natural language operating condition descriptions using a preset semantic mapping template, and the corresponding expert diagnostic conclusions and standard processing actions are matched to form an instruction fine-tuning dataset. Step 2.2: Fine-tune the dataset according to the instructions, and use the masked language model loss function to perform supervised training on the pre-trained LLM model to obtain the initially fine-tuned LLM model; Step 2.3: Based on the initially fine-tuned LLM model, introduce efficient parameter updates using low-rank adaptive methods. While keeping the backbone parameters of the pre-trained LLM model frozen, only inject the low-rank decomposition matrix into the attention mechanism layer of the Transformer architecture for incremental training to obtain the incrementally trained LLM model. Step 2.4: Based on the incrementally trained LLM model, construct a reward model with energy consumption reduction rate and congestion relief speed as reward factors. Achieve model alignment through direct preference optimization or proximal policy optimization algorithms to obtain the LLM model after SFT supervised fine-tuning training.

[0011] Furthermore, step 2.2 specifically includes: Based on the instruction-based fine-tuning dataset, the pre-trained LLM model is supervisedly trained using the masked language model loss function, so that the initially fine-tuned LLM model can learn the physical logic, nonlinear pressure response relationship, and semantics of technical terms in the drip irrigation system.

[0012] Furthermore, the intelligent toolkit in step 3 includes one of the following: backwashing, pickling, sterilization, and pressure regulation.

[0013] Furthermore, in step 6, based on the control results of the LLM model trained under SFT supervision in actual operation, the control cases generated by the intelligent agent are captured in real time and vectorized for evidence storage. When the new experience reaches the preset threshold, the incremental fine-tuning task is automatically triggered based on the vectorized evidence storage to drive the LLM model to continuously iterate on new composite congestion conditions.

[0014] The present invention also proposes a computer storage medium and a computer device, wherein the computer storage medium stores a computer program, and when the computer program is executed by a processor, it implements a control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agent.

[0015] The computer device includes: a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements a control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agents.

[0016] The beneficial effects of this invention are: 1. This invention utilizes a Transformer-based multivariate time series forecasting model to provide early warning of congestion risks. It also employs an intelligent agent to adaptively select the optimal cleaning and control strategies based on system status, thereby shifting from "passive maintenance" to "proactive prediction and autonomous optimization." This solution integrates physical and data-driven methods, not only improving system operating efficiency and resource utilization but also reducing maintenance costs. It holds significant theoretical and practical value for promoting the development of smart agriculture and the Internet of Things in agriculture.

[0017] 2. This invention has a wide range of applications, especially in smart agriculture and agricultural Internet of Things. Attached Figure Description

[0018] Figure 1 This is a structural block diagram of a self-cleaning drip irrigation system based on neural network prediction and intelligent agents. Figure 2 A flowchart illustrating the control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agent; Figure 3 This is a schematic diagram comparing this application with LSTM and CNN models. Detailed Implementation Specific Implementation Method 1 like Figure 1 As shown, the structure of the self-cleaning drip irrigation system based on neural network prediction and intelligent agent described in this embodiment includes a data acquisition and processing module, a neural network model training module, and an intelligent agent decision-making module connected in sequence.

[0020] The data acquisition and processing module is used to acquire multivariate time series data and preprocess the multivariate time series data; The neural network model training module is used to perform feature extraction and temporal dependency modeling on multivariate time series data using a time series modeling method based on the Transformer Encoder structure. It uses a binary classification cross-entropy loss function to train the neural network model and optimizes the parameters through backpropagation. The intelligent agent decision-making module takes the prediction results of the neural network model, the current system state, and the neural network prediction labels as inputs to decompose the congestion problem, determine the type of action to be executed, and observe the results after the action is executed. When the predicted congestion probability exceeds a set threshold, the agent will trigger the corresponding operation; under low-risk conditions, it will maintain the normal operation of the system.

[0021] The self-cleaning drip irrigation system adopts a two-stage architecture of "data-driven prediction + intelligent agent decision-making". First, in the data collection phase, a sensor network (including pressure sensors, flow meters, and water quality sensors) deployed at key locations within the drip irrigation system continuously monitors the system's operating status, collecting multivariate time-series data. The collected data mainly includes indicators such as pressure, flow rate, turbidity, conductivity, and temperature, recorded at fixed time intervals (e.g., 10–30 seconds). Subsequently, the raw data undergoes preprocessing, including missing value imputation, outlier removal, normalization, and time alignment. Simultaneously, the TST algorithm is used to construct clogging labels based on abnormal dripper flow rate drops or manual annotation results, transforming the problem into a binary classification or probabilistic prediction task. To adapt to neural network input, the continuous time-series data is reconstructed into model input samples using a sliding window approach, i.e., using multidimensional sensor data from a past period (e.g., 60 time steps) to predict the clogging status at future moments.

[0022] First, during the data collection phase, a sensor network (including pressure sensors, flow meters, and water quality sensors) deployed at key locations within the drip irrigation system continuously monitors the system's operational status, collecting multivariate time-series data. The collected data primarily includes indicators such as pressure, flow rate, turbidity, conductivity, and temperature, recorded at fixed time intervals (e.g., 10–30 seconds). Subsequently, the raw data undergoes preprocessing, including missing value imputation, outlier removal, normalization, and time alignment. Simultaneously, the TST algorithm is used to construct clogging labels based on abnormal dripper flow rate drops or manual annotation results, transforming the problem into a binary classification or probabilistic prediction task. To adapt to neural network input, the continuous time-series data is reconstructed into model input samples using a sliding window approach, i.e., using multidimensional sensor data from a past period (e.g., 60 time steps) to predict the clogging status at future moments.

[0023] The TST prediction model established in this invention, namely the neural network model, also known as Time Series Transformer, is mainly constructed by the following functional layers stacked together: The input projection layer is used to map the raw values ​​of multidimensional sensors to a high-dimensional feature space. The Positional Encoding Layer is used to inject temporal sequence information into unordered feature vectors. The encoder module, or Encoder Block, is the core component of the model. It contains a multi-head self-attention sub-layer and a position-wise feed-forward network sub-layer. Normalized and residual connection layers, also known as Norm & Residual Layers, are distributed among the sub-layers and are used to stabilize deep network training. The Output Classification Layer, consisting of a global pooling layer and a fully connected linear layer, is responsible for outputting the final blocking probability. It includes the following steps: Step 1: Multivariate Input Preprocessing and Linear Projection. The acquired M-dimensional sensor time series X, with dimensions L * M (where L is the window length), is standardized. A linear transformation is then used to map the original features to a d_model-dimensional latent space. The calculation formula is as follows: (1); In formula (1), The projection matrix is ​​learnable. This is a bias term used to achieve preliminary dimensional fusion of features from multiple sensor sources.

[0024] Step 2: Positional Encoding Injection. Positional encoding, short for Positional Encoding, is used to inject... H 0 Injection Location Information PE The input representation Z0 is obtained by using sine and cosine encoding. This step ensures that the model can identify the dynamic trends of pressure and flow over time.

[0025] The calculation method for the input representation is as follows: (2); Step 3: Multi-head self-attention feature extraction. Z0 is input into a multi-layer attention sub-layer. The core mechanism captures temporal dependencies through multi-head self-attention. The formula for capturing temporal dependencies is: (3); In formula (3), Q , K , VThese are the query, key, and value matrices, respectively. The multi-head mechanism allows the model to simultaneously focus on fluctuations at different frequencies (such as instantaneous pressure surges and slow flow decay).

[0026] Step 4: After the feedforward network and residual connections pass through the attention layer, a nonlinear mapping is performed using the position-by-layer feedforward network. Residual connections and layer normalization are introduced to accelerate convergence and prevent gradient vanishing. (3); This process is repeated N times to extract deep blockage symptoms layer by layer.

[0027] Step 5: Global Pooling and Classification Prediction. Global average pooling is performed on the output of the last layer, compressing all temporal features into a fixed-length vector g. This vector is then passed through a fully connected layer and a sigmoid activation function to output the predicted probability y_hat. (4); Step 6: Loss Function Definition and Backpropagation. The binary classification cross-entropy loss function is used to measure the deviation between the predicted value and the true label y. (5); In formula (5), y is the predicted value, and y is the actual value.

[0028] The Adam optimizer is used to backpropagate based on the loss, iteratively updating the network weight parameters until the model's F1 score on the validation set reaches the expected performance.

[0029] When the TST model predicts congestion, the system triggers the resource scheduling logic of the MCP module. The MCP module, through its built-in standardized resource connector, loads and retrieves the physical topology, historical maintenance records, and environmental meteorological data of the target pipeline segment in real time. The module transforms this multi-dimensional background information into a structured context recognizable by the model and feeds it back to the Agent decision-making layer. This process enables the decision-making layer to logically verify and constrain the TST model's predictions in conjunction with real-time physical conditions, thereby endowing the system with physical-level environmental awareness and significantly reducing the false alarm rate.

[0030] The module has a pre-built standardized toolset that encapsulates complex low-level hydraulic control commands, such as rotary flushing and pump station frequency regulation, into abstract interfaces. After the decision vector generated by the Agent decision chain is issued, the MCP module is responsible for converting it into specific protocol messages suitable for different manufacturers and models of PLCs (Programmable Logic Controllers or actuators).

[0031] By employing a pluggable architecture based on the MCP protocol, this invention decouples the prediction logic from the execution logic, ensuring that the prediction model can quickly adapt to drip irrigation sensors and cleaning devices using various protocols without modifying the underlying code. This not only improves the response accuracy and system power management efficiency for sudden blockages but also significantly enhances the system's deployment flexibility and scalability in complex farmland environments.

[0032] At the decision-making level, an intelligent control module based on React Agent is constructed. This agent takes neural network predictions, the current system state, and neural network prediction labels as input, and dynamically generates control decisions based on a pre-defined think-execute-observate logic. Specifically, when the predicted probability of blockage exceeds a set threshold, the agent will trigger corresponding operations, such as initiating backflushing, performing chemical cleaning, or adjusting pump pressure; under low-risk conditions, it will maintain normal system operation to reduce energy and resource consumption. Through this "perception-decision-execution" cyclical mechanism, the React Agent achieves real-time response and adaptive control of the drip irrigation system's operating status.

[0033] The system's overall workflow revolves around blockage prediction and adaptive control, forming a branch-decision-driven intelligent control mechanism. First, after system startup, a pre-trained BERT model makes an initial assessment of the current pipeline state, evaluating the potential for blockage. When the model determines that "blockage may occur," the system enters the intelligent control path, where the ReactAgent takes over the decision-making process. The Agent autonomously and intelligently determines the cause of the blockage based on current pipeline state data (flow rate, water pressure, and drug concentration), and pinpoints the blockage point based on neural network prediction labels. Then, the Agent invokes the MCP module to intervene in the pipeline structure at the blockage point based on different blockage causes (using different intervention methods for different causes). To regulate water pressure, during the control execution phase, the system performs differentiated adjustments before and after the blockage point: reducing water flow and velocity in the area before the blockage point to reduce pressure buildup; and increasing water flow in the area after the blockage point to restore end-point water supply capacity, ultimately completing a closed-loop control cycle.

[0034] When the initial assessment indicates "no blockage," the system enters the normal adjustment path. It analyzes the pipeline pressure change trend using a predictive model and implements tiered control based on pressure levels: when the pressure is low, the system increases water flow to maintain normal water supply; when the pressure is high, it reduces water flow and velocity to avoid potential risks, thus achieving dynamic balance control of the system. This entire process combines conditional judgment with agent-based decision-making, realizing an intelligent operation mechanism from blockage warning to precise regulation. Specific Implementation Method Two like Figure 2As shown, the steps of the control method for the self-cleaning drip irrigation system based on neural network prediction and intelligent agent described in this embodiment include: S1: Perform multimodal semantic state representation; The numerical indicators such as pressure, flow rate, and water quality collected by sensors are fused with the blockage probability predicted by the neural network, and then the LLM transforms them into semantic descriptive text containing the current operating context. S2: Conduct deep reasoning based on SFT-optimized LLM cognitive abilities; LLM combines its pre-tuned and internalized domain knowledge base for drip irrigation conditions to perform logical analysis of the context, generate explanatory reasoning text to determine the physical or chemical properties of the blockage, and deduce its potential impact on the water supply capacity of the pipeline network.

[0036] SFT in S2 stands for Supervised Fine-Tuning, and its specific steps include: S201: Construct a multimodal expert dataset. By collecting sensor time-series data, equipment maintenance logs, and expert decision records from the historical operation of the drip irrigation system, and using a preset semantic mapping template, convert numerical indicators such as "pressure-flow-water quality" and pipeline status into natural language operating condition descriptions, and pair them with corresponding expert diagnostic conclusions and standard processing actions to form a high-quality instruction fine-tuning dataset. S202: Conduct supervised fine-tuning based on domain knowledge, and use the masked language model loss function to conduct supervised training on the pre-trained LLM to internalize the physical logic, nonlinear pressure response relationship and professional terminology semantics of the drip irrigation system, so as to ensure that the generated reasoning text conforms to the logic of water conservancy engineering. S203: Introduces efficient parameter updates with low-rank adaptive optimization. While keeping the backbone parameters of the pre-trained model frozen, only injects low-rank decomposition matrices into the attention mechanism layer of the Transformer architecture for incremental training, so as to reduce computational resource consumption and prevent the forgetting of general knowledge. S204: Implement reinforcement learning optimization based on environmental feedback, construct a reward model with energy consumption reduction rate and congestion relief speed as reward factors, and achieve model alignment through direct preference optimization or proximal policy optimization algorithms to guide LLM to output low-risk and high-efficiency control decisions; S205: Performs online incremental fine-tuning and knowledge closure. The system captures successful control cases generated by the React Agent in real time and stores them in vector form. When the new experience reaches the preset threshold, the incremental fine-tuning task is automatically triggered, driving the model to continuously iterate and autonomously evolve for new complex congestion conditions.

[0037] S3: Perform semantic retrieval and matching of intelligent toolsets; Based on the reasoning conclusions, LLM selects the optimal combination of processing tools from a tool library containing functions such as backwashing, pickling, sterilization, and pressure control through semantic tag matching. S4: Implement LLM-driven parameterized action generation, translating high-dimensional decisions into specific execution instructions and control parameters that conform to tool calling specifications and sending them to the execution mechanism; S5: Perform textual observation of execution results. The system captures the dynamic response data of the pipeline network and converts it into observation results in natural language form, which are then sent back to the LLM. S6: Initiate strategy correction based on LLM introspection mechanism. LLM analyzes decision-making bias and corrects internal reasoning logic by comparing the observed results with the expected goals. If necessary, it jumps back to S2 to restart the iteration. S7: Perform long short-term memory coupling and knowledge accumulation, store the complete logical chain into the vector database memory module, and end the control cycle when the system reaches the preset stable operating state.

[0038] like Figure 3 As shown, this implementation method trains 10 sets of sample data. The presence of blockage is further determined by parameters related to pressure, flow rate, turbidity, and conductivity. Please refer to Table 1 for details. Table 1

[0039] This embodiment compares the TST model described in this application with the LSTM model and the CNN model. Table 2 shows that when the TST model described in this application is trained, its accuracy steadily increases from 96.77% in the 10th round to 99.48% in the 60th round.

[0040] Table 2

[0041] In summary, this implementation method achieves early warning of congestion risk through a Transformer-based multivariate time series prediction model, and utilizes an intelligent agent to adaptively select the optimal cleaning and control strategy based on the system state, thereby realizing a shift from "passive maintenance" to "proactive prediction and autonomous optimization." This solution integrates physical and data-driven methods, not only improving system operating efficiency and resource utilization but also reducing maintenance costs, and has significant theoretical and practical value for promoting the development of smart agriculture and the Internet of Things in agriculture.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A self-cleaning drip irrigation system based on neural network prediction and intelligent agent, characterized in that, include: The data acquisition and processing module is used to acquire multivariate time series data and preprocess the multivariate time series data to obtain preprocessed time series data. The neural network model training module is used to perform feature extraction and temporal dependency modeling based on the preprocessed time series data using a time series modeling method based on the TransformerEncoder structure, and to train the neural network model using a binary classification cross-entropy loss function and optimize the parameters through backpropagation to obtain the congestion prediction results. The intelligent agent decision-making module takes the congestion prediction results, the current system state, and the neural network prediction labels as inputs, determines the type of action to be executed based on the decomposition and thinking results of the congestion problem, and makes decisions based on the observation results after the execution of the action. When the predicted congestion probability exceeds a set threshold, the corresponding operation is triggered to maintain the normal operation of the system under low-risk conditions.

2. The self-cleaning drip irrigation system based on neural network prediction and intelligent agent according to claim 1, characterized in that, The data acquisition and processing module selects at least one of the following methods for preprocessing based on the acquired multivariate time series data: missing value imputation, outlier removal, normalization, and time alignment, to obtain preprocessed time series data. The intelligent agent decision-making module reconstructs continuous time series data into model input samples using a sliding window approach, and uses multivariate time series data of a predetermined number of time steps to predict the congestion status at future moments, thus obtaining neural network prediction labels.

3. The self-cleaning drip irrigation system based on neural network prediction and intelligent agent according to claim 1, characterized in that, The neural network model includes: The input mapping layer is used to map the raw values ​​from the multidimensional sensor to a high-dimensional feature space; The positional encoding layer is used to inject temporal sequence information into unordered feature vectors; The encoder module contains a multi-head self-attention sublayer and a position-by-position feedforward network sublayer; it is used for temporal dependency capture, nonlinear mapping and deep feature extraction, and to ensure training stability. Normalized and residual connection layers are distributed among the sub-layers to stabilize deep network training; The output classification layer, consisting of a global pooling layer and a fully connected linear layer, is responsible for outputting the final blocking probability.

4. A control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agent, applied to the self-cleaning irrigation system based on neural network prediction and intelligent agent as described in any one of claims 1-3, characterized in that, include: Step 1: The preprocessed multivariate time series data is fused with the congestion probability predicted by the neural network. The LLM model then transforms the fusion result into semantic description text containing the current operating context. Step 2: Based on the semantic description text, the LLM model trained under SFT supervision and combined with the drip irrigation operating condition domain knowledge base is used to perform logical analysis of the context, generate explanatory reasoning text to determine the blockage attribute, and deduce the potential impact of the blockage attribute on the water supply capacity of the pipeline network. Step 3: Based on the inference results of potential impact, the LLM model selects the current optimal combination of processing tools from the intelligent tool library through semantic tag matching; Step 4: Based on the combination of processing tools, the LLM model translates the high-dimensional decision into execution instructions and control parameters that conform to the tool calling specifications and sends them to the execution mechanism; Step 5: Based on the execution results of the actuator, capture the dynamic response data of the pipeline network and convert it into observation results in natural language form, then feed it back to the LLM model; Step 6: Based on the comparison between the returned observation results and the expected goals, the LLM model autonomously analyzes the decision-making bias and corrects the internal reasoning logic. If the expected goals are not achieved, the iteration is restarted based on the corrected reasoning logic. Step 7: Based on the inference logic after iterative convergence, store the complete logic chain into the vector database memory module until the system reaches the preset stable operating state and then the control loop ends.

5. The control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agent according to claim 4, characterized in that, Step 2, training the LLM model, includes: Step 2.1: Based on the sensor time-series data, equipment maintenance logs and expert decision records from the historical operation of the drip irrigation system, the pressure-flow-water quality numerical indicators and pipeline status are converted into natural language operating condition descriptions using a preset semantic mapping template, and the corresponding expert diagnostic conclusions and standard processing actions are matched to form an instruction fine-tuning dataset. Step 2.2: Fine-tune the dataset according to the instructions, and use the masked language model loss function to perform supervised training on the pre-trained LLM model to obtain the initially fine-tuned LLM model. The masked language model loss function adopts the binary classification cross-entropy loss function. Step 2.3: Based on the initially fine-tuned LLM model, introduce efficient parameter updates using low-rank adaptive methods. While keeping the backbone parameters of the pre-trained LLM model frozen, only inject the low-rank decomposition matrix into the attention mechanism layer of the Transformer architecture for incremental training to obtain the incrementally trained LLM model. Step 2.4: Based on the incrementally trained LLM model, construct a reward model with energy consumption reduction rate and congestion relief speed as reward factors, and achieve model alignment through direct preference optimization or proximal policy optimization algorithms to obtain the LLM model trained under SFT supervision. The binary classification cross-entropy loss function is: （1）； In formula (1), y is the predicted value, and y is the actual value.

6. The control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agent according to claim 5, characterized in that, Step 2.2 specifically includes: Based on the instruction-based fine-tuning dataset, the pre-trained LLM model is supervisedly trained using the masked language model loss function, so that the initially fine-tuned LLM model can learn the physical logic, nonlinear pressure response relationship, and semantics of technical terms in the drip irrigation system.

7. The control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agent according to claim 4, characterized in that, The intelligent toolkit in step 3 includes one of the following: backwashing, pickling, sterilization, and pressure control.

8. The control method for a self-cleaning drip irrigation system based on neural network prediction and intelligent agent according to claim 4, characterized in that, In step 6, based on the control results of the LLM model trained by SFT supervision in actual operation, the control cases generated by the intelligent agent are captured in real time and vectorized for evidence storage. When the new experience reaches the preset threshold, the incremental fine-tuning task is automatically triggered according to the vectorized evidence storage to drive the LLM model to continuously iterate on new composite congestion conditions.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the control method of the self-cleaning drip irrigation system based on neural network prediction and intelligent agent as described in any one of claims 4-8.

10. A computer device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the control method for the self-cleaning drip irrigation system based on neural network prediction and intelligent agent as described in any one of claims 4-8.

Images