A voice-controlled intelligent water-saving irrigation control system

Abstract

The application relates to the technical field of automatic control of agricultural irrigation, and particularly discloses an intelligent water-saving irrigation control system controlled by voice, which comprises a voice instruction analysis module, a multi-source data fusion sensing module, a dynamic irrigation strategy generation module, an irrigation execution control module and a system configuration and learning module. The semantic and demand intensity of a user voice instruction are analyzed, real-time soil moisture, weather forecasts and crop growth model data are fused, and precise irrigation instructions are dynamically generated based on a water-saving optimization target to drive an execution mechanism to complete irrigation. The application realizes the combination of the convenience of voice control and the scientificity of irrigation decision-making, and significantly improves the utilization efficiency of water resources.

Description

Technical Field

[0001] This invention belongs to the field of agricultural irrigation automation control technology, specifically relating to a voice-controlled intelligent water-saving irrigation control system. Background Technology

[0002] Smart agriculture technology is an important direction for the development of modern agriculture. Its core lies in utilizing advanced technologies such as the Internet of Things and artificial intelligence to achieve precise perception, intelligent decision-making, and automatic control of the agricultural production process, aiming to improve resource utilization efficiency and ensure the quality and yield of agricultural products. As a key component of smart agriculture, the intelligent water-saving irrigation control system aims to achieve water conservation and efficient utilization through the automated and intelligent management of the irrigation process.

[0003] In existing technologies, voice-controlled irrigation systems have been applied, allowing users to conveniently trigger irrigation operations via voice commands (such as "water the vegetable garden"). However, such systems typically only perform simple keyword matching on voice commands, binding specific commands to preset fixed irrigation areas or fixed durations, without intelligently associating and comprehensively analyzing voice commands with real-time environmental data (such as soil moisture and weather forecasts) and crop growth needs (such as crop type and growth stage).

[0004] This leads to a serious disconnect between the system's operations and actual irrigation needs. For example, the system may still execute the "water the vegetable garden" command even when the soil is already saturated or after rainfall, resulting in a serious waste of water resources and deviating from the original intention of water-saving irrigation. Summary of the Invention

[0005] The purpose of this invention is to provide a voice-controlled intelligent water-saving irrigation control system to solve the technical contradiction in the prior art where voice commands are disconnected from real-time environmental data and crop needs, resulting in rigid irrigation operations and serious water waste.

[0006] This invention provides a voice-controlled intelligent water-saving irrigation control system, which includes a voice command parsing module, a multi-source data fusion sensing module, a dynamic irrigation strategy generation module, an irrigation execution control module, and a system configuration and learning module.

[0007] The voice command parsing module receives and parses users' natural language irrigation commands, extracting semantic intent, target area information, and implicit irrigation demand intensity. The multi-source data fusion sensing module collects and processes data from sensor networks deployed in the irrigation area, weather forecast data, and preset crop growth model data in real time, generating a quantitative assessment of current irrigation demand.

[0008] The dynamic irrigation strategy generation module receives the command parsing results from the voice command parsing module and the quantitative evaluation results from the multi-source data fusion perception module. Based on the preset water-saving optimization objective function, it performs dynamic planning calculations of irrigation strategies under multiple constraints, generating a precise irrigation command sequence that includes specific irrigation areas, irrigation duration, irrigation flow rate, and start time.

[0009] The irrigation execution control module receives and executes the precise irrigation command sequence generated by the dynamic irrigation strategy generation module, drives the corresponding solenoid valves and water pumps to perform irrigation operations, and provides feedback on the execution status. The system configuration and learning module manages crop growth model parameters and water-saving optimization objective function parameters, and adaptively optimizes and updates the model parameters based on historical irrigation effect data.

[0010] Furthermore, the voice command parsing module includes a speech recognition unit, a semantic understanding unit, and a command standardization unit. The speech recognition unit uses a deep neural network-based acoustic and language model to convert the input speech signal into text commands. The semantic understanding unit performs named entity recognition and dependency parsing on the text commands to identify core action entities, regional entities, and modifier entities. Core action entities include "irrigate," "stop," and "query." Regional entities include user-defined geographical region names.

[0011] Modifying entities include adverbs or phrases that characterize the intensity of demand. The instruction standardization unit maps the output of the semantic understanding unit to a structured instruction data object, which contains an instruction type field, a target region list field, and a demand intensity coefficient field. The demand intensity coefficient is a floating-point number between 0 and 1, and its initial value is assigned based on the identified modifying entities. When no modifying entities are identified in the instruction, the coefficient is assigned a default value of 0.5.

[0012] Furthermore, the multi-source data fusion sensing module includes an environmental data acquisition unit, a crop demand calculation unit, and a comprehensive demand assessment unit. The environmental data acquisition unit connects to soil moisture sensors, temperature sensors, light sensors, and a network meteorological data interface to acquire real-time soil volumetric water content, ambient temperature, light intensity, and precipitation probability and forecast data for the next 24 hours at each monitoring point, using a preset sampling period. The crop demand calculation unit pre-stores water requirement models for different crop types at different growth stages, with crop coefficients, reference crop evapotranspiration, and soil moisture stress coefficients as core parameters.

[0013] The crop demand calculation unit calculates the theoretical water requirement of the crop at the current moment based on the current crop type and growth stage information provided by the system configuration and learning module, combined with real-time environmental data provided by the environmental data acquisition unit. The comprehensive demand assessment unit integrates environmental data and crop demand data to generate a quantitative assessment of irrigation demand.

[0014] This unit first calculates the soil moisture deficit based on real-time soil volumetric moisture content and the upper and lower limits of suitable moisture content for crops. Second, it corrects the soil moisture deficit by incorporating future precipitation probability and precipitation forecast data, resulting in a net irrigation demand index. Finally, it weights and fuses the net irrigation demand index with the theoretical water requirement of crops to output a global irrigation demand urgency score, ranging from 0 to 100.

[0015] Furthermore, the dynamic irrigation strategy generation module is the core decision-making unit of the system, internally running an irrigation strategy optimization engine based on model predictive control. The optimization cycle of this engine is synchronized with the minimum control cycle of the irrigation execution control module. Within each optimization cycle, the engine receives the current structured instruction data object and an irrigation demand urgency score.

[0016] The engine's built-in water-saving optimization objective function aims to minimize the total irrigation water consumption, maximize irrigation uniformity, and minimize the deviation from the user's demand intensity coefficient. Constraints include pipeline water pressure constraints, solenoid valve switching frequency constraints, soil moisture content not exceeding field capacity constraints, and single irrigation duration not exceeding the maximum allowable duration constraints.

[0017] The dynamic irrigation strategy generation module first determines the set of irrigation sub-regions to be optimized based on the target region list field. Next, it constructs an initial irrigation strategy by using the irrigation demand urgency score as the initial irrigation volume benchmark for each sub-region and the demand intensity coefficient in the user command as the irrigation volume adjustment weight. Then, the irrigation strategy optimization engine iteratively optimizes the initial irrigation strategy while satisfying all constraints, calculating the optimal irrigation duration sequence and corresponding valve opening sequence for each sub-region over several future control cycles. These sequences together constitute the precise irrigation command sequence. When the command type is stop, the dynamic irrigation strategy generation module generates an all-zero irrigation command sequence and sends it to the irrigation execution control module to immediately terminate all irrigation operations.

[0018] Furthermore, the irrigation execution control module includes an instruction sequence parsing unit, a drive unit, and a status feedback unit. The instruction sequence parsing unit receives a precise irrigation instruction sequence and decomposes it into independent low-level control commands based on time slices and regions. The drive unit, according to the low-level control commands, controls the opening and closing of solenoid valves on the corresponding irrigation branches via relay circuits or industrial buses, and adjusts the output flow of the water pump via frequency converters or pressure regulating valves to achieve precise control of irrigation duration and flow rate. The status feedback unit monitors the on / off status of each solenoid valve, pipeline pressure, and water pump operating current in real time, and packages this status information to feed back to the system configuration and learning module.

[0019] Furthermore, the system configuration and learning module includes a parameter management unit and a model adaptation unit. The parameter management unit provides a human-computer interaction interface for users or administrators to set and update system parameters such as crop type, growth stage, mapping relationship between irrigation area and valve, upper and lower limits of soil moisture content, and maximum allowable irrigation duration.

[0020] The model adaptation unit continuously collects historical irrigation data, including the instructions for each irrigation, the executed irrigation strategy, changes in soil moisture before and after irrigation, and subsequent real-time weather data. Employing a reinforcement learning algorithm, the unit uses changes in soil moisture and crop growth after irrigation as reward signals to fine-tune key parameters in the crop water requirement model and optimize the weight coefficients in the water-saving objective function. This allows the system's decision model to adapt to specific field microenvironments and user preferences, continuously improving water-saving effectiveness and instruction compliance.

[0021] As one embodiment of the present invention, the voice command parsing module supports the processing of commands for ambiguous regions. When the semantic understanding unit identifies the region entity as "all" or "the entire garden," the command standardization unit will query the system configuration and learning module for all configured irrigation regions and fill the target region list field with the identifier list of these regions.

[0022] In one embodiment of the present invention, the comprehensive demand assessment unit in the multi-source data fusion sensing module has a weighted fusion algorithm in which the weight of the net irrigation demand index is set to 0.7 and the weight of the theoretical water requirement of crops is set to 0.3. This weight ratio can be adjusted according to the drought resistance characteristics of different crops through the system configuration and learning module.

[0023] In one embodiment of the present invention, the irrigation strategy optimization engine of the dynamic irrigation strategy generation module has a prediction time domain of 6 control cycles and a control time domain of 3 control cycles. In each optimization cycle, the engine executes only the instructions of the first control cycle within the control time domain and performs rolling optimization in the next cycle.

[0024] As one embodiment of the present invention, the model adaptive unit of the system configuration and learning module adopts the near-end policy optimization algorithm for its reinforcement learning algorithm, and the update cycle is set to perform model parameter update once after every 30 irrigation operations are completed.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention fundamentally changes the disconnect between commands and execution in traditional voice irrigation systems by constructing a collaborative working system that integrates voice command parsing, multi-source data fusion perception, and dynamic strategy generation. The system can deeply analyze the semantics and intent of user commands, while simultaneously integrating multi-dimensional data such as real-time soil moisture, weather forecasts, and crop growth models. Under a unified water-saving optimization framework, it performs dynamic planning, transforming a simple "water the vegetable patch" voice command into a refined and personalized irrigation plan that fully considers current water scarcity, future rainfall, and the actual water requirements of the crops. This fundamentally avoids ineffective irrigation when the soil is already wet or before or after rainfall, achieving a deep integration of the convenience of voice control and the scientific basis of water-saving irrigation, significantly improving water resource utilization efficiency. 2. This invention introduces a dynamic irrigation strategy generation mechanism based on model predictive control, upgrading irrigation decisions from static, open-loop triggering actions to a dynamic, closed-loop optimization process. This mechanism not only responds to current user commands and data but also proactively considers short-term system state changes and seeks optimal solutions under multiple physical constraints such as pipeline water pressure and valve lifespan. This rolling optimization and feedback correction characteristic enables the system to proactively adapt to environmental disturbances. For example, when new weather warning data is received during irrigation, the subsequent irrigation strategy can be dynamically adjusted, thereby minimizing water consumption and maximizing irrigation uniformity while meeting crop needs, ensuring the economy and reliability of irrigation operations. 3. This invention designs a system configuration and learning module with self-learning capabilities, enabling the system to surpass the initially set fixed model parameters. By continuously collecting irrigation performance data and applying reinforcement learning algorithms, the system can automatically fine-tune the crop water requirement model and optimize target weights, continuously approximating the actual water dynamics of a specific field and the user's actual preferences. This means that the system's decision-making intelligence will continuously evolve with usage time, becoming more accurate and water-saving over time. This effectively solves the model mismatch problem caused by soil spatial variability, crop variety differences, or climate change, giving the system long-term adaptability and robustness, and reducing the cost of later manual maintenance and parameter adjustments. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical architecture of the voice-controlled intelligent water-saving irrigation control system proposed in this invention. Figure 2This is a schematic diagram of the core principle framework of the dynamic irrigation strategy generation module based on model predictive control in this invention; Figure 3 This is a flowchart illustrating the logical process of the multi-source data fusion sensing module generating an irrigation demand urgency score in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of voice command parsing and dynamic strategy generation working together in this invention; Figure 5 This is a schematic diagram illustrating the principle framework of the system configuration and learning module in this invention for achieving adaptive optimization of model parameters. Detailed Implementation

[0027] Please refer to Figures 1 to 5 This invention provides a voice-controlled intelligent water-saving irrigation control system. Please refer to the appendix. Figure 1 The system is physically deployed in a central control server or a high-performance embedded gateway device, and logically consists of a voice command parsing module, a multi-source data fusion and sensing module, a dynamic irrigation strategy generation module, an irrigation execution control module, and a system configuration and learning module. These modules exchange data and transmit commands through an internal message bus or shared memory, forming a complete closed loop from voice command input to physical device execution, and then to system self-optimization.

[0028] The voice command parsing module serves as the system's human-computer interaction entry point, responsible for converting user-issued natural language voice commands into structured command data that machines can understand. This module is deployed on terminal devices with audio input capabilities, such as smart speakers, smartphones, or dedicated field control terminals. Internally, it comprises three sub-units: a voice recognition unit, a semantic understanding unit, and a command standardization unit. These three sub-units execute sequentially in a pipeline manner.

[0029] The speech recognition unit receives raw speech signals from a microphone array or audio interface. The unit first preprocesses the input analog speech signal, including pre-emphasis, framing, and windowing operations, to enhance high-frequency features and reduce spectral leakage. Subsequently, the Mel-frequency cepstral coefficients of each frame of the speech signal are extracted as acoustic feature vectors. The speech recognition unit employs a pre-trained deep neural network acoustic model, composed of multiple convolutional layers and long short-term memory network layers, to map the acoustic feature vector sequence into a phoneme probability sequence.

[0030] Simultaneously, the unit loads a language model trained on a large-scale Chinese text corpus. This model employs a recurrent neural network or transformer architecture to decode phoneme sequences based on context and output the most probable text instruction string. For example, if a user says "Water the vegetable garden to the east this afternoon," the speech recognition unit, after processing as described above, outputs the corresponding text string "Water the vegetable garden to the east this afternoon." The speech recognition unit possesses online learning capabilities, enabling it to fine-tune the acoustic model based on the pronunciation habits of specific users, thereby improving recognition accuracy.

[0031] The semantic understanding unit receives text instruction strings output by the speech recognition unit. The core task of this unit is to perform deep semantic analysis of the text and identify key information entities within the instructions. The semantic understanding unit incorporates a natural language processing engine, which first performs word segmentation and part-of-speech tagging on the text. Next, the engine performs named entity recognition, identifying named entities in the text through a sequence labeling model based on a bidirectional long short-term memory network combined with conditional random fields. In this system, named entities are mainly divided into three categories: core action entities, region entities, and descriptive entities.

[0032] Core action entities represent the fundamental operations of user intent, and their vocabulary includes "irrigate," "water," "sprinkler irrigation," "drip irrigation," "stop," "close," "query," and "view." Region entities represent the geographical area to which the irrigation operation is applied, and their content is usually a region name that the user defines during the system configuration phase, such as "East Vegetable Garden," "West Orchard," "Greenhouse No. 1," "All," and "The Entire Garden."

[0033] Modifying entities are used to represent the implicit requirements of users for the intensity or degree of irrigation operations, usually manifested as adverbs or specific phrases, such as "slightly," "a little," "fully," "thoroughly," and "a lot." After completing named entity recognition, the semantic understanding unit further performs dependency parsing to construct a syntactic dependency tree of the text to determine the modification and modified relationships between the core action entity, regional entities, and modifying entities, thereby accurately understanding the instruction structure. For example, for the text "Water the vegetable garden on the east side a little more this afternoon," the semantic understanding unit identifies the core action entity as "water," the regional entity as "the vegetable garden on the east side," and the modifying entity as "thoroughly." Through dependency parsing, it is determined that "thoroughly" modifies the core action "water," and "the vegetable garden on the east side" is the object of the action "water."

[0034] The instruction standardization unit receives the output from the semantic understanding unit. Its function is to map unstructured semantic analysis results into unified, structured instruction data objects for subsequent processing. The instruction data object is a data structure with fixed fields, primarily including three key fields: instruction type, target area list, and demand intensity coefficient. The instruction type field is an enumerated value, mapped based on the identified core action entities.

[0035] If the core action entity belongs to the "Irrigation" category, the instruction type field is assigned the value "Irrigation Instruction". If it belongs to the "Stop" category, it is assigned the value "Stop Instruction". If it belongs to the "Query" category, it is assigned the value "Query Instruction". The target area list field is a string array used to store unique identifiers for one or more irrigation areas.

[0036] The instruction standardization unit populates this field based on the identified regional entity content. When the regional entity is a specific configured regional name, such as "Eastern Vegetable Garden," the unit directly queries the system configuration and learning module for the regional identifier corresponding to that name and fills it into the array. When the regional entity is a vague description, such as "all" or "the entire garden," the instruction standardization unit initiates a query request to the system configuration and learning module to obtain a list of identifiers for all activated irrigation regions and fills the entire list into the target region list field.

[0037] The demand intensity coefficient field is a floating-point number, strictly limited to a range of 0 to 1, used to quantify the user's subjective requirements for irrigation intensity. The instruction standardization unit has a built-in mapping table from modifier entities to demand intensity coefficients. For example, recognizing "slightly" or "a little" maps to 0.3, recognizing "normal" or no modifier maps to 0.5, recognizing "sufficiently" maps to 0.7, and recognizing "thoroughly watered" or "large amount" maps to 0.9. The instruction standardization unit queries this mapping table and assigns values ​​based on the modifier entities recognized by the semantic understanding unit. If the instruction does not contain any identifiable modifier entities, the field uses the default value of 0.5.

[0038] Finally, the instruction standardization unit outputs a complete instruction data object. For example, for the instruction mentioned above, the output object might be: Instruction type field = "Irrigation Instruction", Target area list field = ["zone_east_veg"], Demand intensity coefficient field = 0.9. This object is encapsulated into a message of a specific format and sent to the dynamic irrigation strategy generation module via an internal communication protocol.

[0039] The multi-source data fusion sensing module is the system's environmental sensing and data fusion center, responsible for providing the objective quantitative basis needed for irrigation decisions. Please refer to the appendix. Figure 3This module runs continuously, performing a data acquisition, calculation, and evaluation process at a fixed sampling period, such as every 10 minutes. Internally, the module consists of an environmental data acquisition unit, a crop requirement calculation unit, and a comprehensive requirement assessment unit.

[0040] The environmental data acquisition unit is responsible for acquiring real-time and predictive data from the physical world and external data sources. This unit connects to a sensor network deployed within the irrigation area via wired or wireless means. Typical sensors include soil moisture sensors, which use frequency domain reflectance or time domain transmission principles to measure the volumetric water content of the soil, with a measurement accuracy of ±2%. Data is uploaded via analog input modules or digital interfaces such as RS485 Modbus protocol.

[0041] Temperature and light sensors are used to monitor the microclimate environment near the crop canopy. In addition, the environmental data acquisition unit regularly accesses data from authoritative meteorological service providers via a web application programming interface to obtain gridded weather forecast data for the next 24 hours or even longer. Among the data items that are crucial for irrigation decisions are precipitation probability, forecast precipitation, temperature, humidity, wind speed, and sunshine duration.

[0042] The environmental data acquisition unit performs verification, filtering, and outlier removal on the raw data. For example, for soil moisture data, the unit checks whether it is within the physically possible range of 0% to 100% and uses a moving average filter to smooth short-term fluctuations. The processed data is organized into a standard-format environmental data package, containing the timestamp of each monitoring point, soil volumetric water content, ambient temperature, light intensity, and precipitation probability and forecast values ​​for the next six 4-hour time slots.

[0043] The core of the crop demand calculation unit is an embedded crop water requirement model. This model is based on the Penman-Montis formula and its derivatives recommended by the Food and Agriculture Organization of the United Nations. Its core is to calculate the reference crop evapotranspiration, and then correct it with the crop coefficient and soil moisture stress coefficient to obtain the actual water requirement of the crop.

[0044] The parameter management unit of the system configuration and learning module provides static parameters to the crop demand calculation unit, including the types of crops currently planted in the irrigated area and the current growth stage of the crops, such as seedling stage, vegetative growth stage, flowering stage, and maturity stage. Based on the crop type and growth stage, the crop demand calculation unit queries the corresponding crop coefficient benchmark value from the built-in database. Simultaneously, the unit receives real-time environmental data such as temperature, humidity, wind speed, and sunshine duration from the environmental data acquisition unit and uses the Penman-Montis formula to calculate the reference crop evapotranspiration for the current period.

[0045] The calculation formula involves multiple meteorological parameters, and its core principle, described in plain text, is as follows: the reference crop evapotranspiration is determined by both net radiation and aerodynamic terms. The net radiation term depends on solar radiation, air temperature, and humidity. The aerodynamic term depends on wind speed, air temperature, and humidity. The crop demand calculation unit multiplies the calculated reference crop evapotranspiration by the crop coefficient to obtain the theoretical water requirement of the crop without water stress.

[0046] In addition, the unit calculates the soil water stress coefficient based on the relationship between real-time soil volumetric water content, crop wilting point water content, and field water holding capacity. This coefficient begins to decrease linearly when the soil water content is below a threshold, thereby reducing the theoretical water requirement and finally outputting a value of the crop's theoretical water requirement that is more consistent with the actual water stress situation, in millimeters per day.

[0047] The integrated demand assessment unit is a crucial component of multi-source data fusion. Its task is to integrate environmental data and crop demand data into a single, quantifiable score indicating the urgency of irrigation needs. Please refer to the appendix. Figure 3 The execution logic of this unit is divided into three steps.

[0048] The first step is to calculate the soil moisture deficit. The unit reads the real-time soil volumetric moisture content of each monitoring point from the environmental data packet and compares it with the upper and lower limits of the suitable moisture content for crops in the area set in the system configuration. The soil moisture deficit is defined as the degree to which the current moisture content is lower than the lower limit of the suitable moisture content. After normalization, a value between 0 and 1 is obtained, where 1 indicates severe water shortage.

[0049] The second step involves refining the net irrigation demand index. This step incorporates future weather forecast data as a correction factor. It retrieves the probability and forecast amount of precipitation for the next 24 hours. If effective precipitation is forecast, the amount of soil moisture replenishment from this precipitation is estimated using a precipitation infiltration model, and the soil moisture deficit calculated in the first step is adjusted accordingly to obtain the net irrigation demand index. The core idea of ​​this step is to avoid unnecessary irrigation before rainfall.

[0050] The third step involves weighted fusion to generate the final score. The comprehensive demand assessment unit weights and fuses the net irrigation demand index with the theoretical crop water requirement output by the crop demand calculation unit. In a typical configuration, the weight of the net irrigation demand index is set to 0.7, and the weight of the theoretical crop water requirement is set to 0.3. This reflects the system's decision-making tendency to prioritize responses to real-time soil moisture and recent weather conditions. The weighted calculation result is linearly scaled and mapped to an integer range of 0 to 100, thus obtaining the irrigation demand urgency score. The higher the score, the more urgently the area needs irrigation. This score is based on a regional unit, with each configured irrigated area generating an independent score value for each assessment period.

[0051] The dynamic irrigation strategy generation module is the intelligent decision-making center of the system. Please refer to the appendix. Figure 2 The core of this module is an irrigation strategy optimization engine based on model predictive control. Model predictive control is an advanced process control method characterized by using dynamic models to predict the future behavior of the system and continuously optimizing the control input over a future period while satisfying various constraints.

[0052] In this system, the execution cycle of the optimization engine, i.e., the control cycle, is consistent with the minimum control cycle of the irrigation execution control module, for example, set to 15 minutes. This means that the engine is triggered to perform an optimization calculation every 15 minutes.

[0053] At the start of each optimization cycle, the dynamic irrigation strategy generation module receives the latest instruction data object from the voice instruction parsing module, as well as the latest irrigation demand urgency score for all irrigation areas from the multi-source data fusion perception module. Internally, the module first determines the instruction type.

[0054] If the instruction type is "stop instruction," the module will immediately generate a precise irrigation instruction sequence of all zeros. This sequence instructs all valves to close, water pumps to stop, and is sent to the irrigation execution control module to forcibly terminate any ongoing irrigation operations. If the instruction type is "query instruction," the module generates a status report data packet based on the current area scores and system status, and provides feedback to the user via voice synthesis or a graphical interface, without generating an irrigation instruction.

[0055] For the most common "irrigation instructions," the dynamic irrigation strategy generation module initiates a complete strategy optimization process. First, the module determines the set of irrigation sub-regions that need to participate in the optimization calculation based on the target region list field in the instruction data object. Next, the module uses the current irrigation demand urgency score of these sub-regions as the baseline for the basic irrigation demand.

[0056] Then, the demand intensity coefficient from the user's instructions is introduced as an adjustment factor. Specifically, the module multiplies the irrigation demand urgency score of each sub-region by the demand intensity coefficient to obtain the preliminary irrigation demand value adjusted by the user's subjective wishes. This value constitutes the initial expected trajectory of the model predictive control optimization problem.

[0057] Subsequently, the irrigation strategy optimization engine begins operation. The engine incorporates a simplified dynamic model of the irrigation system to predict changes in soil moisture content under varying valve openings and irrigation durations. The engine's optimization objective is defined by a water-saving optimization objective function, which is a weighted sum of multiple objectives and primarily comprises three sub-objectives: Minimize the total irrigation water consumption within the prediction time domain; To maximize the uniformity of final soil moisture content in each sub-region; Minimize the deviation between the actual irrigation volume and the initial irrigation demand value adjusted by the user demand intensity coefficient.

[0058] These three sub-objectives are harmonized by weight coefficients, which are stored in the system configuration and learning module and can be adjusted through model adaptive unit learning.

[0059] The optimization process is subject to a series of hard and soft constraints. Hard constraints must be strictly satisfied, including: Maximum and minimum operating water pressure constraints for the pipeline system to prevent pipe bursts or irrigation failures; The maximum number of times each solenoid valve can be switched on and off per unit time is limited to protect the mechanical life of the valve. The predicted soil moisture content in any area must not exceed its field capacity to avoid deep seepage and nutrient loss; The duration of a single continuous irrigation session in any sub-region must not exceed the system's preset maximum allowable duration, such as 4 hours. Soft constraints are penalized as part of the optimization objective, such as hoping for gradual changes in valve opening.

[0060] The engine's prediction time domain is set to 6 control cycles, meaning it predicts the system state 90 minutes in advance. The control time domain is set to 3 control cycles, meaning it optimizes and calculates control commands for the next 45 minutes. In each optimization cycle, the engine solves a constrained optimization problem, calculating the optimal irrigation duration sequence and the corresponding valve opening sequence for each sub-region within the control time domain in each control cycle.

[0061] The valve opening sequence determines the irrigation flow rate. These sequences collectively constitute a precise irrigation command sequence for a future period. However, model predictive control employs a rolling optimization strategy, meaning that in each cycle, only the command calculated for the first control cycle within the control time domain—the command to be executed immediately—is output to the irrigation execution control module. In the next control cycle, the engine acquires the latest system status, including the latest measured soil moisture values ​​and new weather forecasts, and then re-executes the aforementioned optimization calculations. Both the prediction and control time domains roll forward for each cycle, repeating this process continuously. This mechanism enables the system to dynamically adjust its strategy based on the latest information, exhibiting strong anti-interference and adaptive capabilities.

[0062] The irrigation execution control module serves as a bridge connecting digital commands and physical actuators. This module is typically deployed near programmable logic controllers (PLCs) or industrial IoT gateways located close to the field. The module includes an instruction sequence parsing unit, a drive unit, and a status feedback unit.

[0063] The instruction sequence parsing unit receives a precise irrigation instruction sequence from the dynamic irrigation strategy generation module. This sequence is a time-ordered list, with each entry corresponding to a control cycle. Each entry contains a list of valve identifiers to be opened in that cycle, the planned opening duration for each valve, and the desired pipeline pressure or flow rate setpoint. The instruction sequence parsing unit first verifies the format validity and temporal continuity of the instruction sequence.

[0064] The unit then decomposes the received sequence into time slices. For the control cycle to be executed, the unit generates a set of independent low-level control commands. Each low-level control command explicitly specifies the unique address of the solenoid valve, its action (open or closed), and the absolute time point or delay time for executing the action. For cases requiring flow regulation, the command also includes the target frequency or opening value for the frequency converter or electric regulating valve.

[0065] The drive unit is the component that directly interacts with the execution hardware. Based on the low-level control commands generated by the instruction sequence parsing unit, it sends switching signals to designated solenoid valves via digital output relays or industrial fieldbuses such as Modbus TCP or CAN bus. For water pump control, the drive unit sends speed setpoints to the frequency converter via analog output modules or dedicated communication protocols, thereby precisely adjusting the pump speed and achieving closed-loop control of pipeline pressure and flow. The drive unit has command queuing and conflict detection mechanisms to ensure that conflicting commands are not sent to the same valve simultaneously. Additionally, the unit has built-in safety interlock logic; for example, when the main pipeline pressure is detected to be below a safety threshold, all irrigation operations will automatically stop and a fault will be reported.

[0066] The status feedback unit is responsible for real-time monitoring of the irrigation system's operating status. This unit reads the auxiliary contact signals of each solenoid valve via a digital input module to confirm whether the actual valve opening and closing status matches the command. Pressure transmitters are used to collect real-time pressure values ​​from the main pipeline and key branch pipes. Current transformers monitor the operating current of the water pump motor to determine if the pump is overloaded or running dry.

[0067] The status feedback unit collects these status data at a frequency equal to or higher than the control cycle, such as once per minute, and packages them into status feedback data packets. These packets contain timestamps, device status, pressure, current, and other readings. This data packet is uploaded in real-time to the system configuration and learning module via a communication link, serving as the basis for system monitoring, fault diagnosis, and historical data recording.

[0068] The system configuration and learning module is the core of the system's management and its intelligent evolution engine. Please refer to the appendix. Figure 5 This module provides system initialization, parameter configuration, operation monitoring, and model self-learning functions. The module includes a parameter management unit and a model adaptation unit.

[0069] The parameter management unit provides a graphical or command-line user interface. Through this interface, users or system administrators can perform a range of configuration operations. These include, but are not limited to: Define and edit irrigation zones, assign each zone an easily identifiable name, and establish a mapping relationship between the zone and one or more physical solenoid valves; Select or enter the types of crops grown in this area; Set the current growth stage of the crop; set the physical parameters of the soil in the area, such as field water holding capacity, wilting point, and upper and lower limits of suitable moisture content; Configure the physical constraints of the irrigation system, such as the maximum and minimum working pressure of the pipeline network, the maximum valve opening and closing frequency, and the maximum duration of a single irrigation session; Adjust the weighting parameters of the decision-making model, such as the weighting coefficients in the comprehensive demand assessment unit and the weights of each sub-objective in the water-saving optimization objective function. All configuration information is persistently stored in a structured database for other modules to query and access in real time.

[0070] The model adaptive unit is key to achieving continuous system intelligence. This unit continuously collects the massive amounts of data generated during system operation, forming a historical irrigation dataset.

[0071] Each record in the dataset represents a complete irrigation event, containing information in multiple dimensions: the original voice command text that triggered the irrigation and the parsed command data object, the precise irrigation command sequence calculated and output by the dynamic irrigation strategy generation module, the actual execution status sequence fed back by the irrigation execution control module, the soil moisture changes at key monitoring points before and after irrigation, and the actual weather conditions after irrigation.

[0072] The model adaptive unit employs a reinforcement learning framework to optimize the system's internal model. It treats the entire irrigation control system as an agent, each irrigation decision as an action, and the improvement in soil moisture state and crop growth response after irrigation as reward signals. The unit utilizes the advanced reinforcement learning algorithm, the proximal policy optimization algorithm.

[0073] The algorithm update cycle is not based on fixed intervals, but rather triggered by a certain number of irrigation operations completed. For example, it is set to update the model parameters every 30 irrigation operations completed. Within the update cycle, the model adaptive unit uses the collected historical dataset to calculate the performance of the current strategy, i.e., the performance of the current crop water requirement model parameters and the weights of the water-saving optimization objective function.

[0074] Then, the algorithm fine-tunes these model parameters and weights using gradient ascent, enabling the system to make decisions that yield higher rewards in similar future scenarios—namely, decisions that are more water-efficient, better suited to crop needs, and more aligned with user intentions. For example, the algorithm might discover that the initial crop coefficient model overestimated water requirements for a specific field, leading to frequent light irrigation. Through learning, it automatically lowers the crop coefficient for that area, making the model predictions closer to reality and reducing unnecessary irrigation. This self-learning mechanism allows the system to adapt to the spatiotemporal variability of the field microenvironment and the user's unique irrigation habits and preferences, achieving a long-term effect of becoming increasingly accurate with use.

Claims

1. A voice-controlled intelligent water-saving irrigation control system, characterized in that, include: The voice command parsing module is used to receive and parse the user's natural language irrigation commands, extract the semantic intent, target area information, and implicit irrigation demand intensity from the commands; The multi-source data fusion sensing module is used to collect and process data from sensor networks deployed in the irrigation area, weather forecast data, and preset crop growth model data in real time to generate a quantitative assessment of the current irrigation demand. The dynamic irrigation strategy generation module is used to receive the instruction parsing result from the voice instruction parsing module and the quantitative evaluation result from the multi-source data fusion perception module, and perform dynamic planning calculation of irrigation strategy under multiple constraints based on the preset water-saving optimization objective function to generate a precise irrigation instruction sequence that includes specific irrigation area, irrigation duration, irrigation flow rate and start time. The irrigation execution control module is used to receive and execute the precise irrigation instruction sequence generated by the dynamic irrigation strategy generation module, drive the corresponding solenoid valves and water pumps to perform irrigation operations, and provide feedback on the execution status. The system configuration and learning module is used to manage crop growth model parameters and water-saving optimization objective function parameters, and to adaptively optimize and update model parameters based on historical irrigation effect data. The multi-source data fusion sensing module includes an environmental data acquisition unit, a crop demand calculation unit, and a comprehensive demand assessment unit. The environmental data acquisition unit connects to soil moisture sensors, temperature sensors, light sensors, and a network meteorological data interface to acquire real-time soil volumetric water content, ambient temperature, light intensity, and precipitation probability and forecast data for the next 24 hours at each monitoring point, using a preset sampling period. The crop demand calculation unit pre-stores water requirement models for different crop types at different growth stages, and is used to calculate the water requirement based on the current crop type and growth conditions provided by the system configuration and learning module. The system uses long-term information and real-time environmental data provided by the environmental data acquisition unit to calculate the theoretical water requirement of the crop at the current moment. The comprehensive demand assessment unit is used to integrate environmental data and crop demand data to generate a quantitative assessment of crop irrigation demand. Based on the real-time soil volumetric moisture content and the upper and lower limits of suitable crop moisture content, the soil moisture deficit is calculated. The soil moisture deficit is corrected by combining future precipitation probability and precipitation forecast data to obtain a net irrigation demand index. The net irrigation demand index is weighted and integrated with the theoretical water requirement of the crop to output a global irrigation demand urgency score. The dynamic irrigation strategy generation module internally runs an irrigation strategy optimization engine based on model predictive control; The optimization cycle of the irrigation strategy optimization engine is synchronized with the minimum control cycle of the irrigation execution control module. In each optimization cycle, the irrigation strategy optimization engine receives the current structured instruction data object and irrigation demand urgency score, and performs iterative optimization based on the built-in water-saving optimization objective function and constraints to calculate the optimal irrigation duration sequence and corresponding valve opening sequence for each sub-region in the next few control cycles. These sequences together constitute the precise irrigation instruction sequence.

2. The voice-controlled intelligent water-saving irrigation control system according to claim 1, characterized in that, The voice command parsing module includes a voice recognition unit, a semantic understanding unit, and a command standardization unit; The speech recognition unit is used to convert the input speech signal into text commands using an acoustic model and a language model based on a deep neural network. The semantic understanding unit is used to perform named entity recognition and dependency parsing on the text instructions, and to identify core action entities, region entities and modifier entities in the instructions; The instruction standardization unit is used to map the output of the semantic understanding unit into a structured instruction data object, which includes an instruction type field, a target region list field, and a demand intensity coefficient field.

3. The voice-controlled intelligent water-saving irrigation control system according to claim 2, characterized in that, The instruction standardization unit is also used to: when the regional entity identified by the semantic understanding unit is a fuzzy regional representation, query the system configuration and learning module for all configured irrigation regions, and fill the list of identifiers of these regions into the target region list field; The demand intensity coefficient field is a floating-point number between 0 and 1. Its initial value is mapped and assigned according to the identified modifier entities. When no modifier entity is identified in the instruction, the coefficient is assigned a default value of 0.

5.

4. The voice-controlled intelligent water-saving irrigation control system according to claim 1, characterized in that, The water-saving optimization objective function aims to minimize the total irrigation water consumption, maximize irrigation uniformity, and minimize the deviation from the user demand intensity coefficient. The constraints include pipeline water pressure constraints, solenoid valve switching frequency constraints, soil moisture content not exceeding field capacity constraints, and single irrigation duration not exceeding the maximum allowable duration. The irrigation strategy optimization engine adopts a rolling optimization strategy. Its model predictive control has a prediction time domain of 6 control cycles and a control time domain of 3 control cycles. In each optimization cycle, the irrigation strategy optimization engine only executes the instructions of the first control cycle in the control time domain and recalculates the optimization in the next cycle.

5. The voice-controlled intelligent water-saving irrigation control system according to claim 1, characterized in that, The irrigation execution control module includes an instruction sequence parsing unit, a drive unit, and a status feedback unit; The instruction sequence parsing unit is used to receive the precise irrigation instruction sequence and decompose it into independent low-level control commands according to time slices and regions; The drive unit is used to control the opening and closing of the solenoid valve on the corresponding irrigation branch through a relay circuit or industrial bus according to the underlying control command, and to adjust the output flow of the water pump through a frequency converter or pressure regulating valve. The status feedback unit is used to monitor the on / off status of each solenoid valve, pipeline pressure, and water pump operating current in real time, and package this status information to feed back to the system configuration and learning module.

6. The voice-controlled intelligent water-saving irrigation control system according to claim 1, characterized in that, The system configuration and learning module includes a parameter management unit and a model adaptation unit; The parameter management unit is used to provide a human-computer interaction interface for users or administrators to set and update system parameters such as crop type, growth stage, mapping relationship between irrigation area and valve, upper and lower limits of soil moisture content, and maximum allowable irrigation duration. The model adaptive unit is used to continuously collect historical irrigation data and uses a reinforcement learning algorithm to fine-tune the key parameters in the crop water requirement law model, using the changes in soil moisture and crop growth after actual irrigation as reward signals, and to optimize the weight coefficients in the water-saving optimization objective function.

7. A voice-controlled intelligent water-saving irrigation control system according to claim 6, characterized in that, The reinforcement learning algorithm used by the model adaptive unit is the near-end policy optimization algorithm; The update cycle of the model adaptive unit is set to perform a model parameter update once after every 30 irrigation operations are completed.

Images