Intelligent greenhouse management method and system based on multiple sensors

Abstract

The invention provides an intelligent greenhouse management method and system based on multiple sensors, and relates to the technical field of agricultural environment monitoring and automatic regulation and control. According to the method, multiple types of sensor nodes including air parameters, soil parameters, illumination parameters and seedling growth parameters are arranged in a greenhouse, multi-dimensional acquisition is performed on environment states, unified time synchronization and formatting processing is performed on acquired data, an event mapping sequence is constructed when a system detects a farming operation event, and the event mapping sequence is subjected to data synchronization and formatting processing. A response curve is generated by calculating the change slope, amplitude and time delay characteristic quantity of environmental parameters before and after events, and an environmental standard response template is formed based on the statistical result of multiple events. Furthermore, the system obtains deviation information through comparison of response characteristics and a standard template, automatic diagnosis of the state of the sensor and the operation state of the execution mechanism is achieved, the system generates a control instruction based on the model to complete dynamic execution, and strategy closed-loop updating is conducted according to execution feedback.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural environmental control technology, and in particular to a smart greenhouse management method and system based on multiple sensors. Background Technology

[0002] With the large-scale development of facility agriculture, greenhouse environmental control is gradually shifting from single-parameter adjustment to multi-parameter coordinated regulation. Existing smart greenhouses generally use multiple types of sensors to monitor air temperature and humidity, light intensity, carbon dioxide concentration, and soil parameters in real time, and drive actuators such as ventilation, irrigation, and supplemental lighting based on threshold rules or simple feedback models. Therefore, current technology has some shortcomings: First, existing greenhouse systems typically treat external operations (such as irrigation, opening windows, rolling up curtains, fertilization, etc.) as independent events, without establishing a causal mapping relationship between changes in environmental parameters before and after the events. This makes it difficult for the system to distinguish between "natural environmental changes" and "equipment-driven changes," thus making it difficult to form a precise control basis applicable to sensitive stages such as seed and seedling cultivation.

[0003] Secondly, because the multidimensional parameters collected at different times and under different environments have strong nonlinearity and strong coupling, traditional filtering, interpolation or steady-state calibration methods cannot describe the dynamic response laws such as the rate of humidity increase after the irrigation event is triggered and the slope of temperature decrease after the ventilation event is triggered. This makes the system increasingly dependent on manual adjustment and makes it difficult to improve the degree of automation.

[0004] Therefore, we propose a smart greenhouse management method and system based on multiple sensors.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a smart greenhouse management method and system based on multiple sensors, thereby solving the technical problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A smart greenhouse management method based on multiple sensors includes the following steps: S1. Deploy multiple types of environmental sensor nodes in the smart greenhouse to collect environmental parameters; record the operation time, operation type and operation duration of agricultural operations such as irrigation, ventilation, curtain rolling, and supplemental lighting in real time in the greenhouse control system to form a sequence of agricultural operation events that correspond one-to-one with the environmental parameters. S2. Based on the sequence of agricultural operation events, extract the environmental change trajectory within the corresponding time window from the real-time data collected by multiple sensor nodes; construct the environmental response curve under the action of each agricultural operation event according to the sensor type. S3. Statistically model the environmental response curves corresponding to various types of agricultural operation events to form a standard response template for describing the environmental change pattern under normal operating conditions; when subsequent agricultural operations occur, compare the actual environmental response curve with the standard response template to extract response deviation information. S4. Based on the response deviation information, determine whether there are abnormal states such as no response, slow response, or drift response in the multi-sensor nodes; if the sensor response is normal, further identify the operating deviation of the actuator based on the spatial distribution characteristics of the response deviation. S5. For sensor nodes with offset trends, perform online self-calibration of zero point and slope based on response deviation information; input each response deviation and the corresponding environmental control results into the reinforcement learning model, so that the reinforcement learning model updates the control strategy according to the reward function, and obtains a multi-parameter adaptive control model.

[0008] S6. Based on a multi-parameter adaptive control model, it automatically generates control commands for irrigation, ventilation, curtain rolling, and supplemental lighting and executes them in real time. After execution, it continuously receives environmental feedback information, stores the feedback results and control strategies accordingly, and realizes closed-loop updates between sensor status, electrical execution status and environmental changes to ensure long-term stability and precise optimization of the seed and seedling cultivation environment.

[0009] S1 specifically includes: deploying multiple types of environmental sensor nodes within the seed and seedling cultivation greenhouse. These nodes collect environmental parameters such as canopy air temperature, canopy air humidity, light intensity, root zone soil temperature, root zone soil humidity, and root zone soil pH. A real-time acquisition channel for each environmental parameter is established via a gateway or main control terminal, ensuring all environmental data is uploaded to the greenhouse control system according to a fixed sampling period and a unified timestamp. Within the greenhouse control system, agricultural operations such as irrigation, ventilation, curtain rolling, and supplemental lighting are managed as events, recording the operation type, trigger time, and duration for each operation. The time series of environmental parameters are linked to the agricultural operation events according to their timestamps, forming a mapping sequence of agricultural operation events and environmental data for subsequent analysis. This mapping sequence is stored as input data in the system data cache for use in the environmental response curve construction step, enabling subsequent steps to accurately extract the environmental change trajectory before and after the agricultural operation.

[0010] S2 specifically includes: in the mapping sequence, for each agricultural operation event, determining the corresponding analysis time window based on its operation trigger time; extracting continuous data points of multiple environmental parameters from the analysis time window, forming a parameter data trajectory according to the sampling time order; calculating the slope, amplitude, and response delay of environmental changes based on the parameter data trajectory, and constructing an environmental response curve corresponding to the agricultural operation event; classifying the curves according to the parameter type (temperature, humidity, light, soil), so that each type of sensor corresponds to at least one set of response curves for an agricultural operation event; storing all response curve sets as a "raw response curve library" to provide basic data for the subsequent generation of standard response templates.

[0011] S3 specifically includes: performing statistical analysis on the original response curve library, calculating the average slope, average amplitude, and average delay of each response curve under the same type of agricultural operation event, and obtaining standard response model parameters; constructing a standard response template to describe the normal operating environment change pattern based on the standard response model parameters; when a new agricultural operation event occurs, extracting the real-time environmental response curve of the event from the mapping sequence; comparing the real-time environmental response curve with the standard response template item by item, calculating the response amplitude deviation, slope deviation, and delay deviation; outputting the deviation results as response deviation information and storing it in the system deviation data area to provide input for subsequent sensor state discrimination steps.

[0012] S4 specifically includes: reading the response deviation information and determining whether the deviation of each environmental parameter exceeds the preset normal deviation threshold range; when the response amplitude deviation, slope deviation, or time delay deviation of a certain sensor continues to exceed the limit, the sensor is determined to be in a state of no response, slow response, or drift response; when all sensor response states are normal, the actual operating effect of the actuator is analyzed based on the spatial distribution of deviation information among different sensor nodes; when the response deviation information shows significant spatial non-uniformity, it is determined that the irrigation system may have local drip irrigation blockage, or that the vent or roller shutter mechanism has insufficient opening or incomplete execution operation deviation; the sensor state discrimination results and the actuator operation deviation identification results are summarized into a "system operation deviation report" to provide training data for the reinforcement learning model.

[0013] S5 specifically includes: for sensors determined to have a drift trend, calculating self-calibration compensation amounts based on response amplitude deviation and slope deviation, including zero-point compensation and slope compensation; applying the compensation amounts to the measurement model of the corresponding sensor node to achieve online self-calibration of the sensor, so that the collected data reconciles with the changing trend of the standard response template; inputting the deviation changes before and after sensor self-calibration and the control effects of each agricultural operation into the reinforcement learning model as a state, action, and reward triplet; updating the environmental control strategy based on the reward function of the reinforcement learning model, so that the multi-parameter control process can automatically optimize the control intensity and duration as the deviation changes; and saving the updated control strategy as a multi-parameter adaptive control model for subsequent dynamic execution control steps.

[0014] S6 specifically includes: calling the multi-parameter adaptive control model to generate environmental control commands such as irrigation, ventilation, curtain rolling, and supplemental lighting based on the seed and seedling cultivation stage and the current environmental state; issuing the environmental control commands to the corresponding actuators so that the actuators can perform actions according to the commands; collecting real-time environmental parameter changes after execution to form a mapping sequence of "agricultural operation events and environmental data" as closed-loop feedback information; storing the closed-loop feedback information in correspondence with the current control strategy and inputting it as a new training sample into the reinforcement learning model to continuously optimize the control strategy; completing a closed-loop update so that the system can synchronously maintain the accuracy of sensor status, the reliability of actuators, and the effectiveness of environmental control during long-term operation, thereby ensuring a stable growth environment throughout the entire seed and seedling cultivation process.

[0015] The beneficial effects of this invention are as follows: This invention extracts the dynamic change sequence of multiple types of environmental parameters before and after agricultural operation events, and calculates the response slope, amplitude, and time delay characteristics. This invention can establish an event-driven environmental response model, distinguish between changes caused by natural fluctuations and those caused by equipment triggers, and enable the system to obtain a highly reliable causal relationship understanding in complex coupled environments, which is something that traditional steady-state monitoring systems cannot achieve.

[0016] This invention forms a standard response template by learning the statistical characteristics of multiple similar events, enabling the system to normalize and compare multidimensional environmental responses under different seasons, climates, or crop stages. This provides a unified scale for subsequent deviation diagnosis, sensor calibration, and actuator status identification, solving the problem of incomparable and unreusable dynamic processes in existing greenhouse systems.

[0017] This invention can not only identify the drift, sluggishness or no response state of the sensor from the degree of deviation of the response curve from the standard template, but also identify the inaccurate opening, slow action or regional failure of the actuator according to the spatial distribution characteristics of the deviation, so as to realize the whole-link health diagnosis of the "greenhouse sensing layer + execution layer" and greatly reduce the accumulation of hidden faults in long-term operation.

[0018] This invention utilizes real-time deviation information to automatically calculate the zero-point compensation and slope compensation coefficient of the sensor, generating calibrated data input. This allows the sensor to maintain high accuracy during long-term operation without manual disassembly and calibration, solving the pain points of high sensor maintenance frequency and high labor costs in traditional greenhouse systems.

[0019] This invention incorporates calibrated environmental parameters, execution actions, and feedback rewards into a reinforcement learning model. This invention enables the system to continuously update its control strategy based on different crops, different growth stages, and different external disturbances, achieving multi-scenario adaptive capabilities across seasons and varieties. It breaks through the limitations of traditional threshold control and empirical rule control, which cannot dynamically evolve.

[0020] This invention reconstructs the environmental response mapping by analyzing the feedback information from the execution results, and uses it to update the strategy model. This enables the system to complete parameter self-updating and strategy iteration after each execution cycle, forming a complete closed-loop intelligent control system. This significantly improves the stability, accuracy, and response speed of regulation, and is especially suitable for application scenarios that are extremely sensitive to environmental changes, such as seed and seedling cultivation. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a smart greenhouse management method based on multiple sensors according to the present invention; Figure 2 This is a schematic diagram of the framework of a smart greenhouse management system based on multiple sensors according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: As Figure 1 As shown, this embodiment provides a smart greenhouse management method based on multiple sensors, including the following steps: S1. Environmental data acquisition and agricultural operation event recording steps: Multiple types of environmental sensor nodes are deployed in the smart greenhouse to collect environmental parameters such as canopy air temperature, air humidity, light intensity, root zone soil temperature, soil moisture and pH value; and the operation time, operation type and operation duration of agricultural operations such as irrigation, ventilation, curtain rolling, and supplemental lighting are recorded in real time in the greenhouse control system to form a sequence of agricultural operation events that correspond one-to-one with the environmental parameters.

[0024] S2. Environmental response curve construction steps: Based on the agricultural operation event sequence, extract the environmental change trajectory within the corresponding time window from the real-time acquisition data of the multi-sensor nodes; and construct the environmental response curve under the action of each agricultural operation event according to the sensor type, wherein the environmental response curve includes the slope, change amplitude and response delay of the environmental parameters over time.

[0025] S3. Response template generation and deviation information extraction steps: Statistically model the environmental response curves corresponding to various types of agricultural operation events to form a standard response template for describing the environmental change pattern under normal operating conditions; and when subsequent agricultural operations occur, compare the actual environmental response curve with the standard response template to extract response deviation information, including response amplitude deviation, slope deviation, and time delay deviation.

[0026] S4. Sensor status judgment and actuator operation deviation identification steps: Based on the response deviation information, determine whether there are abnormal states such as no response, slow response or drift response of the multi-sensor nodes; if the sensor response is normal, further identify the actuator operation deviation based on the spatial distribution characteristics of the response deviation, including abnormal working conditions such as drip irrigation branch blockage, insufficient actual opening degree of the vent or incomplete roller shutter execution.

[0027] S5. Optimization steps of self-calibration and control model based on reinforcement learning: For sensor nodes with offset trends, perform online self-calibration of zero point and slope based on the response deviation information; and input the response deviations and corresponding environmental control results into the reinforcement learning model, so that the reinforcement learning model updates the control strategy according to the reward function, thereby obtaining a multi-parameter adaptive control model for the entire seed and seedling cultivation process.

[0028] S6. Dynamic execution control and closed-loop update steps: Based on the multi-parameter adaptive control model, control commands such as irrigation, ventilation, curtain rolling, and supplemental lighting are automatically generated and executed in real time; after execution, environmental feedback information is continuously received, and the feedback results are stored in correspondence with the control strategy to realize closed-loop update between sensor status, electrical execution status and environmental changes, so as to ensure the long-term stability and precise optimization of the seed and seedling cultivation environment.

[0029] S1 specifically includes the following sub-steps: S110. Deployment and Data Acquisition Baseline Definition Steps for Multiple Types of Environmental Sensor Nodes: Deploy multiple types of environmental sensor nodes in the seed and seedling cultivation greenhouse, and set the data acquisition baseline positions according to environmental data acquisition needs, including: : Indicates the average height of the current seedlings (unit: cm), obtained through manual measurement or optical ranging; : Indicates the installation distance between the canopy sensor and the seedling canopy (unit: cm), and its value ranges from 2 to 5 cm; : This indicates the installation height of the canopy air environment sensor (unit: cm), and its calculation formula is:

[0030] : Indicates the burial depth of the root zone soil sensor (unit: cm), with a value range of 3–8 cm. This range covers the growth area of ​​the main root system of the seedlings, and the average plant height of the seedlings. The height can be obtained through manual measurement with a ruler, ultrasonic ranging sensor measurement, or a combination of a top camera and a simple height estimation algorithm. This method of obtaining height can be implemented by ordinary technicians.

[0031] Install canopy air temperature sensors, canopy air humidity sensors, and light intensity sensors at the above locations, as well as bury root zone soil temperature sensors, root zone soil humidity sensors, and root zone soil pH sensors, so that various sensors can obtain representative growth environment parameters.

[0032] To ensure the representativeness of environmental parameters, the deployment of various sensor nodes should follow the principle of uniform coverage by zone. This means dividing the greenhouse into several grid units along both the length and width directions, and placing at least one canopy air sensor node and one root zone soil sensor node within each grid unit to guarantee comprehensive spatial sampling. The root zone soil sensor should be buried at a depth of 3-8 cm; this range was chosen because it covers the main distribution layer of most seed and seedling root systems, making the collection of soil temperature, humidity, and pH data more representative.

[0033] Various types of sensor nodes can be powered by batteries or DC power and interact with the gateway via wireless communication (such as LoRa, BLE or Wi-Fi).

[0034] S120. Environmental Data Acquisition Channel Establishment and Unified Timestamp Marking Steps: To ensure the comparability of all sensor data, real-time acquisition channels for various environmental parameters are established through gateway devices, and all uploaded data is marked with a unified timestamp, including: t: Represents the acquisition time (unit: seconds), provided by the system clock; : Indicates the sampling period (unit: seconds), and its value ranges from 10 to 120 seconds.

[0035] The system operates according to a cycle. All sensor nodes collect data and store it according to timestamp t, so that air temperature, air humidity, light intensity, soil temperature, soil humidity and pH value at any time can form a strictly aligned data sequence. The unified timestamp can be generated by the gateway's system clock and synchronized with the NTP network time or local time synchronization device periodically to ensure that the data uploaded by different sensors are aligned under the same time base.

[0036] S130. Event-based recording steps for agricultural operations: All manually or automatically executed agricultural operations are recorded as events within the greenhouse control system, including: : Represents the i-th agricultural operation event; This indicates the type of agricultural operation, including irrigation, ventilation, rolling shutters, supplemental lighting, etc. : Indicates an event Trigger time (unit: seconds); : Indicates the duration of the event (unit: seconds).

[0037] The type and trigger time of each agricultural operation Duration The data is recorded as structured event data, providing an accurate event reference time base for subsequent response curve construction steps.

[0038] Agricultural operation events can be recorded in a structure format, with fields including: event number, event type, trigger time, duration, and execution intensity parameters (such as vent opening angle, irrigation flow rate level, etc.). This structure format is a commonly used data organization method.

[0039] S140, Steps for binding environmental parameter sequences with agricultural operation events: In order to build a mapping relationship between agricultural operation events and environmental changes, environmental parameter sequences are bound to agricultural operation events according to timestamps; in : Represents the set of environmental parameters (including temperature, humidity, light intensity, soil parameters, etc.) collected at time t; : Indicates an event The corresponding "agricultural operation event-environment data" mapping set.

[0040] The mapping relationship is defined as follows:

[0041] in: : Indicates the duration of the observation window before the event is triggered (in seconds), with a value range of 60–180 seconds; : Indicates the observation duration of the response after the event is triggered (unit: seconds), and its value ranges from 300 to 900 seconds; the above range provides a clear method for ordinary technicians to extract environmental change data before and after the event.

[0042] To ensure data integrity within the analysis time window, the system generates... Previously, the data time series was checked. When there were packet loss or missing tests at certain time points, linear interpolation could be used to fill in the missing points to ensure that the curve construction process was continuous and reproducible.

[0043] S150, Data storage and subsequent retrieval preparation steps for the mapping sequence: For each of the above mapping sets... The data is stored in the system's data cache, where: : Represents the data structure used to store all mapping sequences; : Represents the index number of each mapping sequence.

[0044] Storage rules are defined as follows:

[0045] All mapping sequences are stored in chronological order and used for subsequent environmental response curve construction steps. This enables the system to accurately construct environmental response curves based on each agricultural operation, ensuring that the input for subsequent algorithm processing is complete, traceable, and reproducible.

[0046] Data cache It can be organized using a time-series table or a key-value index structure, each... Each event corresponds to a specific time window, enabling ordinary technicians to directly retrieve its corresponding mapping sequence based on the order of events or event number.

[0047] S2 specifically includes the following sub-steps: S210, Steps for extracting the response analysis time window: from the mapping set Extract the analysis time window, where: The i-th agricultural operation event is already defined. :event The trigger time (in seconds) is already defined; The length of the observation window before the event is triggered (in seconds), which ranges from 60 to 180 seconds and is already defined; The length of the response window after the event is triggered (in seconds), which ranges from 300 to 900 seconds and is already defined; :event The analysis time window (second range). The analysis window is defined as:

[0048] At this point, the principles, physical significance, and scope of time window selection have been fully disclosed, facilitating reproduction by those skilled in the art. The analysis time window is then determined. Afterwards, the system performs an integrity check on each sampling time point within the window. When data is missing at certain sampling time points, linear interpolation or nearest neighbor interpolation can be used to fill in the missing points to ensure the continuity of the time series and make subsequent response curve calculations stably reproducible.

[0049] To ensure that the data density within the time window is sufficient to form an analyzable curve, the sampling period... With window length The following conditions must be met: That is, the window must contain at least 5 sampling points so that the response curve can show an effective trend. This ratio is public and no specific sampling value is specified.

[0050] S220, Steps for constructing time series trajectories of environmental parameters: From Extract environmental parameter values ​​within the time window to construct a time series trajectory, where: Environmental parameters of type j (such as air temperature, air humidity, light intensity, soil temperature, soil humidity, pH, etc.) collected at time t, with units corresponding to the measured quantities; j: environmental parameter category index, satisfying... ,in Indicates the total number of parameters (greater than 6); : and the event The time series trajectory of the corresponding j-th type of environmental parameter. The time series trajectory is defined as:

[0051] This step clarifies the structure of the trajectory, the meaning of the parameters, and the range of values, and the disclosure is complete and sufficient.

[0052] In constructing time series trajectories At that time, the system automatically sorts all sampling points in ascending order according to the timestamp t to ensure that any... Strict temporal relationships are maintained to ensure the accuracy of calculations for features such as slope and amplitude.

[0053] To reduce the impact of environmental noise on the calculation of response curve characteristic quantities, time series can be analyzed. First-order moving average filtering, three-point smoothing, or Savitzky-Golay smoothing are optional processing methods.

[0054] S230, Calculation steps for environmental response curve characteristic quantities: For the time series trajectory... Calculate three types of response characteristics: response slope, response amplitude, and response delay. Formula for calculating response slope:

[0055] Formula for calculating response amplitude:

[0056] Response delay calculation formula:

[0057] in: : The j-th type of environmental parameter in the event The slope of the response; : The j-th type of environmental parameter in the event The response amplitude below; : The j-th type of environmental parameter in the event The response latency is as follows; : The time (in seconds) at which the curve of the j-th type of environmental parameter begins to change; : Any two time points used to calculate the slope of the response, and satisfying And all belong to .

[0058] and It can be selected based on adjacent data points in the sequence, for example, by taking adjacent sampling times. , Alternatively, a longer time interval can be used to estimate the overall trend. Ordinary technicians can choose an appropriate time interval based on the rate of change of the response curve.

[0059] The This can be determined by judging when the slope of the parameter curve first exceeds a preset change threshold, i.e., when the following conditions are met:

[0060] At that time, the corresponding time Recorded as .in The threshold used to detect the initial change of the curve can be obtained by estimating the noise amplitude, and no specific value is limited.

[0061] S240. Response curve classification step according to environmental parameter type: Based on the physical properties of each type of parameter, the response characteristic quantities are classified, and the classification set is defined as follows: : A set of air-related features (air temperature, air humidity); : Set of illumination features; : A set of soil characteristics (soil temperature, soil moisture, soil pH); :event The response feature classification set. Where:

[0062] Each category contains three feature quantities corresponding to the parameters.

[0063] S250, Steps for constructing the raw response curve library: To be used for subsequent standard response template generation, the categorized response curves are stored in the raw response curve library, where: Three characteristic quantities.

[0064] S250, Steps for constructing the raw response curve library: To be used for subsequent standard response template generation, the categorized response curves are stored in the raw response curve library, where: : Represents the data structure of the original response curve library; : Indicates an event The index of the response curve for the j-th parameter. The storage rule is:

[0065] in Stored in chronological and parameter category order, allowing ordinary technicians to directly call and reproduce the data without additional knowledge.

[0066] Original response curve library Common data structures such as key-value index tables, two-dimensional matrices, hash tables, or time series tables can be used to organize the data. Each record contains an event number, parameter category j, and corresponding feature value. This enables those skilled in the art to quickly retrieve all response information for a specific agricultural operation based on the event number.

[0067] S3 specifically includes the following sub-steps: Before performing statistical analysis on multiple agricultural operation events, the system ensures the sample size. satisfy Ideally, at least five trials should be conducted to ensure the statistical mean is representative. This sample size requirement is standard practice in this field when estimating the mean.

[0068] S310. Statistical analysis steps for multi-event response characteristics: To construct a standard response template, analyze the original response curve library. Statistical analysis was conducted on multiple agricultural operation events. Before performing statistical analysis, the system can perform anomaly detection on event samples, such as detecting anomalies in the response magnitude of a certain event. Values ​​exceeding three times the standard deviation of the mean of all samples can be considered outliers and removed to enhance the stability of the statistical results. The statistical calculation formula is as follows: Average response slope:

[0069] Average response amplitude:

[0070] Average response time

[0071] in: : The i-th agricultural operation event, which is already defined; j: The index of the j-th type of environmental parameter; :event The response slope of the j-th type of environmental parameter is already defined; :event The response magnitude of the j-th type of environmental parameter is already defined; :event The response delay of the j-th type of environmental parameter is already defined; : The number of events used for statistical analysis (must be ≥3 times to ensure statistical stability); : The average response slope of the j-th type of parameter; : The average response magnitude of the j-th type of parameter; : The average response delay of the j-th type of parameter.

[0072] Among them, the average response slope The unit is consistent with the unit of parameter change divided by time; average response amplitude The units are consistent with the original parameters; average response delay The unit is seconds. Consistency in units enables ordinary technicians to accurately perform the calculation of characteristic quantities.

[0073] The system chooses the mean as the standard feature because the same type of agricultural events are repeatable under similar environmental conditions. The mean can describe the typical response trend of the event, and those skilled in the art can reproduce the standard pattern of environmental change based on this.

[0074] S320, Construction steps of the standard response template: Construct the standard response template based on the average characteristic values ​​obtained in step S310, wherein: : The data structure representing the standard response template; : Represents the standard response template entry for the j-th type of parameter. The standard response template is defined as:

[0075] The template means the following: : Indicates the typical rate of increase / decrease caused by the event; : Indicates the typical range of environmental changes caused by the event; : Indicates the timing characteristics of the typical response start.

[0076] This template is used in subsequent steps such as "deviation judgment," "anomaly identification," and "self-calibration." Standard Response Template It can be organized in the form of a key-value table or a multidimensional array, where the key is the parameter category j and the value is the corresponding feature set. This data structure is easily implemented in software by those skilled in the art.

[0077] S330, Steps for extracting real-time environmental response curves: When a new agricultural operation event occurs... When this occurs, repeat the process in S210–S250 to obtain the real-time response characteristics: Real-time response slope; Real-time response amplitude; Real-time response latency. Also defined: The deviation of the real-time response slope from the standard value; The deviation between the real-time response amplitude and the standard value; The deviation between the real-time response latency and the standard value; this step ensures that the data input is sufficient and can automatically connect to S340.

[0078] To ensure real-time response characteristics In accordance with the calculation method of template features, the system adopts the same time window, sorting strategy and smoothing method as S210–S250, thereby ensuring that the feature quantities have consistent comparability.

[0079] S340. Comparison steps between real-time response and standard response template: To identify abnormal situations, the real-time response features are compared item by item with the standard response template: Response slope deviation: ; Response amplitude deviation: ; Response delay deviation: ; The positive and negative meanings of deviation: This indicates that the response speed is faster than normal. This indicates a slow response time, which may be due to congestion or insufficient execution. This indicates that the magnitude of environmental change is relatively weak, which may indicate that the actions taken at the execution end are insufficient. This indicates a response delay, which may be due to sensor drift or execution lag.

[0080] S350, Structured Output Steps for Deviation Information: The above three deviation quantities are uniformly organized into structured deviation information for use in subsequent sensor state determination steps, wherein: : Indicates an event The deviation information structure; : Indicates an event The deviation entries for the j-th type of parameter. The structured representation is as follows:

[0081] All deviation entries ultimately form the overall deviation information:

[0082] The structure is clearly defined, its implementation is well-defined, and its data types are consistent.

[0083] Deviation Information Structure It can be organized as a JSON object, dictionary structure, or table format, with fields including parameter category, response slope deviation, amplitude deviation, and time delay deviation, making it directly usable as input for the discrimination steps of S410–S450 sensors. The specific data structure format does not constitute a limitation on the scope of protection.

[0084] S4 specifically includes the following sub-steps: S410, Sensor Deviation Threshold Reading and Preliminary Status Screening Steps: To determine whether the sensor has an abnormal response, read the deviation thresholds corresponding to various environmental parameters from the system parameter library, including: : Response slope deviation threshold for the j-th type of parameter (unit: same) ); : Response amplitude deviation threshold for the j-th type of parameter (unit: same) ); The time delay deviation threshold for the j-th type of parameter (unit: same) ); : The j-th type of parameter in the event The deviation entries in the text are already defined; :from The three types of deviations extracted from them have already been defined.

[0085] Deviation threshold The threshold can be obtained by calculating the standard deviation or percentiles of characteristic values ​​from historical normal event samples. For example, two or three times the standard deviation can be selected as the threshold, or the 95th percentile can be selected as the threshold range. This method of obtaining the threshold is a statistical method that can be directly implemented by ordinary technicians and does not limit the specific threshold value.

[0086] The system performs a preliminary screening based on the following conditions: If This indicates an abnormal response speed; if This indicates an abnormal response amplitude; if This indicates an abnormal response delay.

[0087] Only the threshold sign and logic need to be disclosed here; the specific numerical range does not need to be specified. The aforementioned deviation... The units are consistent with the corresponding response slope, response amplitude, and response delay, making the judgment conditions clear and directly reproducible.

[0088] When determining sensor status, the system can prioritize a multi-indicator joint judgment strategy. When multiple deviations of a parameter exceed a threshold, the confidence level of its corresponding status is higher. For example, simultaneously satisfying... and When this occurs, it can be preferentially identified as a drift state.

[0089] S420. Refined determination steps for sensor anomaly types: Determine the specific anomaly type of the sensor based on the deviation pattern, as defined below: : The status label of the sensor corresponding to the j-th type of parameter; "Normal": all deviations have not exceeded the threshold; "No Response": the amplitude deviation is significantly zero or minimal; "Slow Response": response delay deviation. Significant delay; "drift response": both the slope deviation and amplitude deviation of the response show a systematic shift.

[0090] The judgment rules are as follows: No response state (the sensor barely reacts): ; Slow response state: ; When response delay deviation Positive and exceeding the threshold When the time shift occurs, it indicates that the start time of parameter change is significantly later than the normal response template, and therefore it can be determined as a slow response state. This determination is based on time offset characteristics, which is easy for ordinary technicians to understand and implement.

[0091] Drift response status: ; Drift differs from normal slight fluctuations in that it is characterized by a persistent, unidirectional shift in both slope and amplitude deviations; that is, the overall trend of the data shows a systematic difference from the normal response curve. Ordinary technicians can distinguish drift from random noise based on this trend characteristic.

[0092] Normal state: If none of the above conditions are met, then: ; The mathematical basis and judgment objectives for each state are publicly available and clearly defined, allowing ordinary technicians to implement them directly. The amplitude approaching zero can be determined using the following conditions: ;in This is a small threshold estimated from the sensor noise amplitude, for example, 1–2 times the historical noise range. This threshold is only used to explain the detection method for no-response states and does not limit the specific value.

[0093] S430, Spatial Distribution Characteristics Analysis Steps under Normal Sensor Conditions: When all sensor conditions satisfy:

[0094] At this point, the deviation information does not originate from the sensor itself, but may come from differences in the operation of the actuator. Therefore, we define: The spatial coordinates (in meters) of the j-th sensor were set during the installation of the greenhouse. :event The following is a set of spatial distributions of all sensor biases:

[0095] The system analyzes the locality of deviation based on spatial distribution: if the deviation is concentrated in a certain area (e.g., the same...) If the deviation is within the specified area, it may be a regional failure of the actuator; if the deviation shows a continuously decreasing distribution, it may be a weak response caused by insufficient execution.

[0096] The region can be obtained by dividing the area based on the physical location of the sensors during deployment. For example, a greenhouse can be divided into several spatial grids, with each sensor within a grid considered a local area. This region division method is a common spatial segmentation approach used to help determine the spatial concentration of deviations.

[0097] S440, Steps for identifying actuator operating deviation types: Based on The spatial distribution characteristics are used to identify operational deviations of the actuator. The definition is as follows: :event The operating status label of the actuator; The types of actuators include: irrigation branches, ventilation openings, and roller shutter mechanisms; the types of deviations include: blockage, insufficient opening, and incomplete execution.

[0098] The effective range of actuators (such as irrigation, ventilation, and roller shutters) is defined during system configuration. For example, a certain irrigation branch corresponds to a specific planting row, and a certain ventilation opening affects the air environment of a specific area. Based on this range mapping relationship, when the deviation is concentrated in the corresponding area, the deviation can be associated with a specific actuator.

[0099] The identification rules are as follows: Irrigation branch blockage: If soil parameters (such as humidity) satisfy the following within a specific spatial region R: ;

[0100] The judgment is as follows:

[0101] Insufficient ventilation: If the air temperature / humidity in the area near a certain ventilation opening meets the following conditions:

[0102] but:

[0103] Incomplete roller blind operation: If the response amplitude of lighting parameters is too low:

[0104] And if it is located in the area affected by the roller shutter, then:

[0105] Each of the above judgments can be made directly by a person skilled in the art based on the deviation information.

[0106] S450, Steps for generating a system operation deviation report: Transfer sensor status labels With the status label of the executing agency Combined to generate a deviation report, where: :event Deviation report structure; The definition is as follows:

[0107] The bias report serves as input data for subsequent reinforcement learning model training and self-calibration strategies in S510–S550. This step provides a complete output structure, making the entire S410–S450 process a fully open and closed-loop process.

[0108] Deviation Report

[0109] The report can be organized using JSON, a dictionary, a table, or a key-value structure. Its fields include parameter categories, status labels, and actuator operating status, allowing the report to be directly used as input to the reinforcement learning module (S510–S550). This structure is a common approach in software implementation and does not constitute a limitation on the scope of protection.

[0110] S5 specifically includes the following sub-steps: S510, Step 1: Calculate self-calibration compensation based on deviation information: Since many types of sensors are susceptible to temperature drift, humidity adsorption or aging during long-term operation, resulting in a deviation between the response slope and the output amplitude, the system needs to automatically calculate the compensation amount to perform self-calibration to ensure that the data acquired by the subsequent reinforcement learning module is accurate and reliable.

[0111] To correct for sensors exhibiting a tendency to drift, a self-calibration compensation amount is calculated using the deviation. The definition is as follows: : Zero-point calibration compensation amount for the j-th type of sensor (units consistent with the corresponding sensor dimensions); : Slope calibration compensation coefficient of the j-th type sensor (dimensionless); , The slope deviation and amplitude deviation of the response from S350 are already defined; The scaling factor for the compensation coefficient (dimensionless, its range is unlimited, but must be >0). To ensure feasibility, the self-calibration formula is disclosed as follows: Zero-point compensation amount:

[0112] Note: When the amplitude deviation At that time, the sensor output is too low, and the zero-point calibration is positive; when At that time, the sensor output is too high, and the zero-point calibration is negative.

[0113] Slope compensation coefficient:

[0114] Note: If the response slope is too slow ,but The slope increases; If the response slope is too fast ,but The slope decreases. Compensation factor and The values ​​can be obtained through historical data fitting, empirical parameter tuning, or simple grid search. They are not limited to fixed constants and can be dynamically adjusted according to different sensor types. This selection method is one that can be directly implemented by ordinary technicians.

[0115] To ensure the numerical stability of the compensation calculation process, the calculated values ​​can be... and Set upper and lower limits, such as limiting them to ±10% of the sensor's range, to avoid overcalibration caused by abnormal deviations.

[0116] S520. Application steps of self-calibration compensation in sensor measurement model: To ensure real-time calibration of the sensor output, a compensation amount is applied to the original measurement model. Definition: : The raw measurement value of the j-th type sensor at time t; : Measurement values ​​after self-calibration; , The compensation amount from S510 is already defined.

[0117] The output value after self-calibration is defined as:

[0118] Note: Zero-point compensation is applied first, followed by slope compensation; the calibrated output will be used as input to the subsequent control model and reinforcement learning module. (Self-calibrated output value) The original measurements will be uniformly replaced. It serves as the input for all subsequent modules (including feature extraction, state construction, and control strategy execution), thereby ensuring the consistency of the data chain within the system.

[0119] S530. Steps for constructing the state, action, and reward dataset for a reinforcement learning model: To enable the reinforcement learning model to update its control policy through experience, it is necessary to construct (state, action, reward) triples. Defined as follows: The system's state vector at time t, including the calibrated sensor output. Current status of the implementing agency, etc.; The control actions taken by the system at time t, such as irrigation start duration, ventilation start angle, and roller shutter descent range; The reward value generated based on the control effect reflects the degree of environmental improvement; : Reinforcement learning experience set (Experience Buffer); Enew: The currently processed agricultural operation event. State definition:

[0120] The principles for defining returns are as follows:

[0121] The evaluation function is an achievable function, the specific form of which can be selected according to the greenhouse management needs. A linear weighted method can be used, for example: ; in Weights can be set by ordinary technical personnel according to application needs; there are no fixed restrictions on the specific format. Experience storage method: .

[0122] To avoid experience set Unlimited growth, with a maximum capacity that can be set. When the capacity exceeds the limit, the oldest recorded sample can be removed first, thus achieving rolling updates. This is a common implementation method for reinforcement learning experience pools.

[0123] State vector In addition to including the calibrated environmental parameter values, it may also include information such as the actuator's opening percentage, the target environment setting, and the current time period (e.g., morning, noon, or night), enabling the reinforcement learning model to make decisions from multiple dimensions. The above state composition is only a possible example and does not limit the actual fields.

[0124] S540, Steps for updating control policies based on reinforcement learning: Utilizing experience sets Update the policy of the reinforcement learning model. Definition: : Parameters are The control strategy function (dimensionless); Learnable parameters of the policy function; Learning rate (dimensionless).

[0125] The updated formula is as follows: .in The mathematical expectation operator, Indicates about parameters The gradient operator; when performing gradient updates, gradient clipping can be used to limit the norm of the gradient within a preset threshold to avoid model instability due to excessively large gradients during training. When the reward value stabilizes or the change amplitude falls below the set threshold after several consecutive policy updates, the policy is considered to have converged. This type of convergence criterion is a common method in reinforcement learning training and does not limit the specific implementation.

[0126] S550, Steps for forming a multi-parameter adaptive control model: After the strategy update is completed, the system solidifies the latest strategy into a real-time executable environmental control model. Definition: The final generated multi-parameter adaptive control model; The policy function with the latest parameters. The regulation model is defined as:

[0127] illustrate: This will serve as the decision-making basis for subsequent dynamic execution control steps of S610–S650; control model The model can be stored in the controller as a matrix, weight tensor, or parameter configuration file, and invoked by the controller each time an instruction is generated, enabling the model to run in real time on the greenhouse equipment. The specific storage format of the model does not constitute a limitation on the scope of protection; all processes are open and transparent, and do not involve any algorithmic black boxes.

[0128] S610, Steps for generating environmental control commands based on the adaptive control model: Utilizing the multi-parameter adaptive control model generated by S550. It generates execution control instructions based on the current environment state. The definitions are as follows: : Multi-parameter adaptive control model, from S550; control model It can be invoked once at a fixed period (such as every 10 seconds, 30 seconds or 1 minute), or it can be triggered immediately when a rapid change in the environment is detected. This invocation strategy can be selected by ordinary technicians according to actual needs.

[0129] The system's environmental state vector at time t, including calibrated sensor data. Information such as the operational status of the implementing agency comes from S530; The control action output by the system at time t; action The value can be limited to the allowable output range of the equipment. For example, the opening degree of the vent can be controlled between 0 and 100%, and the switching of the irrigation pump can be represented by discrete actions. This range limit is used to ensure that the action can be performed, but does not restrict the specific form of the control model.

[0130] : A set of control commands, including irrigation commands, ventilation commands, curtain commands, supplemental lighting commands, etc.

[0131] Control actions are defined as:

[0132] The actions are then mapped into directly executable control commands:

[0133] in Action-command mapping functions are used to convert policy actions into device commands, such as "ventilator opening angle 30%" or "irrigation duration 20 seconds". Mapping function It can be a linear mapping, a piecewise function, or a lookup table function. For example, when the action represents a target opening percentage, the mapping function can convert it into the PWM duty cycle or motor pulse count required by the device driver. The above mapping methods are standard engineering implementations and are not limited to a specific format.

[0134] S620, Steps for Issuing and Executing Actuator Actions: Based on the set of control commands generated in step S610. The corresponding instructions are then issued to the actual implementing agency. The definition is as follows: The implementing agency in time The system is a set of execution actions; the execution of instructions by the actuator usually involves physical response time, such as the acceleration of a motor, the opening of a valve, or the start-up time of a pump. The system can record the actual start and end times of execution so that subsequent feedback acquisition steps can be correctly aligned with the execution completion time.

[0135] The actuators include: irrigation pumps, vent motors, roller shutter motors, and supplemental lighting drivers.

[0136] The execution action is defined as:

[0137] in These are device drive functions used to translate control commands into actual mechanical actions (such as rotating a motor or adjusting flow). When the device fails to execute a command (e.g., abnormal drive current or position feedback error), the system can generate an error flag and incorporate this result into the feedback data, enabling the reinforcement learning module to correct the abnormal behavior. This error handling process is a device management method familiar to general technicians.

[0138] S630, Execution Feedback Data Acquisition Step: After the actuator completes its action, the system collects environmental feedback data again and constructs a new mapping sequence. The definition is as follows: : Feedback value of the j-th type of environmental parameter after the action is executed (unit is the same as the original parameter); : Feedback mapping sequence after action execution; Feedback collection time window (unit: seconds).

[0139] The window is defined as follows:

[0140] in This represents the feedback observation time, ranging from 60 to 300 seconds (without narrowing the protection range). The feedback sequence is defined as:

[0141] The feedback collection logic is complete and unambiguous.

[0142] To ensure that the feedback data accurately reflects the effectiveness of the actuator, a feedback sampling window is used. length The time can be set according to the typical action time of the actuator. For example, soil moisture changes can be observed 1–2 minutes after irrigation, and air temperature and humidity trends can be observed 30–60 seconds after ventilation.

[0143] The system sorts the feedback data by timestamp and can selectively perform smoothing processing (such as moving average) to reduce the impact of sensor noise on subsequent reward calculations. This type of data preprocessing method is conventional and does not limit the specific algorithm form.

[0144] S640. Steps for updating data records based on execution feedback: To achieve multi-round closed-loop learning, the feedback results are transformed into data usable by the reinforcement learning model. Defined as follows: Reward value generated based on feedback data (reward after execution); : Enhanced learning experience set (already defined); : The new state vector after feedback.

[0145] The reward calculation principles are as follows:

[0146] in It is an implementable function, with no specific form restrictions, but it must generate positive or negative returns based on the degree of improvement in environmental deviation.

[0147] Feedback reward value The following examples can be used:

[0148] in These are weighting coefficients, which can be set by those skilled in the art according to the specific crop growth logic. This function form is merely an example and does not limit the method of calculating returns in this invention.

[0149] The new state vector is defined as:

[0150] : The feedback value of the j-th type of environmental parameter after self-calibration; Final recorded dataset:

[0151] This structure is consistent with typical reinforcement learning (RL), is publicly complete, and is available when the experience set is available. When the maximum capacity is reached, the system can use a "first-in, first-out (FIFO)" or "random replacement" strategy to delete some old samples to ensure the stability of the data structure and long-term operation.

[0152] S650. Steps for establishing a long-term closed-loop update mechanism: To ensure continuous optimization of the control effect during long-term system operation, the model is updated in a closed loop. The definition is as follows: The updated regulatory model; Strategy model parameters; Learning rate (already defined).

[0153] Based on experience set Execution strategy update:

[0154] The updated model is defined as follows:

[0155] The updated control model is used to generate the next control action for S610, enabling the system to form a continuous, self-learning, and self-optimizing long-term closed-loop control process. The old model can be automatically replaced after the strategy update cycle ends. Replacement may be performed only when a return enhancement threshold is met, to ensure the stability of field execution. The model replacement strategy can be selected by those skilled in the art as needed.

[0156] Example 2: Figure 2 As shown, this embodiment provides a smart greenhouse management system based on multiple sensors, including: Multiple types of environmental sensing modules are used to collect various environmental parameters in the greenhouse, such as air temperature and humidity, light intensity, carbon dioxide concentration, seedling height, and soil temperature, humidity, and pH in the root zone. The modules also perform time synchronization, formatting, and unified identification of the data collected by each sensor node. The event detection and mapping module is used to record the event type, trigger time and execution parameters when an agricultural operation event is detected, and to extract environmental parameter data within a preset time window before and after the event, and to construct the event mapping sequence and the corresponding environmental parameter time series trajectory. The environmental response feature extraction module is used to calculate the response slope, response amplitude and response delay features of environmental parameters based on the time series, and to generate a standard response template based on the statistical results of multiple similar events. The deviation discrimination module is used to compare the real-time response characteristics with the standard response template to obtain the slope deviation, amplitude deviation and time delay deviation of various parameters, and to judge the response status of multiple types of sensors based on this, and to identify the operating status of the actuator based on the spatial distribution of the deviation. The self-calibration compensation module is used to calculate the zero-point compensation amount and slope compensation coefficient of multiple types of sensors based on the deviation characteristics, and to perform self-calibration processing on the original sensor output to obtain a unified calibrated environmental parameter input. The reinforcement learning decision module is used to construct a state vector based on calibrated environmental parameters, actuator operating status and preset control targets, make decisions on control actions and calculate rewards, and iteratively update the strategy model through experience set to form a multi-parameter adaptive control strategy model. The execution control and feedback module is used to generate control commands for actuators such as irrigation, ventilation, rolling shutters, and supplemental lighting based on the control strategy model, issue execution actions and monitor operation, and collect environmental feedback data after execution to form a closed-loop update data source. The strategy closed-loop update module is used to update the reinforcement learning model based on execution feedback, generate a new control strategy model, and replace it in the next control cycle, so that the system has the ability to continuously learn and adaptively control.

[0157] The modules are connected sequentially via data links, enabling the system to achieve a complete closed-loop management process including multi-parameter environmental perception, event-driven environmental response learning, sensor self-calibration, deviation diagnosis, reinforcement learning strategy optimization, and dynamic environmental control.

[0158] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0159] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0160] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart greenhouse management method based on multiple sensors, characterized in that, Includes the following steps: S1. Deploy multiple types of environmental sensor nodes in the smart greenhouse to collect environmental parameters; record the operation time, operation type and operation duration of irrigation, ventilation, curtain rolling and supplemental lighting in real time in the greenhouse control system to form a sequence of agricultural operation events that correspond one-to-one with the environmental parameters. S2. Based on the sequence of agricultural operation events, extract the environmental change trajectory within the corresponding time window from the real-time data collected by multiple sensor nodes; construct the environmental response curve under the action of each agricultural operation event according to the sensor type. S3. Statistically model the environmental response curves corresponding to various types of agricultural operation events to form a standard response template for describing the environmental change pattern under normal operating conditions; when subsequent agricultural operations occur, compare the actual environmental response curve with the standard response template to extract response deviation information. S4. Based on the response deviation information, determine whether the multi-sensor nodes have abnormal states such as no response, slow response, or drift response. If the sensor response is normal, the operating deviation of the actuator can be further identified based on the spatial distribution characteristics of the response deviation. S5. For sensor nodes with offset trends, perform online self-calibration of zero point and slope based on response deviation information; input each response deviation and the corresponding environmental control results into the reinforcement learning model, so that the reinforcement learning model updates the control strategy according to the reward function, and obtains a multi-parameter adaptive control model.

2. The intelligent greenhouse management method based on multiple sensors according to claim 1, characterized in that, Also includes: S6. Based on a multi-parameter adaptive control model, it automatically generates and executes control commands for irrigation, ventilation, curtains, and supplemental lighting in real time. After execution, it continuously receives environmental feedback information, stores the feedback results in correspondence with the control strategy, realizes closed-loop updates between sensor status, electrical execution status and environmental changes, and ensures long-term stability and precise optimization of the seed and seedling cultivation environment.

3. The intelligent greenhouse management method based on multiple sensors according to claim 1, characterized in that, S1 specifically includes: Multiple types of environmental sensor nodes are deployed in the seed and seedling cultivation greenhouse. These multiple types of environmental sensor nodes are used to collect environmental parameters such as canopy air temperature, canopy air humidity, light intensity, root zone soil temperature, root zone soil humidity, and root zone soil pH. Establish real-time acquisition channels for various environmental parameters through gateways or main control terminals, so that all environmental data are uploaded to the greenhouse control system according to a fixed sampling period and a unified timestamp; In the greenhouse control system, irrigation, ventilation, curtain rolling, and supplemental lighting operations are managed through event-based management, recording the operation type, trigger time, and duration of each operation.

4. The intelligent greenhouse management method based on multiple sensors according to claim 1, characterized in that, S1 specifically includes: The time series of the environmental parameters are bound to the agricultural operation events according to the timestamps to form an agricultural operation event and environmental data mapping sequence for subsequent analysis; The mapping sequence is stored as an input dataset in the system data cache for use in the environmental response curve construction step, enabling subsequent steps to accurately extract the environmental change trajectory before and after agricultural operations.

5. The intelligent greenhouse management method based on multiple sensors according to claim 1, characterized in that, S2 specifically includes: In the mapping sequence, for each agricultural operation event, the corresponding analysis time window is determined based on its operation trigger time; Continuous data points of multiple environmental parameters are extracted from the analysis time window, and parameter data trajectories are formed according to the sampling time order; Based on the parameter data trajectory, the slope, magnitude, and response delay of environmental changes are calculated, and environmental response curves corresponding to agricultural operation events are constructed. The curves are classified according to the parameter type of the response curve, so that each type of sensor corresponds to at least one set of response curves for agricultural operation events. All response curve sets are stored as a raw response curve library to provide basic data for the subsequent generation of standard response templates.

6. The intelligent greenhouse management method based on multiple sensors according to claim 1, characterized in that, S3 specifically includes: Statistical analysis was performed on the original response curve library to calculate the average slope, average amplitude, and average delay of each response curve under the same type of agricultural operation event, and the standard response model parameters were obtained. Based on the parameters of the standard response model, a standard response template is constructed to describe the changing patterns of the normal operating environment; When a new agricultural operation event occurs, the real-time environmental response curve of the event is extracted from the mapping sequence; The real-time environmental response curve is compared with the standard response template item by item, and the response amplitude deviation, slope deviation and time delay deviation are calculated. The deviation results are output as response deviation information, and the response deviation information is stored in the system deviation data area.

7. The intelligent greenhouse management method based on multiple sensors according to claim 1, characterized in that, S4 specifically includes: Read the response deviation information and determine whether the deviation of each environmental parameter exceeds the preset normal deviation threshold range; When the response amplitude deviation, slope deviation, or time delay deviation of a sensor continues to exceed the limit, the sensor is judged to be in a state of no response, slow response, or drift response. Under the condition that all sensor response statuses are normal, the actual operating effect of the actuator is analyzed based on the spatial distribution of deviation information among different sensor nodes. When the response deviation information shows significant spatial non-uniformity, it is determined that there may be local drip irrigation blockage in the irrigation system, or that there is insufficient opening or incomplete execution of the ventilation opening or roller shutter mechanism. The sensor status judgment results and the actuator operation deviation identification results are summarized into a system operation deviation report.

8. The intelligent greenhouse management method based on multiple sensors according to claim 1, characterized in that, S5 specifically includes: For sensors determined to have a drift trend, the self-calibration compensation amount is calculated based on the response amplitude deviation and slope deviation, including the zero-point compensation amount and the slope compensation amount. The compensation amount is applied to the measurement model of the corresponding sensor node to achieve online self-calibration of the sensor, so that the collected data conforms to the changing trend of the standard response template again. The changes in the deviation of the sensor before and after self-calibration and the control effects of each agricultural operation are input into the reinforcement learning model as a triplet of state, action, and reward. Based on the reward function of the reinforcement learning model, the environmental control strategy is updated so that the multi-parameter regulation process can automatically optimize the regulation intensity and duration as the deviation changes. The updated control strategy is saved as a multi-parameter adaptive control model.

9. A smart greenhouse management method based on multiple sensors according to claim 2, characterized in that, S6 specifically includes: The multi-parameter adaptive control model is invoked to generate environmental control commands for irrigation, ventilation, curtain rolling, and supplemental lighting based on the seed and seedling cultivation stage and the current environmental status. The environmental control command is sent to the corresponding actuator, so that the actuator performs the action according to the command. The real-time environmental parameter changes after the collection and execution are used to form a sequence of agricultural operation events and environmental data mapping, which serves as closed-loop feedback information. The closed-loop feedback information is stored in correspondence with the current control strategy and used as a new training sample to input into the reinforcement learning model to continuously optimize the control strategy. A closed-loop update is completed, enabling the system to maintain the accuracy of sensor status, the reliability of actuators, and the effectiveness of environmental control during long-term operation, thus ensuring a stable growth environment throughout the entire seed and seedling cultivation process.

10. A smart greenhouse management system based on multiple sensors, employing the smart greenhouse management method based on multiple sensors as described in any one of claims 1-9, characterized in that, include: Multiple types of environmental sensing modules are used to collect various environmental parameters in the greenhouse, such as air temperature and humidity, light intensity, carbon dioxide concentration, seedling height, and root zone soil temperature, humidity, and pH. The event detection and mapping module is used to record the event type, trigger time and execution parameters when an agricultural operation event is detected, and to extract environmental parameter data within a preset time window before and after the event. The environmental response feature extraction module is used to calculate the response slope, response amplitude, and response delay features of environmental parameters based on time series, and to generate standard response templates based on the statistical results of multiple similar events. The deviation discrimination module is used to compare the real-time response characteristics with the standard response template to obtain the slope deviation, amplitude deviation and time delay deviation of various parameters. The self-calibration compensation module is used to calculate the zero-point compensation amount and slope compensation coefficient of multiple types of sensors based on the deviation characteristics, and to perform self-calibration processing on the original sensor output. The enhanced learning decision-making module is used to construct a state vector based on calibrated environmental parameters, actuator operating status and preset control targets, make decisions on control actions and calculate rewards; The execution control and feedback module is used to generate control commands for irrigation, ventilation, curtain rolling, and supplemental lighting actuators based on the regulation strategy model, and to issue and monitor the execution actions.

Images