Automatic plant maintenance method and device based on multi-AI model collaboration

Abstract

The invention provides an automatic plant maintenance method and device based on multi-AI model collaboration, belongs to the technical field of intelligent plant maintenance, and solves the technical problem that the intelligent result of automatic maintenance is inaccurate due to lack of precise data support in existing plant maintenance. The method comprises the following steps: performing parallel processing on acquired data by utilizing a multi-AI model; the multi-AI model comprises a plurality of sub-models and a structured cooperative processing module; the sub-model comprises a plant health status evaluation sub-model; a disease identification sub-model; a pest recognition sub-model; an environment suitability analysis sub-model; a plant growth state stage identification sub-model; generating a unified data contract by using an information summarization formatting method; performing environment conflict detection, forming an intelligent maintenance executable strategy, converting an equipment instruction, executing automatic control and supervision, performing periodic monitoring management and adaptive adjustment, and outputting an intelligent maintenance result report. The system is suitable for automatic plant maintenance based on computer vision, the Internet of Things technology and the automatic control technology.

Description

Technical Field

[0001] This invention belongs to the field of intelligent plant maintenance technology, specifically involving the field of intelligent automatic plant maintenance based on multi-AI model collaboration, and is used for automatic plant maintenance using computer vision, Internet of Things technology and automatic control technology. Background Technology

[0002] Plant care is a complex task that faces the following technical challenges: It is difficult to coordinate the control of multiple environmental factors, and the control of a single factor cannot meet the comprehensive growth needs of plants. Traditional threshold or timing control strategies lack intelligence and are difficult to make decisions based on the actual state of the plant. Different plant varieties and growth stages have different needs, and a uniform parameter scheme cannot achieve individualized maintenance. Multiple environmental factors may conflict, and the existing system lacks an effective coordination mechanism. It mainly relies on sensor data, lacks visual analysis of plant appearance characteristics, and anomalies are not detected in a timely manner; The user interaction is highly professional, making it difficult for ordinary users to understand and handle abnormal situations.

[0003] Existing technologies typically employ single-function control or simple threshold judgment for plant care. For example, the "Intelligent Plant Care System" disclosed in Chinese Patent CN106774556A, while capable of basic automatic watering, suffers from significant shortcomings in handling multi-factor collaborative control and intelligent decision-making. It cannot effectively address multi-factor conflicts such as the need for simultaneous cooling while supplemental lighting increases temperature. Furthermore, most existing systems lack visual monitoring capabilities, resulting in a lack of timely response to plant diseases, pests, and changes in growth status. While the "Plant Growth Status Monitoring System Based on Image Visual Analysis" disclosed in Chinese Patent CN117058607A incorporates AI technology, it primarily focuses on the application of a single model, lacking collaborative analysis of multiple AI models and information aggregation and formatting mechanisms. This leads to unstable decision-making and difficulties in strategy execution when facing complex plant care scenarios. Summary of the Invention

[0004] In view of this, the present invention aims to propose a method and device for automatic plant maintenance based on the collaboration of multiple AI models, so as to solve the technical problem that the lack of precise data support in existing plant maintenance leads to inaccurate results of automatic maintenance.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention proposes an automated plant maintenance method based on multi-AI model collaboration, the method comprising: S1. Acquire historical data and collect new data using video surveillance and sensors; the new data includes visual data and sensor data; S2. The newly collected data is processed in parallel using multiple AI models to obtain a structured dataset; the multiple AI models include multiple sub-models and a structured collaborative processing module; the sub-models include: A sub-model for assessing plant health status, used to identify features in plant leaves; A disease identification sub-model is used for the location, classification, and detection of plant diseases in specific areas. A pest identification sub-model is used for pest target location and classification identification; An environmental suitability analysis sub-model is used for nonlinear modeling and dynamic weight adjustment of multiple environmental factors; A sub-model for identifying plant growth stages, used to identify growth stages based on the current morphological characteristics of the plant; S3. Process the structured dataset using information summarization and formatting methods to generate a unified data contract; S4. Generate an intelligent maintenance strategy based on the unified data contract and historical data; S5. Based on the intelligent maintenance strategy, perform environmental conflict detection to form an executable intelligent maintenance strategy; S6. Convert the intelligent maintenance executable strategy into equipment instructions, use the equipment instructions to perform automatic control and supervision, collect execution readback data and generate a receipt; S7. Based on the execution readback data and the intelligent maintenance executable strategy, perform periodic monitoring and management, make adaptive adjustments according to the results of the periodic monitoring and management, and output an intelligent maintenance result report.

[0006] Furthermore, the visual data mentioned in S1 is plant image data; the sensor data includes soil moisture, air humidity, ambient temperature, soil temperature, light intensity, soil nutrients, electrical conductivity, and pH value.

[0007] Furthermore, in the multi-AI model described in S2, The plant health status assessment sub-model uses a ResNet-50 convolutional neural network architecture to identify plant characteristics, and the output data includes health score, confidence level, risk factor list and recommended measures; The disease identification sub-model uses the YOLOv8 target detection framework and combines it with the EfficientNet feature extractor to locate and classify disease areas in plants. The output data includes disease type, detection box coordinates, detection confidence, severity, affected area ratio, overall health status, and treatment urgency. The pest identification sub-model uses the Faster R-CNN two-stage detection framework combined with the ResNeXt-101 backbone network to locate and classify pest targets on plants. The output data includes pest type, developmental stage, detection box coordinates, detection confidence, individual number estimation, activity level, pest severity, affected parts, risk level, detection time and recommended monitoring frequency. The environmental suitability analysis sub-model uses a multilayer perceptron to identify multiple environmental factors, including climate factors, soil factors, and physical factors. The output data includes a comprehensive suitability score and optimization suggestions. The comprehensive suitability score includes a light intensity suitability score, a soil nutrient content suitability score, and a wind speed suitability score. The plant growth stage identification sub-model uses a ResNet-50 convolutional neural network combined with multi-scale feature fusion to identify the morphological features of plants. The output data includes: current growth stage, stage determination confidence, progress within the current stage, key indicators, and development rate assessment.

[0008] Furthermore, the information aggregation and formatting method described in S3 includes data confidence labeling and redundancy removal.

[0009] Furthermore, the intelligent maintenance strategy described in S4 is adaptively adjusted based on the historical execution records of the historical data and the environment, execution, and effects.

[0010] Furthermore, the environmental conflict detection method described in S5 includes: detecting conflict factors; determining the priority of the conflict factors; sorting them according to the priority of life safety and growth stage; adopting a time-sharing plan and determining whether the safety boundary is exceeded; if so, executing a noise reduction strategy; and forming an execution plan.

[0011] Furthermore, the intelligent maintenance strategies described in S5 include maintenance timing, maintenance intensity, duration, and execution cycle.

[0012] Furthermore, the automatic control described in S6 sends control commands, including heating and cooling, irrigation, supplemental lighting, and fertilization, to the device equipped with the sensor via a message queue and a local bus.

[0013] This invention also proposes an automatic plant maintenance device based on multi-AI model collaboration, the device comprising: The data acquisition module is used to acquire historical data and also to acquire new data using video surveillance and sensors; the new data includes visual data and sensor data. A model processing module is used to process newly collected data in parallel using multiple AI models to obtain a structured dataset. The multiple AI models include multiple sub-models and a structured collaborative processing module. The sub-models include: The sub-models are: Plant Health Status Assessment Sub-model for identifying characteristics of plant leaves; Disease Identification Sub-model for locating and classifying plant diseases; Pest Identification Sub-model for locating and classifying pest targets; Environmental Suitability Analysis Sub-model for nonlinear modeling and dynamic weight adjustment of multiple environmental factors; and Plant Growth Stage Identification Sub-model for identifying growth stages based on the current morphological characteristics of the plant. A data formatting module is used to process the structured dataset using information summarization and formatting methods to generate a unified data contract. Initial strategy module; used to generate intelligent maintenance strategies based on the unified data contract and historical data; The execution strategy module is used to perform environmental conflict detection based on the intelligent maintenance strategy and form an executable intelligent maintenance strategy. The control and supervision module is used to convert the intelligent maintenance executable strategy into equipment instructions, use the equipment instructions to perform automatic control and supervision, collect execution readback data and generate a receipt. The strategy adjustment module is used to perform periodic monitoring and management based on the execution readback data and the intelligent maintenance executable strategy, and to make adaptive adjustments according to the results of the periodic monitoring and management, and output an intelligent maintenance result report.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention presents an automated plant maintenance method based on multi-AI model collaboration. Compared to existing methods that are limited in function, have low intelligence, lack intelligent maintenance strategies, poor environmental conflict handling capabilities, insufficient visual monitoring, and high user barriers, this method integrates multi-source data, analyzes it collaboratively using multiple AI models, and generates executable strategies. This method acquires plant appearance and environmental data through cameras and various sensors. These data are then analyzed in parallel by models for health assessment, disease identification, pest identification, environmental suitability, and growth stage identification. An information aggregation and formatting module structures the output, forming a unified data contract. A strategy engine generates parameterized strategies for watering, supplemental lighting, temperature control, and fertilization based on this contract and historical execution results. When multiple factors conflict, a priority, time-sharing, and compromise balancing approach is used to obtain an executable solution. The strategies are automatically executed and controlled, and the results are recorded and reviewed. The monitoring results are used to adaptively adjust the monitoring and execution frequency and generate reports. This method achieves collaborative and closed-loop control of multiple environmental factors, lowers the barrier to entry, and improves maintenance reliability and ease of use.

[0015] The present invention discloses an automated plant maintenance method and device based on multi-AI model collaboration. In terms of multi-model collaborative decision-making, expert models such as pest and disease identification, health assessment, and environmental analysis output in parallel. After information aggregation and formatting, the output is input into the strategy engine, reducing the uncertainty of large model output and making the strategy more stable and usable.

[0016] In terms of personalized maintenance strategies, strategies are adaptively generated based on plant species, current growth status, and historical data, which better meet individual needs compared to fixed thresholds or timed control.

[0017] In terms of multi-factor collaborative control, factors such as water, temperature, light, and nutrients are optimized in a coordinated manner. When factor conflicts are encountered, an executable balance solution can be obtained through priority and time-sharing / compromise mechanisms.

[0018] In terms of vision + sensor fusion, timed visual inspections combined with continuous sensor sampling can detect problems such as abnormal leaf color, wilting, disease spots, and insect traces earlier.

[0019] In terms of closed-loop automatic execution, a closed loop of "strategy-instruction-execution-monitoring-reporting" is formed, and the execution records are traceable, which facilitates continuous optimization and auditing.

[0020] In terms of ease of use, actionable suggestions and troubleshooting guidelines can be pushed through WeChat / APP, allowing ordinary users to complete the process.

[0021] In terms of adaptive scheduling, the monitoring and execution frequency is dynamically adjusted based on the effect evaluation to reduce invalid actions and improve resource utilization efficiency.

[0022] In terms of modular expansion, the model, sensor, and actuator adopt a unified data contract, which facilitates rapid expansion and replacement according to plant category or application scenario.

[0023] In terms of safety and fault tolerance, it includes sensor self-checking, abnormal redundancy, emergency stop and manual takeover processes to reduce the risk of erroneous execution.

[0024] In terms of knowledge accumulation, the structured accumulation of environmental and execution data supports long-term strategy iteration and multi-category migration. Attached Figure Description

[0025] Figure 1 This is a flowchart of an automated plant maintenance method based on multi-AI model collaboration as described in this invention.

[0026] Figure 2 This is a flowchart of the data acquisition process described in this invention.

[0027] Figure 3 This is a flowchart of the multi-environmental factor control method described in this invention.

[0028] Figure 4This is a schematic diagram illustrating the collaborative processing of the disease identification sub-model and the pest identification sub-model described in this invention.

[0029] Figure 5 The flowchart for generating a unified data contract as described in this invention is provided.

[0030] Figure 6 This is a schematic diagram of the information aggregation and formatting method described in this invention.

[0031] Figure 7 This is a schematic diagram of the fields of the unified data contract described in this invention.

[0032] Figure 8 This is a flowchart of the environmental conflict detection process described in this invention.

[0033] Figure 9 This is a schematic diagram illustrating the process of automatically controlling and supervising the execution of the intelligent maintenance executable strategy until a report is generated.

[0034] Figure 10 This is a flowchart of the automatic control and supervision process described in this invention.

[0035] Figure 11 The present invention utilizes the device instructions to perform automatic control and supervision, collects execution readback data, and generates a receipt flowchart.

[0036] Figure 12 This is a diagram showing the topology of the device and sensor equipped with the sensor described in this invention.

[0037] Figure 13 This is the adaptive adjustment flowchart described in this invention.

[0038] Figure 14 This is a schematic diagram of the alarm notification push of the security mechanism described in this invention.

[0039] Figure 15 This is a flowchart of the exception handling mechanism in the security mechanism described in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.

[0041] Specific Implementation Method 1: This implementation method describes an automated plant maintenance method based on multi-AI model collaboration, such as... Figure 1 As shown, the method includes: S1. Acquire historical data and collect new data using video surveillance and sensors; the new data includes visual data and sensor data; S2. The newly collected data is processed in parallel using multiple AI models to obtain a structured dataset; the multiple AI models include multiple sub-models and a structured collaborative processing module; the sub-models include: A sub-model for assessing plant health status, used to identify features in plant leaves; A disease identification sub-model is used for the location, classification, and detection of plant diseases in specific areas. A pest identification sub-model is used for pest target location and classification identification; An environmental suitability analysis sub-model is used for nonlinear modeling and dynamic weight adjustment of multiple environmental factors; A sub-model for identifying plant growth stages, used to identify growth stages based on the current morphological characteristics of the plant; S3. Process the structured dataset using information summarization and formatting methods to generate a unified data contract; S4. Generate an intelligent maintenance strategy based on the unified data contract and historical data; S5. Based on the intelligent maintenance strategy, perform environmental conflict detection to form an executable intelligent maintenance strategy; S6. Convert the intelligent maintenance executable strategy into equipment instructions, use the equipment instructions to perform automatic control and supervision, collect execution readback data and generate a receipt; S7. Based on the execution readback data and the intelligent maintenance executable strategy, perform periodic monitoring and management, make adaptive adjustments according to the results of the periodic monitoring and management, and output an intelligent maintenance result report.

[0042] It also includes S8. Security Mechanism; when steps S1 to S7 are executed, the security mechanism continuously performs anomaly detection, and when an anomaly is detected, the anomaly handling mechanism of the security mechanism is triggered; the anomaly handling mechanism includes emergency stop procedure, sensor anomaly detection, data anomaly detection, alarm notification and automatic recovery.

[0043] In this embodiment, the method begins with step one, in which historical data is acquired and new data is collected using video surveillance and sensors; the new data includes visual data and sensor data. Figure 2As shown, the vision + sensor fusion and strategy generation method demonstrates the workflow relationship between multi-source data acquisition, AI collaborative analysis, information aggregation and formatting, strategy generation, conflict resolution, and automatic execution. The visual data is plant image data; the sensor data includes soil moisture, air humidity, ambient temperature, soil temperature, light intensity, soil nutrients, electrical conductivity, and pH value. Specifically, the acquired data is multi-source data. Through the acquisition of visual and sensor data, it can be deployed in a collaborative environment of local controller and cloud services, with functions including: Image input function; timed camera captures and uploads plant images; users can supplement uploaded images via WeChat / APP; the system performs preprocessing, including noise reduction, cropping, and format conversion; Sensor acquisition function; acquires plant images and environmental and soil-related sensor data; collects soil moisture, air humidity, temperature, light intensity, soil nutrients, electrical conductivity, pH status, etc.; uploads data with timestamps and device identification. The collected data is multi-source data, which includes at least three of the following: soil moisture, air humidity, ambient temperature, soil temperature, light intensity, soil nutrients, electrical conductivity, and pH status.

[0044] Plant Image Acquisition and Recognition Process: 1) Image Acquisition: Timed camera captures plant images; user-uploaded images are used as needed; appropriate lens distance and lighting conditions are maintained to ensure clear leaf and flower / fruit features. 2) Image Upload: Images are uploaded after edge preprocessing; timestamps and device IDs are added. 3) Collaborative Analysis: Health assessment, pest and disease identification, environmental suitability analysis, and growth stage identification models output conclusions in parallel. These conclusions are then formatted into an information summary module, generating a maintenance strategy. Finally, the maintenance strategy is converted into standardized equipment instructions that can be directly executed by various actuators. 4) Information Summary Formatting and Strategy Generation: The information summary formatting module outputs information uniformly, and the strategy generation model generates maintenance strategies based on information from multiple models. 5) Instruction Conversion: The instruction conversion model converts the information generated by the strategy module into instructions that can be executed by various actuators and distributes them. The health assessment, pest and disease identification, environmental suitability analysis, and growth stage identification models output conclusions in parallel; the information summary formatting generates maintenance strategies, which are then converted into standardized equipment instructions that can be directly executed by various actuators. 6) Information aggregation and strategy generation; the information aggregation formatting module provides unified structured output; the strategy engine combines historical data and a knowledge base to generate executable maintenance strategies. Multi-AI model collaborative analysis: parallel identification and evaluation of health status, diseases, pests, environmental suitability, and growth stages; information aggregation formatting: structuring and standardizing the outputs of each model to form a unified data contract; intelligent maintenance strategy generation: generating parameterized strategies for watering, supplemental lighting, temperature control, and fertilization based on the unified data contract and historical execution results; environmental conflict handling: obtaining executable compromise solutions by prioritizing and employing a time-sharing and compromise balancing mechanism; priority order: life safety takes precedence over water, temperature, light, and nutrients. Automatic control: converting strategies into equipment instructions for execution and recording feedback; periodic monitoring management: adaptively adjusting monitoring and execution frequency based on results and generating reports.

[0045] Then proceed to step two, in which, as follows: Figure 3 and Figure 4 As shown, multiple AI models are used to process newly collected data in parallel to obtain a structured dataset; the multiple AI models include multiple sub-models and a structured collaborative processing module; the sub-models include: A sub-model for assessing plant health status, used to identify features in plant leaves; A disease identification sub-model is used for the location, classification, and detection of plant diseases in specific areas. A pest identification sub-model is used for pest target location and classification identification; An environmental suitability analysis sub-model is used for nonlinear modeling and dynamic weight adjustment of multiple environmental factors; A sub-model for identifying plant growth stages is used to identify growth stages based on the current morphological characteristics of the plant.

[0046] In the aforementioned multi-AI model The plant health status assessment sub-model uses a ResNet-50 convolutional neural network architecture to identify plant characteristics, and the output data includes health score, confidence level, risk factor list and recommended measures; The disease identification sub-model uses the YOLOv8 target detection framework and combines it with the EfficientNet feature extractor to locate and classify disease areas in plants. The output data includes disease type, detection box coordinates, detection confidence, severity, affected area ratio, overall health status, and treatment urgency. The pest identification sub-model uses the Faster R-CNN two-stage detection framework combined with the ResNeXt-101 backbone network to locate and classify pest targets on plants. The output data includes pest type, developmental stage, detection box coordinates, detection confidence, individual number estimation, activity level, pest severity, affected parts, risk level, detection time and recommended monitoring frequency. The environmental suitability analysis sub-model uses a multilayer perceptron to identify multiple environmental factors, including climate factors, soil factors, and physical factors. The output data includes a comprehensive suitability score and optimization suggestions. The comprehensive suitability score includes a light intensity suitability score, a soil nutrient content suitability score, and a wind speed suitability score. The plant growth stage identification sub-model uses a ResNet-50 convolutional neural network combined with multi-scale feature fusion to identify the morphological features of plants. The output data includes: current growth stage, stage determination confidence, progress within the current stage, key indicators, and development rate assessment.

[0047] Specifically, Automatic moisture control; 1) Humidity monitoring; Equipped with a high-precision humidity sensor to monitor soil moisture in real time; Sensor data is fed back to the main control board in real time, corresponding to the connection relationship between the controller, sensor, and irrigation actuator. 2) Automatic irrigation control; The main control board determines whether irrigation is needed based on the maintenance plan and real-time humidity data; When the soil moisture is lower than the specified threshold, the water pump is automatically started, and the irrigation volume and frequency are automatically adjusted according to the actual soil moisture and plant water requirements, continuously monitoring the feedback data from the humidity sensor to ensure that the soil moisture is maintained within the ideal range. 3) Safety protection mechanism; Safety mechanism to prevent over-irrigation; Automatically stops irrigation and sends an alarm when abnormal moisture increases; Safety design to prevent water leakage; Automatic Light Control: 1) Light Monitoring: Employs high-precision light sensors to monitor ambient light intensity in real time; monitors the overall light level of natural and artificial light sources; and corresponds to the connection relationships with the controller, sensors, and supplementary lighting actuators. 2) LED Light Source Control: Controls light conditions through an adjustable LED light source system; the main control board automatically adjusts the brightness and duration of the light source according to the plant's light requirements; automatically supplements artificial light sources when natural light is insufficient; and precisely controls the light switching time according to the plant's biological clock requirements. 3) Photoperiod Simulation: Simulates natural photoperiods to promote healthy plant growth; supports different plant requirements for diurnal photoperiods; Automatic Temperature Control: Temperature Monitoring: Monitors ambient temperature in real time through temperature sensors; monitors temperature changes in the plant's growth environment; and corresponds to the connection relationships with the controller, sensors, and heating / cooling actuators. Temperature Regulation Control: Automatically adjusts the heating / cooling system according to the maintenance plan; maintains the optimal temperature range for plant growth; and supports different plant requirements for temperature.

[0048] Automatic fertilizer control: 1) Nutrient monitoring; configuring soil nutrient sensors to monitor soil nutrient content; assessing the current nutrient status in the soil and the connection relationship between the controller, sensors, and fertilizer actuators. 2) Automatic fertilizer application control; applying appropriate fertilizers on time and in the correct amount according to the maintenance plan; automatically adjusting fertilizer type, application amount, and application frequency; combining soil nutrient sensor data to avoid over-fertilization or nutrient deficiency; dynamically adjusting fertilization strategies based on plant growth status.

[0049] The multiple AI models call multiple sub-models in parallel, namely expert models: plant health status assessment, disease identification, pest identification, environmental suitability analysis, plant growth stage identification, etc., and output conclusions and confidence levels.

[0050] 1. Health assessment model Model Architecture: An improved ResNet-50 convolutional neural network architecture is adopted, combined with an attention mechanism to enhance the ability to recognize detailed features of plant leaves. Input Data Format: Image data: RGB format, recommended resolution 224×224 pixels; Sensor data: Environmental parameter vectors such as temperature, humidity, light intensity, and soil pH; Training Dataset Composition: Healthy plant samples: ≥10,000 labeled images; Sub-healthy plant samples: ≥8,000 labeled images; Environmental data: Corresponding sensor readings; Expert annotation: Plant health level (1-10 points); Output Data Structure: {"health_score": 8.5, / / Health score (0-10)"confidence": 0.92, / / Confidence level (0-1)"risk_factors": [ / / List of risk factors {"factor": "leaf_yellowing", "severity":0.3}, {"factor": "growth_slow", "severity": 0.1}],"recommendation": "increase_nutrients" / / Recommendations} Health assessment algorithm pseudocode: Algorithm: PlantHealthAssessment;Input: image_data, sensor_data;Output: health_assessment_result / / Data preprocessing and validation if validate_input_data(image_data, sensor_data) == False:returnerror_result() / / Feature extraction and fusion image_features=extract_image_features(image_data); sensor_features=process_sensor_data(sensor_data); combined_features = feature_fusion(image_features, sensor_features) / / Health assessment calculation health_score=calculate_health_score(combined_features);confidence=calculate_confidence(health_score) / / Result Output return format_assessment_result(health_score, confidence) Performance characteristics: Scoring accuracy: Possesses excellent health scoring capabilities; Processing efficiency: Meets real-time assessment requirements; Confidence control: Provides reliable result confidence. 2. Disease Identification Model Model Architecture: An improved YOLOv8 object detection framework, combined with the EfficientNet feature extractor, is used to achieve disease region localization and classification. Disease Classification System: Leaf diseases: including leaf spot, leaf blight, powdery mildew, etc.; Root diseases: including root rot, stem rot, and other root-related diseases; Viral diseases: including mosaic, leaf curl, yellowing, and other viral diseases; Bacterial diseases: including soft rot, bacterial wilt, etc. Input data format: Image data: RGB format, recommended resolution 640×640 pixels; Multi-angle shooting: Supports images from the front, side, back, etc.; Image enhancement: Adaptively adjusts contrast and saturation; Training dataset composition: Disease samples: ≥1,500 labeled images for each disease; Healthy controls: ≥5,000 images of disease-free plants; Labeling format: Bounding box + disease category + severity level; Output data structure: {"detections": [{"disease_type": "leaf_spot", / / disease type"bbox": [x1,y1, x2, y2], / / detection box coordinates"confidence": 0.89, / / detection confidence"severity": "moderate", / / severity"affected_area_ratio": 0.15 / / affected area ratio}],"overall_health": "diseased", / / overall health status"treatment_urgency": "medium", / / treatment urgency"recommended_treatment": ["fungicide_spray", "isolation"]} Pseudocode for disease detection algorithm: Algorithm: DiseaseDetection;Input: plant_image;Output: disease_detection_result / / Input data validation and preprocessing if validate_input_image(plant_image) == False:return error_result() / / Image Preprocessing and Feature Extraction preprocessed_image = preprocess_image(plant_image) feature_maps = extract_detection_features(preprocessed_image) / / Disease Area Detection and Optimization detection_results = detect_disease_regions(feature_maps) filtered_detections = filter_low_confidence(detection_results) final_detections = remove_duplicate_detections(filtered_detections) Severity assessment and treatment recommendations severity_analysis = assess_disease_severity(final_detections) treatment_plan = generate_recommendations(severity_analysis) return format_disease_detection_result(final_detections, treatment_plan) False positive handling mechanism: Confidence filtering: screening detection results based on dynamic thresholds; Multi-frame verification: improving detection stability through multi-frame images; Morphological correction: optimizing results based on plant morphological characteristics; Expert knowledge fusion: logical verification combined with botanical rules; Performance indicators: Detection capability: possessing good disease detection capability; Processing efficiency: meeting real-time detection requirements.

[0051] 3. Pest Identification Model Model Architecture: The Faster R-CNN two-stage detection framework, combined with the ResNeXt-101 backbone network, is adopted to achieve accurate localization and classification of pest targets. Pest Classification System: Foliar Pests: including aphids, spider mites, spider mites, whiteflies, etc.; Boring Pests: including longhorn beetle larvae, stem borers, root-knot nematodes, etc.; Underground Pests: including grubs, cutworms, root aphids, etc.; Flying Pests: including mosquitoes, flies, thrips, whiteflies, etc.; Input Data Format: Image Data: RGB format, recommended resolution 800×600 pixels; Multi-scale Input: Supports microscopic images at different magnifications; Multi-frame Images: Multiple frames of images within a short time are used to improve detection stability; Training Dataset Composition: Pest Samples: Images of each stage of each pest—egg, larva, and adult—with more than 800 labeled images per stage; Labeling Information: Pest type + developmental stage + quantity statistics; Output Data Structure: {"pest_detections": [{"pest_type": "aphid", / / pest type "life_stage": "adult", / / developmental stage "bbox": [x1, y1, x2, y2], / / detection box coordinates "confidence": 0.91, / / detection confidence "count": 15, / / estimated number of individuals "activity_level": "high" / / activity level}],"infestation_severity": "moderate", / / severity of pest infestation "affected_plant_parts": ["leaves", "stems"], / / affected plant parts "risk_level": "medium", / / risk level "detection_timestamp": "2024-01-15T10:30:00Z", / / detection time "control_recommendations": ["biological_control", "targeted_spray"], / / recommended control methods "monitoring_frequency": "daily" / / recommended monitoring frequency} Pest detection algorithm pseudocode: Algorithm: PestDetection;Input: plant_image_sequence;Output: pest_detection_result / / Input data validation and preprocessing if validate_image_sequence(plant_image_sequence) == False: return error_result() / / Multi-frame image processing for each image in plant_image_sequence: enhanced_image = image_enhancement(image) roi_regions = region_proposal_network(enhanced_image) feature_vectors = feature_extraction(roi_regions) classifications = pest_classifier(feature_vectors) refined_boxes = bbox_regression(roi_regions, feature_vectors) / / Multi-frame information fusion and result generation multi_frame_detections = temporal_fusion(all_detections) pest_counts = count_individuals(multi_frame_detections) severity_assessment = evaluate_infestation_level(pest_counts) activity_analysis = assess_current_activity_level(multi_frame_detections) control_strategy = recommend_control_methods(severity_assessment) return format_pest_detection_result(multi_frame_detections, severity_assessment, activity_analysis, control_strategy) False positive handling mechanism: Confidence threshold: Screening of pest detection results based on dynamic threshold; Multi-frame consistency: Verifying the stability of detection results through multi-frame images; Morphological constraints: Logical verification based on insect morphological characteristics; Biological laws: Screening results based on the life cycle patterns of pests; Environmental adaptability: Adjusting detection sensitivity according to environmental conditions; Performance characteristics: Counting accuracy: Capable of accurately counting dense targets; Processing efficiency: Meets the needs of rapid detection; Missed detection control: Possesses good target discovery capabilities.

[0052] 4. Environmental suitability model Model Architecture: A multilayer perceptron (MLP) combined with an attention weight allocation mechanism is used to achieve nonlinear modeling and dynamic weight adjustment of multiple environmental factors. Environmental Factor System: Climate factors: including climatic environmental parameters such as temperature, humidity, light intensity, and CO2 concentration; Soil factors: including soil physicochemical parameters such as pH, EC value, water content, and nutrient content; Physical factors: including physical environmental parameters such as wind speed, air pressure, and soil temperature; Input Data Elements: Sensor data: multidimensional environmental parameter vector; Plant species information: used for differentiated weight allocation; Growth stage information: plant development stage identifier; Output Data Structure: {"overall_suitability": 7.8, / / Overall suitability score (0-10)"confidence": 0.87, / / Assessment confidence"factor_scores": { / / Scores for each factor"temperature": 8.2, / / Environmental temperature suitability score"humidity": 7.5, / / Environmental humidity suitability score"light_intensity": 6.8, / / Light intensity suitability score"co2_concentration": 7.9, / / CO2 concentration suitability score"soil_ph": 9.1, / / Soil pH suitability score"soil_moisture": 8.3, / / Soil moisture suitability score"soil_nutrients": 7.0, / / Soil nutrient content suitability score"soil_ec": 8.5, / / Soil electrical conductivity suitability score"wind_speed": 7.6, / / Wind speed suitability score"soil_temperature":} 8.1 / / Soil temperature suitability score},"limiting_factors": [ / / Limiting factors {"factor": "light_intensity", "severity": "moderate"},{"factor": "nutrients", "severity": "mild"}],"optimization_suggestions": [ / / Optimization suggestions {"action": "increase_light", "priority": high"},{"action": "add_fertilizer", "priority": "medium"}]} Suitability assessment algorithm pseudocode: Algorithm: EnvironmentalSuitabilityAssessment; Input: sensor_data, plant_species, growth_stage; Output: suitability_assessment / / Input data validation and preprocessing if validate_input_data(sensor_data, plant_species, growth_stage) ==False: return error_result() / / Data normalization and feature extraction normalized_data = normalize_sensor_data(sensor_data) species_weights = load_species_specific_weights(plant_species); stage_modifiers = get_growth_stage_modifiers(growth_stage); combined_features = apply_species_stage_weights (normalized_data, species_weights, stage_modifiers) / / Multi-layer perceptron evaluation calculation hidden_layer1 = relu(linear1(combined_features)) attention_weights = softmax(attention_layer(hidden_layer1)) weighted_features = attention_weights * hidden_layer1 hidden_layer2 = relu(linear2(weighted_features)) overall_suitability = sigmoid(output_layer(hidden_layer2)) / / Factor analysis and recommendation generation factor_scores=calculate_individual_factor_scores(normalized_data,species_weights, stage_modifiers); limiting_factors=identify_limiting_factors(factor_scores, threshold=6.0); optimization_suggestions=generate_optimization_suggestions(limiting_factors); confidence=calculate_assessment_confidence(model_uncertainty) / / Output the result: return format_suitability_output(overall_suitability, factor_scores, limiting_factors, optimization_suggestions,confidence) Multi-factor weighting method: Basic weights: Static weights based on plant physiology research; Dynamic adjustment: Weight coefficients are dynamically adjusted according to growth stage; Attention mechanism: Weight allocation of important factors learned by the model; Species specificity: Differentiated weights for different plant species; Suitability scoring criteria: Excellent (8-10 points): All key factors are within the optimal range; Good (6-8 points): Main factors are suitable, a few factors deviate slightly; Average (4-6 points): Some factors are unsuitable and need adjustment; Poor (2-4 points): Multiple key factors are seriously deviated from the optimal range; Very poor (0-2 points): Environmental conditions are seriously unsuitable for plant growth; Dynamic adjustment mechanism: Real-time update: Environmental suitability is reassessed every 12 hours; Threshold alarm: An alarm is triggered when the suitability score is lower than a set threshold; Performance characteristics: Assessment accuracy: Supports high-precision environmental suitability assessment; Response efficiency: Meets the requirements for rapid response.

[0053] 5. Growth stage identification model; Model Architecture: An improved ResNet-50 convolutional neural network combined with multi-scale feature fusion is used to achieve growth stage recognition based on current morphological features. Input Data Format: Image data: RGB format, recommended resolution 512×512 pixels; Morphological parameters: current quantitative measurement data such as plant height, number of leaves, stem diameter, etc.; Environmental data: current environmental condition parameters; Output Data Structure: {"current_stage": "vegetative_growth", / / current growth stage"stage_confidence": 0.94, / / stage confidence"stage_progress": 0.65, / / progress within the current stage (0-1)"key_indicators": { / / key indicators"height_cm": 25.4,"leaf_count": 12,"development_rate": "normal" / / development rate assessment}} Pseudocode for growth stage identification algorithm: Algorithm: GrowthStageRecognition; Input: current_image, morphological_data, environmental_data; Output: growth_stage_assessment / / Input data validation if not validate_input_data(current_image, morphological_data): return error_result("Invalid input data") / / Image Feature Extraction spatial_features = extract_spatial_features(current_image); morphological_features=extract_morphological_features(current_image) / / Morphological data processing normalized_measurements = normalize_measurements(morphological_data); environmental_features = process_environmental_context(environmental_data) / / Multimodal feature fusion fused_features=fusion_network(spatial_features,morphological_features, normalized_measurements, environmental_features) / / Classification of growth stages stage_predictions = stage_classifier(fused_features); stage_confidence = calculate_confidence(stage_predictions); current_stage=determine_primary_stage(stage_predictions) / / Phase progress assessment stage_progress = assess_current_stage_progress(fused_features,current_stage); return format_stage_result(current_stage, stage_confidence, stage_progress) Morphological feature extraction methods: Contour analysis: plant contour extraction based on image segmentation; Key point detection: detection of leaf nodes, branching points, and flower bud positions; Texture analysis: analysis of leaf surface texture changes; Color features: statistical analysis of leaf and stem color changes; Geometric measurement: automatic measurement of plant height, crown width, and leaf area; Multi-scale feature fusion strategy: Multi-scale convolution: extraction of morphological features at different scales; Feature pyramid: construction of hierarchical feature representation; Attention weight: assigning higher weights to key morphological regions; Stage determination logic: Morphological threshold: comparison of key morphological indicators with standard stage features; Comprehensive evaluation: stage identification through multi-feature fusion; Expert rules: logical rule verification based on botanical knowledge; Anomaly detection mechanism: Developmental anomaly identification: detection of morphological, pathological, and other abnormal development; Artificial intervention marking: marking human intervention events such as pruning and transplanting; Performance features: Stage identification: support for growth stage classification; Progress estimation: providing growth progress assessment.

[0054] Then proceed to step three, in which, as follows: Figure 5 , Figure 6 and Figure 7As shown, the structured dataset is processed using an information aggregation and formatting method to generate a unified data contract. The information aggregation and formatting method in S3 includes data confidence labeling and redundancy removal. Specifically, the structured collaborative processing module is used to standardize the output data of the multiple sub-models, aiming to solve the inconsistency caused by randomness in expression style (such as the mixed use of units of measurement "5g" and "5 grams"), ensuring that the output data has a unified format and clear semantics, forming a structured dataset. The information aggregation and formatting method includes extracting model output, parsing structured or text results, field alignment and naming standardization, unit unification and scale conversion, labeling confidence and removing outliers, performing redundancy merging and initial conflict screening, and finally generating a unified data contract. This achieves field alignment, unit unification, confidence labeling, redundancy removal, and structuring of the outputs of each sub-model, generating a unified data contract for use by the strategy engine.

[0055] Figure 5 This data flow and time-series interaction diagram illustrates the message flow and temporal sequence from the camera / sensor to multiple AI models, information aggregation and formatting, the policy engine, and the execution and supervision modules. A unified data field is defined: factor, target value, upper and lower limits, confidence level, time period, duration, and execution priority, for consistent use by the policy engine and the execution and supervision modules; this unified data contract forms the basis for the input data format of execution control. The core function of information aggregation and formatting is to convert the heterogeneous output data of the five sub-models into a unified format that the strategy generation engine can understand. Input data types include: Health assessment model output: JSON format health score, confidence level, and risk factors; Disease identification model output: detection box coordinates, disease type, and severity assessment; Pest identification model output: pest type, quantity statistics, and activity level analysis; Environmental suitability model output: multi-dimensional environmental factor scores and limiting factor identification; Growth stage identification model output: growth stage identifier and development progress. Formatting algorithms include: Field mapping mechanism: establishing a mapping relationship between the output fields of each model and a unified data structure; Numerical standardization: uniformly converting numerical values ​​of different dimensions into a 0-10 scoring system; Confidence fusion: generating an overall data reliability assessment by combining the confidence levels of each model; Timestamp alignment: ensuring the temporal consistency of data from multiple models; and a unified output data structure. {"timestamp": "2024-01-15T10:30:00Z", / / Data generation timestamp, ISO 8601 format"plant_status": { / / Overall plant status assessment"health_score":8.5, / / Comprehensive health score (0-10 points)"disease_risk":0.3, / / Disease risk assessment (0-1 probability value)"pest_risk":0.2, / / Pest risk assessment (0-1 probability value)"environmental_suitability":7.8, / / Environmental suitability score (0-10 points)"growth_stage":"vegetative_growth" / / Current growth stage identifier},"detailed_analysis":{ / / Detailed analysis results for each model"health_factors":{...}, / / Plant health factor analysis "disease_detections":[...], / / Detailed results of disease detection "pest_detections":[...], / / Detailed results of insect detection "environmental_factors":{...}, / / Environmental factor analysis "growth_indicators":{...} / / Growth indicator monitoring},"data_quality":{ / / Data quality assessment index "overall_confidence":0.89, / / Overall confidence level (0-1 probability value) "model_agreement":0.92 / / Inter-model consistency level (0-1 probability value)}} Formatting algorithm pseudocode: Algorithm:DataFormatStandardization;Input:model_outputs[501-505]; Output: standardized_data / / Data validation and preprocessing for each model_output in model_outputs:validate_data_integrity(model_output); normalize_timestamp(model_output) / / Field mapping and numerical standardization mapped_data = {} for each model_output:mapped_fields = apply_field_mapping(model_output); normalized_values ​​= normalize_to_0_10_scale(mapped_fields); mapped_data.merge(normalized_values) / / Confidence fusion and quality assessment overall_confidence = calculate_weighted_confidence(model_outputs); data_quality = assess_data_consistency(mapped_data) / / Generate output in a uniform format return format_standardized_output(mapped_data, overall_confidence,data_quality) Then, step four is executed. In step four, an intelligent maintenance strategy is generated based on the unified data contract and historical data. The intelligent maintenance strategy adaptively adjusts based on the historical execution records of the historical data, the environment, execution, and effects. Specifically, the intelligent maintenance strategy generation adaptively adjusts based on historical execution records and the three-dimensional data of environment-execution-effect. Its core function is to reconcile contradictions based on current sensor data and multi-AI model analysis results, and generate an instant maintenance strategy that includes time, execution commands, and parameters. Strategy generation is based on the unified data contract and knowledge base, generating parameterized strategies such as watering, supplemental lighting, temperature control, and fertilization, including timing, intensity, duration, and execution cycle.

[0056] Then proceed to step five, in which, as follows: Figure 8 As shown, environmental conflict detection is performed based on the intelligent maintenance strategy to form an executable intelligent maintenance strategy. The method for environmental conflict detection includes: detecting conflict factors; determining the priority of the conflict factors; prioritizing them according to the priority of life safety and growth stage; adopting a time-sharing plan and determining whether the safety boundary is exceeded; if so, executing a noise reduction strategy; and forming an execution plan. The executable intelligent maintenance strategy includes maintenance timing, maintenance intensity, duration, and execution cycle. Specifically, Figure 8This diagram illustrates the environmental conflict handling process / state machine, showing the key steps and state transitions involved in priority determination, time-sharing strategies, and compromise balancing. When conflicting factors exist within the strategy, they are prioritized as follows: life safety > water > temperature > light > nutrition. A time-sharing and compromise balancing mechanism is then used to generate an executable compromise solution. This process includes detecting conflicting factors, calculating priorities, and determining whether they can be satisfied in a time-sharing manner. If they can be satisfied in a time-sharing manner, a time-sharing plan is generated; otherwise, a compromise balancing solution is obtained. Subsequently, the safety boundary is checked, and a degradation strategy is activated if necessary. Finally, an executable solution is output.

[0057] Input data format: Standardized plant status data: from data in a unified format; Plant species are configured with specific plant maintenance rules and parameters; Real-time decision-making algorithm: Weight allocation mechanism: dynamically allocates the weights of each model based on the current data confidence level; Conflict detection algorithm: identifies conflicting suggestions between models, such as the disease model suggesting increased humidity vs. the environment model suggesting decreased humidity; Prioritization: prioritizes maintenance needs according to their current urgency; Balance decision: finds the current optimal balance point among mutually restrictive maintenance measures; Strategy output format: {"strategy_id":"STG_20240115_001", / / Unique identifier for the strategy "generation_time":"2024-01-15T10:30:00Z", / / Strategy generation timestamp, ISO 8601 format "priority_level":"high", / / Execution priority: low / medium / high / urgent "execution_plan":[ / / Specific execution plan array {"action_type":"irrigation", / / Action type: watering / fertilizing / lighting / ventilation, etc. "executor_id":"WATER_PUMP_01", / / Actuator device identifier "start_time":"2024-01-15T11:00:00Z", / / Planned start execution time "duration":300, / / Execution duration (seconds) "parameters":{ / / Execution parameter configuration "flow_rate":2.5, / / Flow rate (liters / minute) "target_moisture":65 / / Target humidity percentage}},{"action_type":"lighting", / / Action type: supplemental lighting control "executor_id": "LED_PANEL_01", / / LED panel device identifier "start_time": "2024-01-15T12:00:00Z", / / Planned start time "duration": 14400, / / Execution duration (seconds, 4 hours) "parameters": { / / Illumination parameter configuration "intensity": 80, / / Illumination intensity percentage "spectrum": "full" / / Spectrum type: full / red / blue}}],"conflict_resolution": { / / Conflict resolution information "detected_conflicts": ["humidity_vs_disease_control"], / / List of detected conflict types "resolution_method": "weighted_compromise", / / Conflict resolution method: weighted compromise "confidence": 0.87 / / Solution confidence (0-1 probability value)}} Conflict Coordination Mechanism: Conflict Type Identification: Environmental Regulation Conflict, Time Scheduling Conflict, Resource Competition Conflict; Coordination Strategies: Time Staggering, Parameter Compromise, Phased Execution; Strategy Generation Algorithm Pseudocode: Algorithm: StrategyGeneration; Input:standardized_data,plant_species_config; Output: care_strategy / / Current Status Requirements Analysis and Priority Assessment care_needs=analyze_current_care_requirements(standardized_data); priorities=calculate_immediate_priority_scores(care_needs); sorted_needs=sort_by_current_priority(priorities) / / Conflict Detection and Collision Identification conflicts=detect_conflicts(sorted_needs); conflict_matrix=build_conflict_matrix(conflicts) / / Real-time strategy coordination and optimization for each conflict in conflicts:resolution=apply_immediate_conflict_resolution(conflict,plant_species_config); optimized_actions=optimize_current_parameters(resolution) / / Generate immediate execution plan execution_plan=generate_immediate_schedule(optimized_actions); resource_allocation=allocate_available_executors(execution_plan) / / Strategy verification and output validated_strategy=validate_immediate_feasibility(execution_plan); return format_strategy_output(validated_strategy) Then proceed to step six, in which, as follows: Figure 9 , Figure 10 , Figure 11 and Figure 12 As shown, the intelligent maintenance executable strategy is converted into device instructions, which are then used to execute automatic control and supervision, collect execution readback data, and generate receipts. The automatic control sends control instructions, including heating and cooling, irrigation, supplemental lighting, and fertilization, to the device equipped with the sensors via message queues and a local bus. Specifically, Figure 10 This is a schematic diagram of the internal structure of the multi-AI model collaboration and information aggregation formatting module, demonstrating the process of parallel output of expert models, field alignment, unit unification, confidence labeling, and strategy generation and conversion into executable commands. Figure 11 The core modules are represented using a modular design approach. Figure 12 This diagram shows the topology of the equipment and sensors, illustrating the irrigation, supplemental lighting, heating / cooling, and fertilization actuators and their connections to the controller and gateway / message queue. The core function of the control command conversion is to translate the maintenance strategies output by the strategy generation engine into standardized equipment commands that can be directly executed by various actuators. Automatic control issues commands to the irrigation, supplemental lighting, heating / cooling, and fertilization equipment via message queues or local buses, and collects and reads back data for monitoring and reporting. Input strategy format: Execution plan data: from structured maintenance strategy; Equipment registration information: protocol type, parameter range, and status information of each actuator; Equipment capability mapping: mapping relationship from abstract actions to specific equipment functions; Equipment protocol adaptation type: MQTT protocol adaptation: a lightweight message transmission protocol for IoT devices; Modbus protocol adaptation: a standard communication protocol for industrial control equipment; TTP / REST interface: Web API calls for smart devices; Serial communication protocol: RS485 / RS232 communication for traditional devices; Command generation algorithm: Action mapping: mapping abstract maintenance actions to specific equipment functions; Parameter conversion: converting strategy parameters into control parameters that the equipment can recognize; Protocol encapsulation: encapsulating command data according to the equipment communication protocol format; Command verification: verifying the legality and security of the generated commands; Standardized command format: {"command_batch_id":"CMD_20240115_001", / / Unique identifier for instruction batch "generation_time":"2024-01-15T10:30:00Z", / / Instruction generation timestamp, ISO 8601 format: "device_commands":[ / / Device command list array {"executor_id":"WATER_PUMP_01", / / Target actuator device identifier "protocol":"modbus", / / Communication protocol type: modbus / mqtt / http etc. "command_data":{ / / Protocol-specific command data "function_code":6, / / Modbus function code: 6 = Write a single register "register_address":1001, / / Target register address "value":250, / / Write value (corresponding to flow rate 2.5L / min*100) "execution_time":"2024-01-15T11:00:00Z" / / Scheduled execution time}}, {"executor_id":"LED_PANEL_01", / / LED panel device identifier "protocol":"mqtt", / / MQTT communication protocol command_data":{ / / MQTT protocol command data "topic":"greenhouse / lighting / led01 / control", / / MQTT topic path "payload":{ / / Message payload content "action":"set_intensity", / / Action to execute: set light intensity "value":80, / / Light intensity percentage "duration":14400 / / Duration (seconds, 4 hours)}}}],"execution_sequence":["WATER_PUMP_01","LED_PANEL_01"], / / Device execution sequence array "rollback_commands":[...] / / Set of rollback commands (used in case of exceptions)} Protocol conversion rules: MQTT command format: Topic path rules, JSON payload structure, QoS level settings; Modbus command format: Function code selection, register address mapping, data format conversion; Security mechanisms: Command signing, access control, exception handling; Command conversion algorithm pseudocode: Algorithm:CommandTranslation;Input:care_strategy,device_registry; Output: executable_commands / / Strategy Analysis and Device Matching action_list=parse_execution_plan(care_strategy); for each action in action_list:target_device=match_device_capability (action, device_registry); validate_device_availability(target_device) / / Parameter conversion and protocol adaptation for each (action, device) pair:device_params=convert_strategy_params (action.parameters,device.param_range) protocol_command=adapt_to_protocol(device_params,device.protocol_type) / / Command generation and verification command_batch=generate_command_batch(protocol_commands) validate_command_safety(command_batch) generate_rollback_commands(command_batch) / / Perform sequence optimization execution_sequence=optimize_execution_order(command_batch) return format_executable_commands(command_batch,execution_sequence) Then proceed to step seven, in which, as follows: Figure 13As shown, based on the execution readback data and the executable intelligent maintenance strategy, periodic monitoring and management are performed. The system adaptively adjusts based on the results of this periodic monitoring and management, and outputs an intelligent maintenance result report. The periodic monitoring and management includes dynamically adjusting the monitoring and execution frequencies based on the execution results, and pushing anomaly handling suggestions and reports via mobile devices. Specifically, Figure 13 The flowchart illustrates the task scheduling and adaptive adjustment process, showcasing the rules and decision points for dynamically adjusting monitoring and execution frequencies based on performance. Strategies are translated into device commands, such as MQTT / local bus, to control irrigation, supplemental lighting, heating / cooling, and fertilization actuators; execution logs are collected, read back, and recorded. The execution closed loop includes data collection, collaborative analysis, summary formatting, strategy generation, conflict resolution, control issuance, device execution, and monitoring and readback. It determines whether the target has been achieved; if not, adaptive adjustments are made; otherwise, a report is generated and archived. Monitoring and execution frequencies are dynamically adjusted based on performance and plant status, forming an adaptive closed loop; anomaly handling guidelines and summary reports are pushed to users via WeChat / APP.

[0058] The adaptive adjustment is a system optimization and learning capability; data-driven optimization; (1) historical data analysis: collect plant growth data and environmental control data; analyze the impact of different environmental parameters on plant growth; optimize the generation algorithm of maintenance plan. (2) user feedback learning: collect user evaluation of maintenance effect; adjust control parameters according to user feedback; continuously improve system performance. Intelligent algorithm optimization; (1) machine learning optimization: train optimization model based on historical data; improve the accuracy of plant identification; optimize the control precision of environmental parameters; (2) adaptive control: automatically adjust control strategy according to plant growth stage; dynamically optimize maintenance plan.

[0059] Finally, proceed to step eight. In step eight, as follows: Figure 14 and Figure 15 As shown, the security mechanism continuously performs anomaly detection during the execution of steps one through seven. When an anomaly is detected, the anomaly handling mechanism of the security mechanism is triggered. The anomaly handling mechanism includes an emergency stop procedure, sensor anomaly detection, data anomaly detection, alarm notification, and automatic recovery. Figure 14 This diagram illustrates the user interaction and alarm push notification process, showcasing the interface flow of receiving abnormal notifications, handling suggestions, and periodic reports in WeChat / APP.

[0060] Emergency stop procedure design: (1) Emergency stop triggering conditions: equipment fault detection; sensor failure; abnormal environmental parameters; abnormal system operation. (2) Emergency stop execution process: immediately stop all automated control operations, including water supply, temperature control, light management, fertilizer application, etc.; all related equipment enters the shutdown state.

[0061] Anomaly detection mechanism: (1) Sensor anomaly detection: Real-time monitoring of various sensor output data; detection of abnormal changes in temperature, humidity, light, and nutrients; periodic self-testing of sensors; evaluation of the rationality of sensor output data. (2) Data anomaly judgment: Sudden increase or decrease in temperature exceeding the preset range; excessive or insufficient water supply; abnormal increase or decrease in light; fertilizer application amount not in compliance with standards.

[0062] Alarm Notification System: (1) Abnormal Alarm: Sends alarm notifications to users via WeChat official account; includes detailed abnormal detection reports; describes the cause of the emergency stop and the system response measures; suggests that users check and manually intervene in accordance with event generation, branch decision and push. (2) Troubleshooting Guidance: Provides troubleshooting guidelines; supports users to contact technical support via WeChat official account; records the troubleshooting process and solutions.

[0063] Automatic recovery mechanism: (1) Recovery after troubleshooting: Re-perform sensor detection and self-test; confirm that all equipment and sensors have returned to normal; automatically release the emergency stop state; restore the regular automated control program. (2) System restart process: Gradually restart each control device; reload the maintenance plan generated by the strategy engine; adjust the maintenance plan according to the latest environmental data. (3) Data recording and analysis: Record the data and operation logs during the emergency stop and recovery process; upload them to the server and store them in the cloud database; users can access the running history through the WeChat official account.

[0064] In general, this embodiment utilizes multi-source data acquisition through timed visual monitoring and various sensors to collect data such as plant images, soil moisture, air humidity, environmental and soil temperature, light intensity, soil nutrients, electrical conductivity, and pH. This invention employs collaborative analysis of multiple AI models, parallelly calling plant health status assessment models, disease identification models, pest identification models, environmental suitability analysis models, and plant growth status / stage identification models to obtain their respective professional conclusions and confidence levels. The formatting method is an information aggregation method that structures and standardizes the output of each sub-model, including field alignment, unit unification, confidence level labeling, and redundancy removal, to eliminate uncertainties in the format, order, and expression of the large language model output, generating a unified data contract. The intelligent maintenance strategy is generated based on the formatted multi-source conclusions, plant category and current status, and historical execution and effect data, generating personalized maintenance strategies and parameters including watering amount and timing, supplemental light intensity and duration, temperature control targets, fertilizer type and dosage, etc. Environmental conflict management addresses conflicting needs among multiple factors by prioritizing plant health and growth stages, employing a time-sharing and compromise mechanism to arrive at an executable compromise solution. Automatic control translates strategies into equipment commands and executes them, controlling actuators for irrigation, supplemental lighting, heating / cooling, and fertilization, while recording execution results and sensor readouts. Periodic monitoring adaptively sets monitoring and execution frequencies based on plant status and performance, forming a closed loop of "monitoring—analysis—strategy—execution—supervision—reporting," and pushes anomaly handling guidelines and maintenance reports to users via WeChat / APP. Safety mechanisms include sensor self-checks, anomaly redundancy, emergency stop, and manual intervention procedures to ensure safety in abnormal situations.

[0065] Specific implementation method two: The automatic plant maintenance device based on multi-AI model collaboration described in this implementation method includes: The data acquisition module is used to acquire historical data and also to acquire new data using video surveillance and sensors; the new data includes visual data and sensor data. A model processing module is used to process newly collected data in parallel using multiple AI models to obtain a structured dataset. The multiple AI models include multiple sub-models and a structured collaborative processing module. The sub-models include: The sub-models are: Plant Health Status Assessment Sub-model for identifying characteristics of plant leaves; Disease Identification Sub-model for locating and classifying plant diseases; Pest Identification Sub-model for locating and classifying pest targets; Environmental Suitability Analysis Sub-model for nonlinear modeling and dynamic weight adjustment of multiple environmental factors; and Plant Growth Stage Identification Sub-model for identifying growth stages based on the current morphological characteristics of the plant. A data formatting module is used to process the structured dataset using information summarization and formatting methods to generate a unified data contract. Initial strategy module; used to generate intelligent maintenance strategies based on the unified data contract and historical data; The execution strategy module is used to perform environmental conflict detection based on the intelligent maintenance strategy and form an executable intelligent maintenance strategy. The control and supervision module is used to convert the intelligent maintenance executable strategy into equipment instructions, use the equipment instructions to perform automatic control and supervision, collect execution readback data and generate a receipt. The strategy adjustment module is used to perform periodic monitoring and management based on the execution readback data and the intelligent maintenance executable strategy, and to make adaptive adjustments according to the results of the periodic monitoring and management, and output an intelligent maintenance result report.

Claims

1. A plant automatic maintenance method based on multi-AI model cooperation, characterized in that, The method comprises: S1. Obtain historical data and collect new data using video monitoring and sensors; the new data includes visual data and sensor data; S2. Process the collected new data in parallel using a multi-AI model to obtain a structured data set; the multi-AI model includes multiple sub-models and a structured collaborative processing module; the sub-models include: a plant health status evaluation sub-model for identifying features of plant leaves; a disease identification sub-model for disease area positioning and classification detection of plants; an insect pest identification sub-model for insect pest target positioning and classification identification; an environmental suitability analysis sub-model for nonlinear modeling and dynamic weight adjustment of multiple environmental factors; a plant growth status stage identification sub-model for growth stage identification based on current morphological characteristics of plants; S3. Process the structured data set using an information summary formatting method to generate a unified data contract; S4. Generate an intelligent maintenance strategy based on the unified data contract and historical data; S5. Perform environmental conflict detection based on the intelligent maintenance strategy to form an intelligent maintenance executable strategy; S6. Convert the intelligent maintenance executable strategy into device instructions to perform automatic control and supervision, collect execution readback data and generate a receipt; S7. Based on the execution readback data and the intelligent maintenance executable strategy, perform periodic monitoring and management, and output an intelligent maintenance result report based on the results of the periodic monitoring and management.

2. The plant automatic maintenance method based on multi-AI model cooperation according to claim 1, characterized in that, The visual data in S1 is plant image data; the sensor data includes soil moisture, air humidity, environmental temperature, soil temperature, light intensity, soil nutrients, electrical conductivity, and pH value. 3.The plant automatic maintenance method based on multi-AI model cooperation of claim 1, wherein, In the multi-AI model in S2, the plant health status evaluation sub-model uses a ResNet-50 convolutional neural network architecture to identify plant features, and the output data includes a health score, a confidence level, a risk factor list, and recommended measures; the disease identification sub-model uses a YOLOv8 target detection framework combined with an EfficientNet feature extractor to position and classify disease areas on plants, and the output data includes disease type, detection frame coordinates, detection confidence, severity, proportion of affected area, overall health status, and treatment urgency; the insect pest identification sub-model uses a Faster R-CNN two-stage detection framework combined with a ResNeXt-101 backbone network to position and classify insect pest targets on plants, and the output data includes pest type, developmental stage, detection frame coordinates, detection confidence, individual number estimate, activity level, pest severity, affected part, risk level, detection time, and recommended monitoring frequency; the environmental suitability analysis sub-model uses a multi-layer perceptron to identify multiple environmental factors, including climate factors, soil factors, and physical factors, and the output data includes a comprehensive suitability score and optimization suggestions, where the comprehensive suitability score includes light intensity suitability score, soil nutrient content suitability score, and wind speed suitability score; The plant growth state stage recognition sub-model adopts a ResNet-50 convolutional neural network combined with multi-scale feature fusion to recognize the morphological characteristics of the plant, and outputs data including the current growth stage, stage determination confidence, progress within the current stage, key indicators, and development rate evaluation.

4. The plant automatic maintenance method based on multi-AI model cooperation according to claim 1, characterized in that, The information summarization and formatting method in S3 includes data confidence labeling and redundancy removal.

5. The plant automatic maintenance method based on multi-AI model cooperation according to claim 1, characterized in that, The intelligent maintenance strategy in S4 is adaptively adjusted based on the historical execution records of the historical data and the environment, execution, and effect.

6. The plant automatic maintenance method based on multi-AI model cooperation according to claim 1, characterized in that, The environment conflict detection method in S5 includes detecting conflict factors, determining the priority of the conflict factors, sorting according to the priority of life safety and growth stage, adopting a time-sharing plan and judging whether it exceeds the safety boundary, if so, executing a noise reduction strategy, and forming an execution scheme.

7. The plant automatic maintenance method based on multi-AI model cooperation according to claim 1, characterized in that, The intelligent maintenance executable strategy in S5 includes maintenance timing, maintenance intensity, duration, and execution cycle. 8.The plant automatic maintenance method based on multi-AI model cooperation of claim 1, wherein, The automatic control in S6 sends control instructions including heating and refrigeration, irrigation, light supplementation, and fertilization to the device carrying the sensor through a message queue and a local bus.

9. A plant automatic maintenance device based on multi-AI model cooperation, characterized in that, The device comprises: a data acquisition module for obtaining historical data and collecting new data using video monitoring and sensors; the new data includes visual data and sensor data; a model processing module for parallel processing of the collected new data using multiple AI models to obtain a structured data set; the multiple AI models include multiple sub-models and a structured collaborative processing module; the sub-models include: a plant health state evaluation sub-model for identifying features of plant leaves; a disease identification sub-model for positioning and classifying detection of plant disease areas; an insect pest identification sub-model for positioning and classifying identification of insect pests; an environmental suitability analysis sub-model for nonlinear modeling and dynamic weight adjustment of multiple environmental factors; and a plant growth state stage recognition sub-model for growth stage recognition based on the current morphological characteristics of the plant; a data formatting module for processing the structured data set using an information summarization and formatting method to generate a unified data contract; an initial strategy module for generating an intelligent maintenance strategy based on the unified data contract and historical data; an execution strategy module for detecting environmental conflicts based on the intelligent maintenance strategy to form an intelligent maintenance executable strategy; a control and supervision module for converting the intelligent maintenance executable strategy into device instructions, executing automatic control and supervision using the device instructions, collecting execution readback data, and generating a receipt; a strategy adjustment module for periodic monitoring and management based on the execution readback data and the intelligent maintenance executable strategy, adaptively adjusting according to the results of the periodic monitoring and management, and outputting an intelligent maintenance result report.

10. A storage medium, characterized by The storage medium stores a computer program, which, when executed, implements the plant automatic maintenance method based on multi-AI model collaboration according to any one of claims 1-8.

Images