Knowledge calibration methods and systems for smart garden knowledge systems

Abstract

This invention relates to the fields of artificial intelligence and the Internet of Things (IoT), specifically to a knowledge calibration method and system for a smart garden knowledge system. By constructing a closed loop of "perception-reasoning-execution-feedback-calibration," and based on real-time environmental data, expert rules, and historical maintenance feedback, it periodically verifies and adaptively corrects entity attributes, relation weights, and rule thresholds in the knowledge graph. It introduces a confidence decay model and a consistency constraint mechanism to ensure the synergistic consistency of knowledge at the semantic, logical, and numerical levels. The system includes modules for multi-source data acquisition, knowledge reasoning, effect collection, deviation calculation, calibration triggering, impact analysis, parameter reestimation, consistency verification, and confidence management. This application, by using actual maintenance effects as the calibration basis, significantly improves the accuracy, robustness, and credibility of decision-making, enabling the continuous evolution and highly reliable service of the smart garden knowledge system.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and the Internet of Things, and in particular to a knowledge calibration method and system for a smart garden knowledge system. Background Technology

[0002] With the increasing application of smart garden systems in urban ecological management, landscape maintenance, and resource optimization, knowledge-based intelligent decision support technology is becoming a core support. These systems typically rely on knowledge graphs to integrate multi-source information such as plant attributes, environmental factors, and maintenance standards, and combine this with sensing devices and artificial intelligence models to achieve precise scheduling of garden operations such as irrigation, pruning, and pest and disease control. However, existing technologies have significant shortcomings in the dynamic evolution capabilities of knowledge systems, the consistency assurance of multi-source heterogeneous knowledge, and the closed-loop verification of knowledge credibility for actual operational results, making it difficult to meet the high reliability and adaptability requirements of smart garden operation and maintenance.

[0003] In existing technologies, such as the knowledge graph-based plant irrigation and maintenance method disclosed in publication number CN115661547B, although a preliminary integration of environmental perception and water demand prediction has been achieved, the content of its knowledge graph mainly relies on preset plant growth habits and environmental preference data, lacking a dynamic verification mechanism for the accuracy of the knowledge itself. When external conditions change significantly or horticultural knowledge is updated, the system cannot automatically detect knowledge deviations and trigger a calibration process. More importantly, the relationship between the perception model and the knowledge graph in this scheme is only a one-way call relationship, without establishing a closed-loop optimization path of "perception-reasoning-execution-feedback-calibration," resulting in the actual irrigation effect not being able to be used to correct the knowledge rules. This structural defect causes the knowledge system to gradually become inaccurate over long-term operation, seriously affecting the robustness and credibility of decision-making, especially in complex and ever-changing open garden scenarios, making it difficult to support high-precision, adaptive intelligent maintenance services. Therefore, there is an urgent need for a smart garden knowledge calibration method and system with dynamic calibration capabilities, a closed-loop feedback mechanism, and multi-source knowledge consistency maintenance functions. Summary of the Invention

[0004] To achieve the above objectives, this invention proposes a closed-loop knowledge evolution framework of "perception-reasoning-execution-feedback-calibration". Its core idea is to implement the reasoning decisions of the knowledge graph and collect the actual results as a "truth standard," thereby driving the quantitative deviation assessment and calibration of parameters such as entities, relationships, and rules in the knowledge graph. Simultaneously, it introduces a confidence level that decays over time for all knowledge units, achieving preventative management of knowledge aging, and ensures the rigor of the calibrated knowledge system through multi-granularity consistency verification.

[0005] On the one hand, this invention provides a knowledge calibration method for a smart garden knowledge system, which includes:

[0006] The system acquires real-time environmental perception data, expert rule data, and historical maintenance operation feedback data. The real-time environmental perception data includes deep-layer soil volumetric water content obtained through a soil moisture sensor array, air temperature and humidity, light intensity, rainfall, and wind speed data obtained from a weather station, and plant canopy temperature distribution obtained through an infrared thermal imager. The expert rule data includes structured knowledge such as irrigation thresholds, pruning cycles, pest and disease control conditions, and fertilizer ratios corresponding to different plant species. The historical maintenance operation feedback data is extracted from the garden operation management system and contains a complete time-series record of the environment and plant status before and after the operation.

[0007] Based on the real-time environmental perception data and expert rule data, a knowledge graph is invoked for reasoning and decision-making, generating and issuing maintenance operation instructions. The reasoning based on the knowledge graph can be jointly driven by a rule-based reasoning engine and a graph neural network to integrate explicit rules and implicit patterns.

[0008] Collect actual operational data generated after executing the maintenance operation instructions, such as quantifiable indicators like soil moisture stability, plant transpiration rate changes, chlorophyll fluorescence parameters, and lesion expansion area;

[0009] The knowledge deviation metric is calculated between the actual operational performance data and the expected baseline value of the corresponding inference path in the knowledge graph. The baseline value is predefined by the plant physiological response model associated in the knowledge graph. The deviation calculation can use a multidimensional weighted Euclidean distance formula to comprehensively evaluate deviations in multiple dimensions such as soil moisture recovery, physiological response lag, disease control, and resource consumption. The weights of each dimension can be dynamically adjusted according to the plant growth stage.

[0010] When the knowledge deviation metric exceeds the dynamically adjusted calibration trigger threshold, the knowledge calibration process is initiated. The dynamic calibration trigger threshold is based on a base threshold for the plant species and incorporates seasonal and historical calibration frequency factors; its calculation formula is as follows:

[0011]

[0012] in, A basic threshold set based on plant species. This is a seasonal correction factor with a value in the range of [-0.2, 0.2], taking a positive value in summer and a negative value in winter; The number of historical calibrations for the target plant over the past 30 days; , The adjustment factors are 0.1 and 0.05, respectively.

[0013] In the calibration process, a graph traversal algorithm is used to analyze the influence range of the knowledge graph, tracing back from the conclusion node that triggered the bias to all related premises, rules, and entity nodes, forming a set of knowledge units to be calibrated. Each knowledge unit includes entity attribute values, relation weight coefficients, and rule condition thresholds.

[0014] Based on the historical maintenance operation feedback data and the actual operation effect data, the weighted least squares regression method is used to re-estimate the parameters of the knowledge unit to be calibrated, and generate calibrated knowledge parameters.

[0015] The calibrated knowledge parameters are verified to be consistent with the original knowledge parameters. The consistency verification includes semantic consistency check, logical conflict detection and numerical boundary compliance judgment. If the verification is successful, the corresponding knowledge unit in the knowledge graph is updated. If the verification fails, the calibrated knowledge parameters are stored in the knowledge temporary storage area to be reviewed and a manual review task is generated.

[0016] The confidence level of all knowledge units in the knowledge graph is dynamically managed based on the time decay factor. The initial confidence level is 1, which decreases exponentially over time. Each successful calibration or expert confirmation operation resets the confidence level to 1. When the confidence level of any knowledge unit is lower than the preset lower limit, it is automatically marked as a state to be calibrated and included in the next round of calibration process.

[0017] As one embodiment of the present invention, the acquisition of real-time environmental perception data specifically includes: acquiring soil volumetric water content data of deep layers through a soil moisture sensor array deployed in the garden area; acquiring air temperature, relative humidity, total solar radiation and cumulative rainfall updated every ten minutes through a weather station; acquiring a surface temperature distribution map of plant canopy through an infrared thermal imager; and acquiring instantaneous wind speed and prevailing wind direction data through a wind speed and direction sensor.

[0018] As one embodiment of the present invention, the acquisition of the historical maintenance operation feedback data includes: extracting all maintenance records for the same individual plant or the same group of plants in the past 12 months from the garden operation management system. Each record includes the operation type, execution time, execution parameters, environmental conditions before the operation, and continuous monitoring data for 72 hours after the operation.

[0019] As one embodiment of the present invention, the knowledge deviation metric is calculated using a multidimensional weighted Euclidean distance formula. Its dimensions include soil moisture recovery deviation, plant physiological response lag time, disease control failure degree and resource consumption overrun ratio. The weights of each dimension are dynamically allocated according to the plant growth stage, and the weight allocation rules are stored in the meta-rule layer of the knowledge graph.

[0020] As one embodiment of the present invention, the influence range analysis is implemented by a graph traversal algorithm, which traces back from the inference conclusion node that triggers the deviation to all the precondition nodes, intermediate rule nodes and basic entity nodes that participate in the inference path, forming a directed acyclic subgraph. All edges and nodes in the subgraph constitute a set of knowledge units to be calibrated.

[0021] As one embodiment of the present invention, in the weighted least squares regression method, the sample weights... The time decay factor of the data Reliability coefficient corresponding to sensor accuracy level and the historical consistency score of this data point The product of the three factors constitutes the time decay coefficient α, which is 0.02 per day.

[0022] As one embodiment of the present invention, the semantic consistency check is performed through ontological constraint rules to ensure that the updated attribute values ​​conform to the predefined data type, value range and unit system; the logical conflict detection is performed by the rule engine to perform a full combination of precondition tests on the updated rule set to verify whether there are contradictory conclusions; the numerical boundary compliance determination is based on the plant physiological limit parameter table to ensure that the calibrated parameters do not exceed the species survival threshold.

[0023] As one embodiment of the present invention, the exponential decay function expression in the dynamic management of confidence is C(t) = exp(-λt), where C(t) is the confidence at time t, λ is the decay rate, and the value of λ is set according to the knowledge unit type: for environmental response rules, λ is 0.05 / day; for inherent plant attributes, λ is 0.01 / day; and for expert experience rules, λ is 0.03 / day.

[0024] According to another aspect of the present invention, a knowledge calibration system for a smart garden knowledge system is provided, comprising:

[0025] The multi-source data acquisition module is used to acquire real-time environmental perception data, expert rule data, and historical maintenance operation feedback data;

[0026] The knowledge reasoning and instruction generation module is used to call the knowledge graph to perform reasoning decisions for the current maintenance task based on the real-time environmental perception data and the expert rule data, generate maintenance operation instructions, and send them to the execution equipment.

[0027] The effect data acquisition module is used to collect the actual operation effect data after the maintenance operation command is executed;

[0028] The knowledge deviation calculation module is used to calculate the deviation between the actual operating effect data and the expected effect benchmark value of the corresponding reasoning path in the knowledge graph, and obtain the knowledge deviation metric value.

[0029] The calibration trigger judgment module is used to start the knowledge calibration process when the knowledge deviation metric value exceeds the preset calibration trigger threshold.

[0030] The influence range analysis module is used to analyze the influence range of entity nodes and their relationships related to the current deviation in the knowledge graph, and to determine the set of knowledge units to be calibrated.

[0031] The knowledge parameter re-estimation module is used to re-estimate the parameters of the knowledge unit to be calibrated based on the historical maintenance operation feedback data and the actual operation effect data, using the weighted least squares regression method, and generate calibrated knowledge parameters.

[0032] The consistency verification and update module is used to verify the consistency between the calibrated knowledge parameters and the original knowledge parameters. If the verification passes, the corresponding knowledge unit in the knowledge graph is updated. If the verification fails, the calibrated knowledge parameters are stored in the knowledge temporary storage area to be reviewed, and a manual review task is generated.

[0033] The knowledge confidence management module is used to dynamically manage the confidence of all knowledge units in the knowledge graph based on the time decay factor. When the confidence of any knowledge unit is lower than the preset lower limit, it is automatically marked as a state to be calibrated and included in the next round of calibration process.

[0034] In one embodiment of the present invention, the multi-source data acquisition module includes a soil moisture sensing unit, a meteorological monitoring unit, an infrared thermal imaging unit, and an operation management database interface unit; the soil moisture sensing unit samples every fifteen minutes; the meteorological monitoring unit integrates temperature, humidity, light, rainfall, and wind speed sensors, with a data update cycle of ten minutes; the infrared thermal imaging unit is installed on a fixed pole or drone platform and scans the garden area daily; the operation management database interface unit connects to the garden operation and maintenance information platform through an application programming interface to synchronize work orders and feedback records in real time.

[0035] As one embodiment of the present invention, the knowledge reasoning and instruction generation module has a built-in rule reasoning engine and a graph neural network joint inferencer. The rule reasoning engine processes explicit expert rules, and the graph neural network joint inferencer processes implicit association patterns. The output results of the two are weighted and fused to generate the final maintenance instruction.

[0036] In one embodiment of the present invention, the effect data acquisition module is deeply integrated with the garden IoT platform. The original sensor data is filtered, denoised, and feature extracted through the edge computing gateway, and a structured effect indicator data stream is output.

[0037] As one embodiment of the present invention, the consistency verification and update module includes three sub-units: ontology constraint checker, rule conflict detector, and physiological boundary checker, which respectively perform consistency verification at the semantic, logical, and numerical levels.

[0038] As one embodiment of the present invention, the knowledge confidence management module maintains a global confidence status table. The table is indexed by the unique identifier of the knowledge unit and records the creation time, last calibration time, current confidence value and decay rate parameter of each unit. The system background service scans the table every hour and updates the confidence value.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This invention addresses the fundamental flaw in existing smart garden systems—the static and inability to dynamically evolve knowledge systems—by constructing a complete closed loop of "perception-reasoning-execution-feedback-calibration." Using actual maintenance results as the calibration basis, this invention achieves strong coupling between knowledge rules and real-world responses, significantly improving the accuracy and robustness of decision-making. By introducing a dynamic management mechanism based on confidence decay, this invention proactively identifies knowledge aging risks, transforming passive correction into proactive prevention. A multi-granularity consistency verification mechanism ensures semantic rigor and logical completeness during knowledge updates, avoiding global knowledge conflicts caused by local corrections. Furthermore, this invention employs weighted least squares regression for parameter reestimation, fully considering the timeliness, reliability, and individual differences of historical data, making the calibration results more scientific and adaptable. In long-term operation, this invention significantly reduces resource waste and maintenance errors caused by inaccurate knowledge, providing a highly reliable knowledge foundation for the sustainable evolution of smart garden systems. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall technical architecture of the knowledge calibration method and system for a smart garden knowledge system proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of knowledge dynamic calibration based on the closed loop of "perception-reasoning-execution-feedback-calibration" in this invention; Figure 3 This is a logical flowchart of the multi-source data fusion, deviation-driven calibration, and dynamic confidence management in this invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Those skilled in the art, guided by the spirit of this invention, can derive other embodiments, all of which will fall within the protection scope of this invention.

[0043] This invention provides a knowledge calibration method for a smart garden knowledge system, comprising the following steps:

[0044] S1: Acquire real-time environmental perception data, expert rule data, and historical maintenance operation feedback data; the real-time environmental perception data includes soil moisture, air temperature and humidity, light intensity, rainfall, wind speed, and plant canopy temperature; the expert rule data includes irrigation threshold ranges corresponding to plant species, pruning cycle specifications, pest and disease control triggering conditions, and fertilizer ratio recommendations; the historical maintenance operation feedback data includes soil moisture recovery curves after historical irrigation, plant leaf water content change rate, new shoot growth rate, and frequency of pest and disease occurrence.

[0045] In step S1, real-time environmental perception data is collected by a multimodal sensor network deployed in the garden area. Soil moisture sensor arrays are arranged vertically at depths of 10cm, 30cm, and 50cm below the surface, collecting volumetric water content data every 15 minutes to form time-series data of soil moisture profiles at different depths. A weather station integrates sensors for air temperature, relative humidity, total solar radiation, and cumulative rainfall, with a data update cycle of 10 minutes to ensure high-frequency synchronous capture of environmental factors. Infrared thermal imagers, installed on fixed observation towers or inspection drone platforms, perform full-area scans of the target plant canopy daily, generating surface temperature distribution maps for inverting transpiration intensity and water stress status. Wind speed and direction sensors record instantaneous wind speed and prevailing wind direction, which are used as microclimate disturbance correction terms in subsequent inference processes. Expert rule data comes from authoritative horticultural literature and local maintenance regulations, and after structured processing, is stored in the rule layer of the knowledge graph in ontology form. Each rule includes preconditions, actions, applicable species range, and initial default confidence value. Historical maintenance operation feedback data is synchronized in real time through the garden operation management system interface. It covers all maintenance records for the same plant or the same plant group within the past 12 months. Each record includes operation type, execution time, execution parameters, environmental conditions before operation and continuous monitoring data within 72 hours after operation, forming a historical basis library for closed-loop calibration.

[0046] S2: Based on the real-time environmental perception data and the expert rule data, the knowledge graph is invoked to perform reasoning and decision-making for the current maintenance task, generate maintenance operation instructions, and send them to the execution equipment.

[0047] In step S2, the knowledge graph serves as the core reasoning engine. Its nodes represent plant entities, environmental factors, maintenance operations, and physiological indicators, while edges represent attribute relationships, causal dependencies, or rule constraints. The reasoning process is collaboratively completed by a rule-based reasoning engine and a graph neural network joint inference engine. The rule-based reasoning engine parses explicit expert rules, such as "trigger irrigation when soil moisture is below 30% and there is no rainfall in the next 24 hours," performing deterministic logical deduction. The graph neural network joint inference engine learns implicit association patterns, such as "the probability of disease occurrence for a specific variety under continuous high temperature and low humidity conditions shows a non-linear upward trend," aggregating neighborhood node features through a message passing mechanism and outputting a probabilistic risk score. The outputs of both are fused according to preset weights, which are dynamically adjusted based on the current task type: irrigation and fertilization tasks emphasize the rule engine output, while pest and disease early warning tasks emphasize the graph neural network output. The final generated maintenance operation instructions include the operation type, target area, execution parameters (such as irrigation water volume, pruning height, and pesticide concentration), and execution priority, and are sent to the corresponding execution devices, such as smart irrigation valves, automatic pruning robotic arms, or plant protection drones, via IoT communication protocols.

[0048] S3: Collect actual operational effect data after executing the maintenance operation command. The actual operational effect data includes the soil moisture stability value, plant transpiration rate change, chlorophyll fluorescence parameters, and lesion expansion area within the preset time window after execution.

[0049] In step S3, the effect data acquisition module is deeply integrated with the garden IoT platform. Edge computing gateways are deployed near each sensor node to perform real-time filtering, noise reduction, and feature extraction on the raw sensor data. Soil moisture stability is defined as the mean value of soil moisture content within 48 hours after the operation when the standard deviation is less than a preset threshold. Changes in plant transpiration rate are obtained by inverting the difference between canopy temperature and air temperature and humidity, calculated using a simplified Penman-Monteith model. Chlorophyll fluorescence parameters are measured periodically by a portable fluorometer, reflecting the maximum quantum efficiency of photosynthetic system II. The area of ​​diseased lesions is identified using a high-resolution visible light image segmentation algorithm to determine the diseased leaf region, calculate the pixel percentage, and convert it to the actual area. All effect indicators are labeled with timestamps, spatial coordinates, and associated operation instruction identifiers, forming a structured effect data stream for subsequent deviation analysis.

[0050] S4: The deviation between the actual operational effect data and the expected effect benchmark value of the corresponding reasoning path in the knowledge graph is calculated to obtain the knowledge deviation metric value; the effect benchmark value is predefined by the plant physiological response rules and environmental factor coupling model associated in the knowledge graph.

[0051] In step S4, the baseline effect value is stored in the effect prediction subgraph directly associated with the inference conclusion node in the knowledge graph. For example, if the inference conclusion is "implement 50mm irrigation," its effect prediction subgraph includes multi-dimensional expected indicators such as "soil moisture is expected to recover to 45%±5%", "transpiration rate recovers to more than 90% of the baseline value within 24 hours", and "chlorophyll fluorescence parameter Fv / Fm is not lower than 0.75". Bias calculation uses a multi-dimensional weighted Euclidean distance formula:

[0052]

[0053] in For the i-th dimension, the actual operating effect data, To correspond to the expected benchmark value, These are the dimensional weights. The weight allocation rules are stored in the meta-rule layer of the knowledge graph and are dynamically adjusted according to the plant's growth stage: during the vegetative growth stage, soil moisture and new shoot growth rate are emphasized; during the flowering and fruiting stage, transpiration rate and chlorophyll parameters are emphasized; and during the dormancy stage, the degree of disease control failure is emphasized. The proportion of resource consumption exceeding the set value is also included in the calculation. For example, the percentage by which actual water consumption exceeds the set value is adjusted upwards with the water resource stress index.

[0054] S5: When the knowledge deviation metric exceeds the preset calibration trigger threshold, the knowledge calibration process is initiated; the calibration trigger threshold is dynamically adjusted according to the plant species, seasonal factors and historical calibration frequency.

[0055] In step S5, the calibration trigger threshold is not a fixed constant, but is adaptively adjusted by the dynamic threshold management module within the system. The adjustment strategy follows these principles:

[0056] Set basic thresholds based on plant species; for example, set higher thresholds for drought-resistant plants and lower thresholds for environmentally sensitive plants.

[0057] Seasonal factors are introduced for correction. During the high temperatures of summer, the threshold can be appropriately relaxed to avoid overcalibration, while during the low temperatures of winter, the threshold is tightened to improve monitoring sensitivity. In addition, the historical calibration frequency must be considered. If a plant individual has triggered multiple calibrations recently (within 30 days), its threshold is temporarily increased by a certain percentage (20%) to prevent overfitting to random noise or transient interference. The system only determines a valid trigger signal and initiates the knowledge calibration process when the calculated knowledge deviation metric continuously exceeds the dynamic threshold for a preset continuous evaluation period (two periods).

[0058] S6: In the calibration process, the influence range analysis is performed on the entity nodes and their relationships related to the current deviation in the knowledge graph to determine the set of knowledge units to be calibrated; the knowledge units include entity attribute values, relationship weight coefficients, rule condition thresholds and inference function parameters.

[0059] In step S6, the scope of influence analysis is implemented using a depth-first graph traversal algorithm. Starting from the inference conclusion node that triggers the bias, all prerequisite nodes, intermediate rule nodes, and basic entity nodes involved in the inference path are traced backward along the directed edges of the knowledge graph until the original perceptual data node or expert rule root node, which cannot be further decomposed, is reached. During the traversal, all nodes and edges traversed are recorded, forming a directed acyclic subgraph. All quantifiable parameters within this subgraph are marked as knowledge units to be calibrated, including but not limited to: the "optimal soil moisture lower limit" attribute value of a certain plant, the "weighting coefficient" between environmental factors and irrigation requirements, the "temperature-humidity joint threshold" in disease early warning rules, and the "message passing attenuation factor" of a specific edge in the graph neural network.

[0060] S7: Based on the historical maintenance operation feedback data and the actual operation effect data, a weighted least squares regression method is used to re-estimate the parameters of the knowledge unit to be calibrated, generating calibrated knowledge parameters. The sample weights of the regression model... Calculated using the following formula:

[0061]

[0062] in, This is the time decay factor, with a value of 0.02 / day; This represents the number of days since the data collection time of the k-th sample was reached. The reliability coefficient corresponding to the accuracy level of the sensor used to collect this data. The historical data consistency score for the plant individual to which the data point belongs is obtained by calculating the residual standard deviation of the recent similar operation effect data of the plant, and mapping it to the interval [0.5, 1.0] after normalization.

[0063] The historical data consistency score is calculated as follows: For the target plant individual, the residuals are calculated based on the effect data of similar maintenance operations over the past 30 days, using the corresponding knowledge graph prediction value as a benchmark; the standard deviation of the residuals is calculated by using the seven most recent valid data points as a sliding window. Consistency score The linear normalization method is used to normalize the values ​​to the interval [0.5, 1.0].

[0064] S8: Perform consistency verification between the calibrated knowledge parameters and the original knowledge parameters. The consistency verification includes semantic consistency check, logical conflict detection, and numerical boundary compliance determination. If the verification passes, update the corresponding knowledge unit in the knowledge graph. If the verification fails, store the calibrated knowledge parameters in the pending knowledge temporary storage area and generate a manual review task.

[0065] In step S8, semantic consistency is checked by the ontology constraint checker, verifying whether the updated attribute values ​​conform to the predefined data type, value range, and unit system. Logical conflict detection is performed by the rule conflict detector, which exhaustively tests all combinations of preconditions on the updated rule set. If two different sets of preconditions lead to contradictory conclusions (e.g., simultaneously triggering irrigation and water restriction), a conflict is identified. Numerical boundary compliance is determined by the physiological boundary checker, referring to the plant physiological limit parameter table. For example, the lethal low temperature for rhododendron roots is -10 degrees Celsius. If the calibrated antifreeze rule threshold is lower than this value, the update is rejected. After all three verifications pass, the knowledge graph transaction manager performs an atomic update operation; if any verification fails, the parameters are stored in the pending knowledge temporary storage area, and a manual review task is pushed to the horticultural expert terminal through the operations and maintenance platform.

[0066] S9: Implement dynamic confidence management based on time decay factor for all knowledge units in the knowledge graph. The initial confidence value is 1, which decreases exponentially over time. Each successful calibration or expert confirmation operation resets the confidence value to 1. When the confidence value of any knowledge unit is lower than the preset lower limit, it is automatically marked as a state to be calibrated and included in the next round of calibration process.

[0067] In step S9, dynamic confidence management based on time decay is implemented for all knowledge units in the knowledge graph. The confidence level of each knowledge unit is... Follows the exponential decay function ,in This is the length of time since the last confidence reset or confirmation operation for this unit. The daily decay rate is set according to the knowledge unit type:

[0068] For environmentally responsive rules (such as irrigation and shading trigger conditions): the lower confidence limit is set to 0.65;

[0069] For inherent plant properties (such as optimal growth temperature range and maximum plant height): the lower confidence limit is set at 0.75;

[0070] For expert experience rules (such as pruning techniques and fertilizer ratios): the lower confidence limit is set at 0.60;

[0071] When the confidence level of any knowledge unit falls below this lower limit, it is automatically marked as a state to be calibrated and included in the next round of calibration process, thereby achieving proactive prevention and control of the risk of knowledge obsolescence.

[0072] The above-described methods and steps constitute a complete knowledge calibration closed loop. Correspondingly, the knowledge calibration system for the smart garden knowledge system includes a multi-source data acquisition module, a knowledge reasoning and instruction generation module, an effect data collection module, a knowledge deviation calculation module, a calibration trigger judgment module, an impact range analysis module, a knowledge parameter re-estimation module, a consistency verification and update module, and a knowledge confidence management module.

[0073] The multi-source data acquisition module integrates a soil moisture sensing unit, a meteorological monitoring unit, an infrared thermal imaging unit, and an operation management database interface unit. The soil moisture sensing unit employs frequency domain reflectance technology, with deployment depth covering the main root distribution layer. The meteorological monitoring unit conforms to World Meteorological Organization standards and has automatic calibration capabilities. The operation management database interface unit connects to the garden operation and maintenance information platform via a secure API, supporting incremental synchronization and breakpoint resume.

[0074] The knowledge reasoning and instruction generation module incorporates a rule-based reasoning engine and a graph neural network joint inference engine. The rule-based reasoning engine supports mixed forward and backward chain reasoning, handling complex rules including logical AND, OR, and NOT operations. The graph neural network joint inference engine employs a two-layer graph attention network (GAT), with each layer containing four attention heads and a node embedding dimension of 128. During training, historical disease occurrence records are used as labels, and cross-entropy loss is employed for optimization. The weighted fusion is implemented through a gating mechanism, with gating weights... ;in, Output vectors for the rule engine. For GAT output vectors, For the sigmoid function, and As learnable parameters, the gating inputs include task urgency, data integrity, and historical inference accuracy.

[0075] The effect data acquisition module achieves localized processing through an edge computing gateway. The gateway is equipped with a lightweight convolutional neural network for image dehazing and lesion segmentation; Kalman filtering is used to smooth the temporal sensor data; and the feature extraction module outputs standardized effect indicators to eliminate individual device differences.

[0076] The knowledge bias calculation module maintains a multi-dimensional bias measurement engine, supporting custom weighting strategies and distance measurement methods. The calibration trigger judgment module implements dynamic threshold scheduling, with threshold updates occurring daily at midnight. The influence range analysis module employs an improved Tarjan algorithm to ensure efficient identification of relevant subgraphs in large-scale knowledge graphs.

[0077] The knowledge parameter reestimation module encapsulates a weighted least squares solver, supporting batch calibration and incremental learning. The consistency verification and update module contains three independent verification sub-units, outputting verification results in a structured report format. The knowledge confidence management module maintains a memory-mapped confidence table, supporting millisecond-level queries and updates to ensure response performance in high-concurrency scenarios.

[0078] Example:

[0079] Taking the autumn irrigation rule calibration of ginkgo trees in a smart park as an example, this embodiment details how to trigger and complete the calibration of relevant knowledge when the system detects that the effect of an irrigation operation does not meet expectations.

[0080] Real-time sensing data was acquired. A soil moisture sensor (30cm depth) measured a volumetric water content of 28%, and the weather station forecast no rainfall in the next 24 hours. A rule stored in the knowledge graph was used: if the ginkgo trees are in their growing season, and the soil moisture is below 30%, and there is no rain in the next 24 hours, irrigation should be triggered, with a suggested irrigation amount of 50mm. The "Irrigate 50mm" command was generated and executed. 48 hours after irrigation, the soil moisture stabilized at 38%. Simultaneously, the transpiration rate calculated from the canopy temperature inversion recovered to 85% of the baseline value.

[0081] According to the effect prediction model associated with this rule in the knowledge graph, the expected soil moisture should recover to 45%±5% and the transpiration rate should recover to over 95%.

[0082] Deviation calculations were performed: soil moisture deviation: |38% - 45%| = 7%, transpiration rate deviation: |85% - 95%| = 10%. The current stage is set as the vegetative growth period of the ginkgo tree, and the weights are set as follows: soil moisture weight. transpiration rate weight Knowledge Bias Measurement .

[0083] Trigger calibration, Ginkgo tree base threshold Autumn coefficient Not calibrated for nearly 30 days Calculate the dynamic threshold. .

[0084] Starting from the node "Irrigation effect not as expected", we traverse backwards to determine the unit to be calibrated as: "soil moisture threshold (currently 30%)" and "recommended irrigation amount (currently 50mm)" in the rules; we collect historical data of the past 10 irrigations for the ginkgo tree, including soil moisture before irrigation, irrigation amount, and soil moisture recovery value after irrigation.

[0085] A regression model was established with "irrigation amount" and "pre-irrigation humidity" as independent variables and "humidity recovery value" as the dependent variable. Calculate the weight of each historical data point. For example, if a data point from 7 days ago was collected by a soil sensor with a reliability of 0.95, and its historical consistency score is 0.9, then its weight... The weighted least squares method was used to solve for the new model parameters. The new model assumes that, under current climatic conditions, to achieve the 45% humidity recovery target, the irrigation amount should be 58 mm, or the soil moisture threshold triggering irrigation should be adjusted to 32%.

[0086] Consistency verification and updates were performed, and the new parameters 58mm and 32% were both positive, which is consistent with the physical meaning.

[0087] The rules related to "ginkgo tree" in the knowledge graph were checked, and no conflicting rules such as "irrigation is prohibited when soil moisture > 30%" were found.

[0088] According to the Plant Physiology Handbook, the maximum irrigation amount for ginkgo trees is 100 mm, and the safe range for soil moisture is 20%-60%. The new parameters are all within the safe limits.

[0089] After successful verification, the system updated the "recommended irrigation amount" for this rule in the knowledge graph to 58mm and the "soil moisture threshold" to 32%.

[0090] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A knowledge calibration method for a smart garden knowledge system, characterized in that, include: Step S1: Acquire real-time environmental perception data, expert rule data, and historical maintenance operation feedback data; Step S2: Based on the real-time environmental perception data and the expert rule data, invoke the knowledge graph to execute the reasoning decision for the current maintenance task, generate maintenance operation instructions, and issue them to the execution equipment; Step S3: Collect the actual operation effect data generated after executing the maintenance operation instructions, the actual operation effect data including at least soil moisture stability value and plant physiological indicators; Step S4: Based on the preset multidimensional weighted Euclidean distance formula, calculate the knowledge deviation metric between the actual operation effect data and the expected effect benchmark value of the corresponding reasoning path in the knowledge graph; Step S5: Compare the knowledge deviation metric with a dynamic calibration trigger threshold. When the knowledge deviation metric exceeds the dynamic calibration trigger threshold, start the knowledge calibration process; wherein, the dynamic calibration trigger threshold is based on the basic threshold of plant species and dynamically adjusted by introducing seasonal factors and historical calibration frequency; Step S6: In the calibration process, perform an influence range analysis on the knowledge graph based on the graph traversal algorithm, and trace back all related entity nodes from the conclusion node that triggered the deviation. Relationship edges and rule nodes form a set of knowledge units to be calibrated; the knowledge unit includes entity attribute values, relationship weight coefficients, and rule condition thresholds; Step S7: Based on the historical maintenance operation feedback data and the actual operation effect data, the weighted least squares regression method is used to re-estimate the parameters of the knowledge units to be calibrated, generating calibrated knowledge parameters; wherein, the sample weights used in the regression model are composed of the product of the time decay factor of the corresponding data, the reliability coefficient corresponding to the sensor accuracy level, and the historical consistency score of the data point; Step S8: The calibrated knowledge parameters and the original knowledge parameters are verified for semantic consistency, logical conflict, and compliance of numerical boundaries; if the verification is successful, the corresponding knowledge unit in the knowledge graph is updated; Step S9: All knowledge units in the knowledge graph are subject to dynamic management of confidence based on time decay; the initial value of the confidence is 1, and it decays exponentially according to the knowledge unit type at different daily decay rates λ; when the confidence of a certain knowledge unit is lower than the preset lower limit, it is automatically marked as a state to be calibrated.

2. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, In step S1, acquiring real-time environmental perception data specifically includes: acquiring layered soil volumetric water content data through an array of soil moisture sensors deployed underground; acquiring air temperature, relative humidity, total solar radiation, and cumulative rainfall data through a meteorological station; and acquiring plant canopy surface temperature distribution data through an infrared thermal imager.

3. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, In step S4, the multidimensional weighted Euclidean distance formula is: ； in, For the i-th dimension, the actual maintenance effect data, The i-th dimension of the expected effect benchmark stored in the knowledge graph. For dynamically assigned dimension weights; The dimensional weights The allocation rules are stored in the meta-rule layer of the knowledge graph and are dynamically switched according to the vegetative growth stage, flowering and fruiting stage, or dormancy stage of the plant.

4. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, In step S5, the dynamic calibration trigger threshold It is obtained by calculation using the following formula: in, A basic threshold set based on plant species. This is a seasonal correction factor with a value in the range of [-0.2, 0.2], taking a positive value in summer and a negative value in winter; The number of historical calibrations for the target plant over the past 30 days; , The adjustment factors are 0.1 and 0.05, respectively.

5. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, In step S6, the graph traversal algorithm is a depth-first search algorithm. The specific process of determining the set of knowledge units to be calibrated includes: starting from the inference conclusion node that triggers the deviation, performing a reverse depth-first traversal along the directed edges of the knowledge graph; tracing back all the precondition nodes, intermediate rule nodes, and basic entity nodes that participate in the inference path to form a directed acyclic subgraph; and determining the quantifiable parameters associated with all nodes and edges in the subgraph as the set of knowledge units to be calibrated.

6. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, In step S7, the sample weights in the weighted least squares regression method The calculation formula is: in, This is the time decay factor, with a value of 0.02 / day; This represents the number of days since the collection time of the k-th data sample was reached. The reliability coefficient corresponding to the accuracy level of the sensor used to collect this data; The historical data consistency score for the plant individual to which the data point belongs is obtained by calculating the residual standard deviation of recent similar operation effect data, and then normalizing it and mapping it to the interval [0.5, 1.0].

7. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, The consistency verification in step S8 includes: semantic consistency check: verifying whether the updated attribute values ​​conform to the predefined data type, value range and unit system through the ontology constraint checker; logical conflict detection: performing a full combination precondition test on the updated rule set through the rule conflict detector to check whether there are rules that lead to contradictory conclusions; numerical boundary compliance judgment: ensuring that the calibrated parameter values ​​do not exceed the survival threshold of the corresponding plant species based on the preset plant physiological limit parameter table.

8. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, In step S9, the confidence decay function is: ,in At the current confidence level, The daily decay rate is the time elapsed since the last confidence reset or confirmation operation, expressed in days. The value is set according to the knowledge unit type: for environment-responsive rules, The value is 0.05 per day; for inherent plant properties, The value is 0.01 per day; for expert experience rules, The value is 0.03 per day.

9. The knowledge calibration method for a smart garden knowledge system according to claim 1, characterized in that, In step S3, the collected plant physiological indicators include at least the change in transpiration rate calculated by canopy temperature inversion, chlorophyll fluorescence parameters, and the lesion expansion area obtained by image recognition.

10. A knowledge calibration system for a smart garden knowledge system, characterized in that, include: The multi-source data acquisition module is used to acquire real-time environmental perception data, expert rule data, and historical maintenance operation feedback data; the knowledge reasoning and instruction generation module is used to call the knowledge graph to make reasoning decisions based on the real-time environmental perception data and the expert rule data, and generate maintenance operation instructions; the effect data acquisition module is used to collect the actual operation effect data generated after the maintenance operation instructions are executed. The knowledge deviation calculation module is used to calculate the knowledge deviation metric between the actual running effect data and the expected effect benchmark value of the corresponding reasoning path in the knowledge graph based on the preset multidimensional weighted Euclidean distance formula; the calibration trigger judgment module is used to compare the knowledge deviation metric with the dynamic calibration trigger threshold, and start the knowledge calibration process when it exceeds the threshold. The influence range analysis module is used in the calibration process to perform influence range analysis on the knowledge graph based on the graph traversal algorithm to determine the set of knowledge units to be calibrated that are associated with the current deviation. The knowledge parameter re-estimation module is used to re-estimate the parameters of the knowledge unit to be calibrated based on the historical maintenance operation feedback data and the actual operation effect data, using a weighted least squares regression method to generate calibrated knowledge parameters; the consistency verification and update module is used to verify the semantic, logical, and numerical boundary consistency of the calibrated knowledge parameters, and update the knowledge graph after the verification is passed; the knowledge confidence management module is used to implement time decay-based dynamic confidence management for all knowledge units in the knowledge graph, and mark them as needing calibration when the confidence level is lower than the lower limit.

Images