Intelligent Regulation Method for Urban Agricultural Crop Growth Environment Based on AI Algorithm

Abstract

This invention belongs to the field of pre-diagnosis and health management technology, specifically involving an intelligent regulation method for the growth environment of urban agricultural crops based on AI algorithms. This method acquires environmental measurement data, crop response data, and environmental control action records within the urban agricultural planting area, aligns them temporally according to crop growth stages, and forms a continuous state sequence. Based on the continuous state sequence, an environmental-driven response benchmark model is established, generating a crop response prediction interval, and constructing measurement-layer residuals, crop response-layer residuals, and control execution-layer residuals accordingly. The three types of residuals are coupled and discriminated to obtain pre-diagnosis and health management results that include measurement drift, environmental disturbances, insufficient execution response, or accumulated crop health risks, and environmental regulation instructions are generated. This method can distinguish between measurement bias, actual disturbances, and abnormal crop responses, forming a closed loop in the growth environment measurement, pre-diagnosis, and regulation processes.

Description

Technical Field

[0001] This invention belongs to the field of pre-diagnosis and health management technology, specifically involving an intelligent regulation method for the growth environment of urban agricultural crops based on AI algorithms. Background Technology

[0002] Urban agricultural crop growth environment regulation typically relies on multi-parameter measurement and closed-loop control processes. Data such as air temperature, air humidity, light intensity, carbon dioxide concentration, substrate moisture content, nutrient solution pH, and electrical conductivity within the planting area are uploaded at fixed sampling intervals. The control unit compares the collected values ​​with preset target ranges and then triggers control actions such as ventilation, supplemental lighting, shading, irrigation, nutrient solution adjustment, or carbon dioxide replenishment based on the deviation. To avoid frequent actions caused by fluctuations in a single sampling, conventional methods perform moving averages, outlier removal, upper and lower limit hysteresis, and time-segmented target switching on the collected data, storing the target ranges corresponding to different growth stages in a rule table. In scenarios approaching the boundary of the target range, the control unit typically suppresses repeated start-stop cycles by extending the sampling period, increasing the action interval, or setting a recovery interval. The technical approach of this type of solution uses whether the measured value exceeds the limit as the basis for judgment. The correlation between measured objects is mostly used for display, alarms, or generating historical curves, and rarely participates in the identification of anomaly sources or establishes a calculable correspondence with crop response changes. Even if historical control actions are recorded, the records are mostly kept as traceability information and are difficult to reflect the temporal relationship between control actions, environmental feedback and crop state changes.

[0003] Approximate AI-based control schemes incorporate crop response data, such as crop images, canopy coverage, leaf color index, transpiration-related estimates, or root zone absorption status, into the aforementioned closed-loop control, and train predictive models using historical samples. A common approach is to use environmental measurement data as input and crop growth level, water shortage risk, or light suitability as output, or to use crop response data as auxiliary variables to refine the environmental target range. After model training, the control unit obtains a risk score based on the current input, and then maps the score to irrigation amount, supplemental lighting duration, or ventilation intensity. For ease of deployment, existing schemes typically concatenate different data sources into a feature vector within the same sampling period, directly superimposing the prediction results onto the threshold control rules, forming a control method that combines scoring and rules. This approach can utilize more data, but the model output is usually a state classification or risk score, lacking a hierarchical expression of measurement bias, environmental disturbances, insufficient control execution, and crop physiological deviations. Crop response data is often used as diagnostic input rather than as a benchmark for verifying environmental measurements and control feedback. The model also struggles to explain whether anomalies are caused by measurement source drift, changes in the real environment, or insufficient feedback from control actions. Subsequent regulation still mainly relies on existing rules to correct current deviations.

[0004] In pre-diagnosis and health management technologies, existing urban agriculture platforms typically file separate records for environmental anomalies, equipment malfunctions, and crop growth anomalies. Environmental anomalies are defined as values ​​exceeding upper or lower limits, equipment malfunctions as failure to reach a predetermined feedback range after a control command is issued, and crop growth anomalies are based on image recognition results or manual records. Health management records often consist of timestamps, anomaly types, durations, and processing records for subsequent querying, statistics, and maintenance. Some solutions adjust alarm levels based on historical anomaly frequencies or merge multiple consecutive anomalies into a single event to avoid duplicate alarms. However, these pre-diagnosis methods still rely heavily on post-event status records, lacking a unified residual structure between environmental data, control action data, and crop response data, and lacking a traceable propagation sequence among anomaly records. When environmental measurement fluctuations, control feedback lags, and crop response deviations occur simultaneously, existing records can only store multiple anomaly events side-by-side, making it difficult to determine the earliest dominant anomaly source. Since data from different sources are not causally correlated within the same time frame, health management records cannot reverse the sample weights of the prediction model and the order of invocation of control strategies, and subsequent sampling cycles still operate according to the original logic.

[0005] The main technical problem arising from this is that existing methods for regulating the growth environment of urban agricultural crops lack a multi-layered residual coupling discrimination mechanism based on crop response. This makes it difficult to link environmental measurements, crop response changes, and control action feedback as verifiable sources of anomalies within the same measurement and processing chain. Since control decisions are primarily based on the current deviation between environmental parameters and target ranges, when crop responses deviate from the normal growth trajectory but environmental parameters have not yet exceeded limits, the regulation method cannot provide a pre-diagnosis before anomalies appear. When environmental measurements drift, jump, or deviate in isolation, the regulation method easily treats the measurement deviation as a real environmental change and generates irrigation, supplemental lighting, ventilation, or nutrient regulation actions accordingly. Furthermore, when control actions have been issued but environmental feedback is insufficient, the regulation method also struggles to distinguish between insufficient execution response and accumulated crop health risks. The common cause of the above situations is not insufficient sampling accuracy, but the lack of a unified construction, temporal alignment and source discrimination of measurement layer residuals, crop response layer residuals and control execution layer residuals. This makes it impossible to form a stable logical closed loop between the pre-diagnosis results and the control instructions. As a result, the control of the growth environment lacks a reliable basis when real disturbances, measurement biases and crop health changes are intertwined, and it is ultimately difficult to maintain the consistency of crop growth environment measurement and control process. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent regulation method for the growth environment of urban agricultural crops based on AI algorithms, which can effectively solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart regulation method for urban agricultural crop growth environment based on AI algorithm includes: acquiring crop growth environment measurement data, crop response data and environmental control action records within urban agricultural planting areas; The crop growth environment measurement data, crop response data, and environmental control action records are time-aligned according to the crop growth stage to form a continuous state sequence. An environmental-driven response baseline model is established based on the continuous state sequence to generate crop response prediction intervals; Based on the actual measurement results and the crop response prediction interval, measurement layer residuals, crop response layer residuals, and control execution layer residuals are generated. The three types of residuals are coupled and discriminated to obtain pre-diagnosis and health management results; Environmental control instructions are generated based on the pre-diagnosis and health management results.

[0008] Preferably, the time alignment according to crop growth stages includes: writing a growth stage identifier, a sampling source identifier, and a control action identifier for each collection time. Air temperature and humidity, light intensity, carbon dioxide concentration, substrate moisture content, nutrient solution status, and crop canopy phenotypic response were placed in the same time window; For missing values, abrupt changes, and duplicate values ​​within the time window, separate checks are established for missing data placeholders, proximity correlation, and source consistency. The correspondence between the start and end times of the control action and the environmental measurement feedback time is marked, and multiple time windows spanned by the same control action are merged into a continuous action segment, forming a continuous state sequence containing environmental state, historical disturbances and action records.

[0009] Preferably, the establishment of the environment-driven response baseline model includes: extracting environmental variable change segments, control action duration segments, and crop response change segments from the continuous state sequence; According to the growth stage, the environmental variable change segment and the control action duration segment are fused into an environmental driving vector; The crop response change segments were used as supervised objects to train the time-series prediction model; Preserve the preceding environmental disturbance sequence, the preceding control action sequence, and the intra-stage sequence identifier in the training samples; The crop response prediction interval is constructed based on the response prediction value, response prediction slope, and prediction fluctuation boundary output by the time series prediction model.

[0010] Preferably, the coupling and discrimination of the three types of residuals includes: constructing a residual propagation sequence according to the generation time of the measurement layer residual, the crop response layer residual, and the control execution layer residual; Calculate the sequential relationship, unidirectional change relationship, and persistence relationship among different residuals; Anomaly source tags are generated based on the sequential relationship, the same-direction change relationship, and the persistence relationship; The anomaly source tags within multiple consecutive time windows are merged into an associated status segment, and a start and end time index is configured for the associated status segment. The anomaly source tags include measurement drift, environmental disturbance, insufficient execution response, and accumulated crop health risk, and a confidence level is configured for each anomaly source tag.

[0011] Preferably, the formation of the continuous state sequence further includes: constructing a homogeneous continuity marker and a heterogeneous consistency marker based on the sampling source identifier; Perform horizontal consistency filtering on interconnected environmental measurement data within the same time window, and perform vertical continuity filtering on data from the same source within adjacent time windows; Data that passed the vertical continuity screening but failed the horizontal consistency screening was marked as a local deviation segment, and data that passed the horizontal consistency screening but failed the vertical continuity screening was marked as a suspected jump segment. Update the data trust markers in the continuous state sequence according to the local deviation segment and the suspected jump segment, respectively, and write the corresponding time window index for the data trust markers.

[0012] Preferably, the generation of the measurement layer residual, the crop response layer residual, and the control execution layer residual includes: using the difference between the actual environmental measurement value and the inferred value of the adjacent environmental variable as the measurement layer residual; The degree of deviation between the actual crop response value and the predicted crop response range is used as the crop response layer residual; The deviation between the actual environmental changes corresponding to the control action record and the action response prediction range given by the environmental drive response benchmark model is used as the control execution layer residual. A unified time window index is used for the three types of residuals, and stage identifiers, time window identifiers, and action segment identifiers are added respectively, wherein the action segment identifier is determined by the start time and end time in the control action record.

[0013] Preferably, the generation of the anomaly source label includes: inputting the residual propagation sequence into a causal attention network; In the causal attention network, measurement residual nodes, response residual nodes, and execution residual nodes are respectively established; Dominant residual nodes are selected based on the time-related weights between nodes; When the time-pointing weights of multiple residual nodes satisfy the same propagation direction, the residual node that appears earliest and continues to span adjacent time windows is retained as the dominant residual node. A horizontal consistency marker is constructed based on the correlation verification results between measurement data of different environments within the same time window, and a vertical continuity marker is constructed based on the continuous verification results of the same measurement object within adjacent time windows. The dominant residual node is jointly encoded with the horizontal consistency marker, the vertical continuity marker, and the growth stage identifier to generate an anomaly source label with confidence level.

[0014] Preferably, the generation of the environmental control instruction includes: selecting a control constraint strategy based on the anomaly source label; When the source of the anomaly is labeled as measurement drift, the associated measurement data is deweighted from the calculation of the control quantity, while the neighboring variable constraints composed of associated environmental measurement data within the same time window are retained. When the abnormal source label is environmental disturbance, the segmented control amount is calculated based on the disturbance duration and the crop response layer residual. When the anomaly source label is insufficient execution response, the duration of the control action is adjusted according to the control execution layer residual; When the abnormal source label is the accumulation of crop health risks, a health protection-type regulatory instruction is generated that limits the continuous increase of the control quantity in the same direction and introduces the priority of the response layer residual.

[0015] Preferably, the update of the anomaly source label with confidence level includes: constructing a label trajectory based on the hold-up state, transition state and decay state of the dominant residual node in multiple consecutive time windows; The label trajectory is matched with the residual change segments of the subsequent crop response layer; When the label trajectory and the residual change segment of the subsequent crop response layer have a corresponding delay in the same direction, the confidence level of the corresponding abnormal source label is increased. When the aforementioned delay correspondence does not exist, the confidence level of the corresponding anomaly source label is lowered and a verification mark is generated; The verification marker is bound to the horizontal consistency marker, the vertical continuity marker, and the growth stage identifier of the corresponding time window.

[0016] Preferably, the formation of the pre-diagnosis and health management results includes: writing the abnormality source label, the confidence level, the label trajectory, the measurement layer residual, the crop response layer residual, the control execution layer residual, and the generated environmental regulation instructions into the crop health status file; According to the recurrence position of the same abnormal source label, the occurrence position of the verification mark, and the residual decay order in the crop health status file, update the training sample weights of the environment-driven response benchmark model and the calling order of the regulation constraint strategy, and associate the updated training sample weights and calling order with the continuous state sequence of the next sampling period.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. After adopting the AI-based intelligent regulation method for urban agricultural crop growth environment, environmental measurement data, crop response data, and control action records are placed in the same continuous state sequence. Air temperature and humidity, light intensity, carbon dioxide concentration, substrate water content, nutrient solution status, and canopy phenotypic response have a unified time window index and growth stage identifier. The environment-driven response benchmark model generates crop response prediction intervals based on historical environmental disturbances and control actions. The deviation between the actual crop response and the prediction interval is transformed into crop response layer residuals. The difference between environmental measurements and associated inferred values ​​is transformed into measurement layer residuals. The deviation between the actual environmental feedback corresponding to the control action and the action response prediction interval is transformed into control execution layer residuals. After the three types of residuals enter the same residual propagation sequence, anomaly source labels can be determined according to the time of generation, unidirectional change relationship, and persistence relationship, and confidence levels can be configured. Therefore, when environmental parameters have not exceeded limits but crop responses have deviated from the normal trajectory, the control method can generate pre-diagnosis and health management results based on the response layer residuals. When measured values ​​experience isolated drifts or jumps, the control method can reduce the impact of abnormal measurement data on control quantity calculations based on the measurement layer residuals and the verification results of related variables. When control action feedback is insufficient, the control method can adjust the duration of actions based on the control execution layer residuals, avoiding misjudging insufficient execution as a risk to crop health. This processing link transforms environmental measurements, crop responses, and control feedback from parallel data into a comparable residual structure, so that control instructions are no longer triggered solely by single measurement deviations, but are constrained by the source of the anomaly, confidence level, and crop health status. Logically, this reduces the probability of erroneously triggering irrigation, supplemental lighting, ventilation, or nutrient regulation, and allows pre-diagnosis results to directly enter the next step of environmental control calculations, maintaining consistency between measurement processing, source identification, and control output.

[0018] 2. Growth stage identifiers, sampling source identifiers, control action identifiers, action duration segments, and data reliability markers in continuous state sequences enable data from different sources to undergo missing measurement placement, proximity correlation verification, source consistency verification, lateral consistency screening, and longitudinal continuity screening before entering the model. After local deviation segments and suspected jump segments are written into the corresponding time windows, anomalous measurement data are not simply deleted, nor do they directly participate in the calculation of control quantities. Instead, they participate in subsequent residual generation and anomalous source identification based on their reliability. A causal attention network forms time-directed weights between measurement residual nodes, response residual nodes, and execution residual nodes, and combines these with lateral consistency markers, longitudinal continuity markers, and growth stage identifiers to generate anomalous source labels. This allows the maintenance, transfer, and decay states of the same anomalous event across multiple time windows to form label trajectories. After the label trajectory is matched with subsequent crop response layer residual change segments, the confidence level can be updated according to the delay correspondence, and the verification marker is bound to the data verification marker of the corresponding time window. The crop health status file stores anomaly source labels, confidence levels, label trajectories, three types of residuals, and environmental control instructions. It updates training sample weights and the order of control constraint strategy invocation based on the recurrence locations of similar anomalies, the locations of verification markers, and the residual decay order. Thus, historical pre-diagnosis results can be transformed into input constraints for the next sampling period, reducing the recurrence of similar misjudgments in continuous control processes. For data fluctuations caused by sampling noise, control feedback lag, or stage switching, this processing method retains traceable data markers and focuses model updates on validated residual chains. It avoids treating short-term isolated fluctuations as normal samples for continuous learning and avoids excluding deviations in actual crop responses as simple measurement errors, thus stabilizing the closed-loop relationship between environmental measurement, health management, and control strategy updates in the data processing chain. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the intelligent regulation method for urban agricultural crop growth environment based on AI algorithm of the present invention; Figure 2 This is a flowchart of the continuous state sequence formation and data trust tag generation process of the present invention; Figure 3 This is a flowchart illustrating the coupling and discrimination of three types of residual generation and anomaly sources in this invention. Figure 4 This is a flowchart illustrating the pre-diagnosis result constraint environment control and health status record update process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please refer to Figure 1 This embodiment provides an AI-based intelligent regulation method for the growth environment of urban agricultural crops. It is used to perform homogenization, residual expression, and source identification on environmental measurement data, crop response data, and environmental control action records within urban agricultural planting areas. The environmental measurement data includes air temperature and humidity, light intensity, carbon dioxide concentration, substrate moisture content, and nutrient solution status. The crop response data includes at least one of canopy cover changes, leaf color changes, transpiration-related responses, or root zone absorption status. The environmental control action records include at least one action type (irrigation, supplemental lighting, ventilation, shading, carbon dioxide supplementation, or nutrient solution adjustment), the start time of the action, the end time of the action, and the target object of the action. During processing, data from different sources are categorized according to crop growth stages. The data is written into a unified timeline to form a continuous state sequence. Then, the environment-driven response benchmark model generates crop response prediction intervals based on environmental conditions and historical action records. The actual measured values, actual crop response values, and action feedback values ​​are converted into measurement layer residuals, crop response layer residuals, and control execution layer residuals, respectively. After coupling and discrimination, the three types of residuals form pre-diagnosis and health management results. The pre-diagnosis and health management results directly constrain the generation of environmental control instructions, so that measurement deviations, real environmental disturbances, insufficient execution responses, and accumulated crop health risks are distinguished in the same processing link. The advantage of this embodiment is that it can transform the parallel collected data into a comparable, traceable, and controllable residual structure, avoiding triggering control actions based solely on the out-of-bounds behavior of a single environmental parameter.

[0022] In this embodiment, the continuous state sequence uses a time window as the smallest processing object. The time window is not limited to a fixed hardware sampling period, but is triggered by environmental measurement updates, crop response updates, and changes in control action state. When a control action spans multiple time windows, the continuous state of the control action within each time window, the remaining time marker of the action, and the action object marker are written into the sequence, enabling the subsequent model to identify the delay correspondence between the control action and the environmental feedback. The construction of the continuous state sequence is expressed by the following formula: ;

[0023] in, Indicates the first A continuous state vector for each time window. Indicates the first Environmental measurement vectors within a time window Indicates the first Crop response vector within a time window Indicates from the first The time window to the first A set of control action segments between time windows Indicates the first The growth stage identifier corresponding to each time window. Indicates the first Data reliability markers for each time window This indicates the number of time windows to look back; for example, the environmental measurement vector within a certain time window is taken as a unit value. The crop response vector is taken as a unit value. By tracing back two time windows, a set of control action segments can be obtained. Growth stage markers Data trust label Then the continuous state vector is written as The action segments in this vector, together with the trusted markers, define the input source for subsequent residual calculations.

[0024] Specifically, the environment-driven response baseline model adopts a time-series prediction structure with stage labels. The training samples consist of environmental variable change segments, control action duration segments, and crop response change segments extracted from historical continuous state sequences. The model input is not a single time window of static measurements, but rather a sequence segment containing preceding perturbations and preceding actions. The model output includes predicted crop response values, predicted response slope, and predicted fluctuation boundaries. In the training data organization, environmental variable change segments are used to express the direction of change in temperature, humidity, light intensity, carbon dioxide concentration, substrate water content, and nutrient solution state; control action duration segments are used to express the delayed impact of regulatory actions on subsequent environmental conditions; and crop response duration segments are used to express the delayed impact of regulatory actions on subsequent environmental conditions. The response change segment is used to limit the model's learning object, and the growth stage identifier is used to distinguish the different strengths of the influence of environmental drivers on crop response in the seedling stage, vegetative growth stage, flowering and fruiting stage, or other production management stages. After training, the model inputs a continuous state vector in the current time window and outputs the corresponding crop response prediction interval. When the actual crop response falls into the prediction interval, it is marked as a stage-consistent response. When the actual crop response deviates from the prediction interval, a response layer residual is formed. The advantage of this embodiment is that by using environmental drivers, action history, and growth stage to jointly limit the reasonable range of crop response, the crop state is no longer just an alarm input, but becomes a benchmark for verifying environmental measurements and action feedback.

[0025] In a preferred embodiment, reference Figure 2For each data acquisition point, time alignment processing writes a growth stage identifier, a sampling source identifier, and a control action identifier. The growth stage identifier is determined from the cultivation plan, historical growth days, and crop response trends. The sampling source identifier distinguishes between air environment data, root zone environment data, nutrient solution data, and image or phenotypic data. The control action identifier records the action type, action object, and action duration. When air temperature and humidity, light intensity, carbon dioxide concentration, substrate moisture content, nutrient solution status, and crop canopy phenotypic response fall within the same time window, the processing end writes a missing value into a missing value placeholder, directly overwriting the original gap without using values ​​from adjacent time windows. Abrupt values ​​are handled by proximity filtering. The system performs a joint check to determine whether other environmental variables associated with the abrupt change data within the same time window exhibit directionally compatible changes. It also performs a source consistency check on duplicate values ​​to determine whether data from the same source remains unchanged for an extended period while adjacent source data has changed within a continuous time window. The time between the start and end of the control action and the environmental measurement feedback time is marked as the pre-action, during-action, and post-action feedback segments. When the same control action spans multiple time windows, it is merged into a single action duration segment. This ensures that the continuous state sequence simultaneously includes environmental states, historical disturbances, and action records. The advantage of this embodiment is that it reduces misalignment caused by different data update frequencies, allowing control actions and environmental feedback to be associated within the same time frame. The continuous state sequence fields and their processing meanings are shown in Table 1.

[0026] Table 1. Fields and Processing Meanings of Continuous State Sequence Growth stage markers Cultivation Programs and Crop Response Trends Write to each time window Distinguishing environmental response boundaries at different stages Sampling source identification Environmental Measurement and Crop Response Data Collection Write by data category Supports horizontal consistency and vertical continuity filtering Control Action Icon Control command execution record Write according to the start and end times of the action. Constitutes a continuous action segment Missing test placeholder Original collection records Preserve the gap position To avoid missing values ​​being mistaken for true stable values Data Trust Marking Consistency and continuity verification results Update by time window Participating in residual generation and control quantity weight reduction Action feedback segment Correspondence between control action and measurement feedback Mark before action, during action, after action Support residual calculation at the control execution layer

[0027] Furthermore, data credibility markers are generated jointly by horizontal consistency screening and vertical continuity screening. Horizontal consistency screening compares interrelated environmental measurement data within the same time window. For example, a decrease in air humidity after ventilation is usually accompanied by changes in carbon dioxide concentration and temperature. If a single data point shows a sudden change while neighboring variables do not show directionally compatible changes, the data point is marked as a local deviation segment. Vertical continuity screening compares the continuity of changes in the same measurement object within adjacent time windows. If data from the same source shows a jump between adjacent time windows, but related environmental variables within the same time window still maintain a reasonable correlation, the data point is marked as a suspected jump segment. After both types of markers are written into the continuous state sequence, subsequent residual calculations do not delete the original measurement values, but retain the original values, verification markers, and credibility, enabling anomaly source identification to distinguish between real environmental changes and measurement source deviations. The measurement layer residuals are calculated using the following formula: ;

[0028] in, Indicates the first Within the first time window Measurement layer residuals of an environmental measurement object Represents actual environmental measurements. This represents a reference value inferred from the correlation of environmental variables within the same time window and the continuity between adjacent time windows. Indicates the first The permissible fluctuation scale of an environmental measurement object at its current growth stage. This indicates a small positive number that prevents the denominator from being zero; for example, a unitized measurement of air temperature within a certain time window. Reference values ​​inferred from adjacent variables The allowable fluctuation scale at the current stage , ,but The results, along with the markers of local deviation segments or suspected jump segments, are written into the measurement layer residual record.

[0029] In this embodiment, the training of the environment-driven response baseline model does not employ isolated sample splicing. Instead, it extracts sequence segments with the same growth stage identifier from a continuous state sequence, retaining the preceding environmental disturbance sequence, the preceding control action sequence, and the intra-stage sequence identifier in each sample. Environmental variable change segments can be represented using adjacent time window differences, cumulative deviation, and change direction encoding. Control action duration segments can be represented using action presence state, action duration window number, and action object encoding. Crop response change segments can be represented by changes in canopy cover, leaf color, transpiration-related response, or root zone absorption state. The time-series prediction model can adopt a gated recurrent structure, a time-series convolutional structure, or an attention-based time-series structure, as long as it can receive sequence segments of fixed or variable length and output the crop response prediction interval. During training, the sample weights are jointly determined by the data credibility markers and the verification markers in the health status profile. Samples with a large number of missing test slots and that have not passed the consistency check have reduced participation intensity, while residual chain samples that have been verified by subsequent crop responses have increased participation intensity. The advantage of this embodiment is that it focuses the model's learning object on the temporal relationship between environmental disturbances, control actions, and crop responses, rather than performing static fitting on a single acquisition moment.

[0030] Preferably, the crop response prediction interval is constructed by the predicted response value, the predicted fluctuation boundary, and the stage-allowed fluctuation scale. The response layer residual is generated only when the actual crop response exceeds the prediction interval, thus avoiding misjudging natural fluctuations within the interval as health risks. The response layer residual is calculated using the following formula: ;

[0031] in, Indicates the first Within the first time window Crop response layer residuals of each crop response object This represents the actual crop response value. This represents the lower bound of the prediction interval output by the environment-driven response baseline model. This represents the upper bound of the prediction interval output by the environment-driven response baseline model. Indicates the first The response scale of a crop at its current growth stage. This indicates taking the maximum value within the parentheses; for example, the normalized actual value of the canopy response. Lower bound of the prediction interval upper bound of the prediction interval Response scale ,but This residual indicates that the crop response is below the reasonable range for the stage, which is derived from both environmental drivers and action history.

[0032] Furthermore, the control execution layer residual is used to describe the deviation between the actual environmental changes and the predicted action response after the control action is issued. The control action record includes the action type, action start time, action end time, and action target object. The action response prediction interval is generated by the environment-driven response baseline model or action feedback sub-model based on the current environmental state, action duration, and growth stage identifier. For irrigation actions, the target feedback object may include substrate moisture content and changes in air humidity. For supplemental lighting actions, the target feedback object may include changes in light intensity and temperature-related changes. For ventilation actions, the target feedback object may include the linked changes in air humidity, temperature, and carbon dioxide concentration. The specific object is determined based on the existing control and measurement objects. The control execution layer residual is calculated using the following formula: ;

[0033] in, Indicates the first Within the first time window The control execution layer residual of each action feedback object This represents the measurement value of the feedback object corresponding to the time window when the action begins. Indicates the process Feedback object measurement values ​​after a time window Indicates action type The predicted change that should occur under the current conditions. Indicates the first The permissible feedback fluctuation scale for an action feedback object The meaning is the same as above; for example, the unitized value of air humidity before ventilation. After two time windows Predicting changes Allow feedback fluctuation scale , ,but The result is written into the control execution layer residual record corresponding to the action segment.

[0034] In a preferred embodiment, reference Figure 3 The coupling discrimination of the three types of residuals takes the residual propagation sequence as input. The residual propagation sequence consists of measurement layer residuals, crop response layer residuals, and control execution layer residuals arranged according to their generation time. Each residual is assigned a time window identifier, growth stage identifier, action segment identifier, data credibility marker, and source category identifier. The processing end calculates the sequential relationship, unidirectional change relationship, and persistence relationship between different residuals. The sequential relationship is used to determine the order in which the residuals appear on the time axis. The unidirectional change relationship is used to determine whether the direction of residual change conforms to the directional constraint of the environment-driven crop response. The persistence relationship is used to determine whether the residuals span adjacent time windows rather than appearing at a single point. When the measurement layer residual appears first and is horizontal... When the consistency marker and the longitudinal continuity marker are not valid, the anomaly source label tends to indicate measurement drift; when the measurement layer residual appears first and forms a directionally compatible change with multiple environmental variables, and the crop response layer residual appears subsequently, the anomaly source label tends to indicate environmental disturbance; when control action records exist but the action feedback object does not produce a corresponding change, the anomaly source label tends to indicate insufficient execution response; when the crop response layer residual continuously expands while the measurement layer residual is small or interpretable, the anomaly source label tends to indicate the accumulation of crop health risks. The advantage of this embodiment is that it limits the anomaly source by the order of residual occurrence and the residual linkage relationship, reducing the discrimination confusion caused by multiple anomaly events being recorded in parallel.

[0035] Preferably, the causal attention network is used to establish time-directed weights between measurement residual nodes, response residual nodes, and execution residual nodes. These nodes are not hardware nodes, but rather data objects in the residual propagation sequence. The time-directed weights are calculated jointly by the source node residual magnitude, source node duration, target node residual magnitude, the time interval between the two nodes, and the growth stage identifier. The causal attention network does not allow future time window data to influence past time window nodes. During calculation, connections between nodes that do not satisfy the temporal order are masked, ensuring that the dominant residual node can only originate from the current time window or an earlier time window. The time-directed weights are calculated using the following formula: ;

[0036] in, Indicates the first Residual nodes within a time window Pointing to residual nodes Time-directed weights This represents the directional score calculated from the residual magnitude, duration, time interval, and growth stage identifier. This indicates that pointing to residual nodes is allowed. The set of candidate source nodes; for example, if the set of candidate source nodes contains three candidate source nodes—measurement residual node, execution residual node, and response residual node—with pointing scores of 1.2, 0.3, and -0.2 respectively, then the weight corresponding to the measurement residual node is... The processing end retains candidate source nodes with higher weights that continuously span adjacent time windows as dominant residual nodes.

[0037] In this embodiment, the generation of anomaly source labels is based on the dominant residual node, combined with lateral consistency markers, vertical continuity markers, and growth stage identifiers for joint encoding. When the dominant residual node is a measurement residual node, and the lateral consistency marker shows that adjacent environmental variables have not undergone compatible changes, and the vertical continuity marker shows that there is a suspected jump segment for the same measurement object, a measurement drift label is generated. When the dominant residual node is a measurement residual node, the lateral consistency marker shows that multiple related environmental variables change together, and subsequent response residual nodes appear in a stage-wise direction, an environmental disturbance label is generated. When the dominant residual node is an execution... When an action feedback object does not change in the predicted direction within the duration of the action, an insufficient execution response label is generated. When the dominant residual node is a response residual node, and the response residual remains or expands in multiple time windows, and the measurement residual cannot explain the change, a crop health risk accumulation label is generated. Each anomaly source label is configured with a confidence level, which is jointly determined by the weight of the dominant residual node, the duration of the residual, the data credibility marker, and the subsequent response verification results. The advantage of this embodiment is that the anomaly source is no longer determined by a single threshold, but is jointly limited by the residual chain, time direction, and data verification.

[0038] In a preferred embodiment, reference Figure 4The generation of environmental control commands uses anomaly source labels and confidence levels as constraint inputs, and retains the mapping relationship between the controlled object and the residual object. When the anomaly source label is measurement drift, the processing end reduces the weight of associated measurement data in the calculation of control quantities, and retains the neighboring variable constraints formed by associated environmental measurement data within the same time window. For example, when a single air humidity data is marked as a suspected jump segment, the ventilation or humidification control quantity is not directly calculated based on this single data, but is combined with temperature, carbon dioxide concentration, substrate moisture state, and crop response residuals as constraints. When the anomaly source label is environmental disturbance, the processing end... The duration of the disturbance and the residual of the crop response layer form a segmented control quantity to avoid reverse fluctuations caused by a one-time control action. When the anomaly source label is insufficient execution response, the processing end adjusts the duration of the control action segment according to the control execution layer residual and records the objects with insufficient action feedback. When the anomaly source label is accumulated crop health risk, the processing end generates a health protection-type control instruction that restricts the continuous increase of the control quantity in the same direction and introduces the priority of the response layer residual. The advantage of this embodiment is that different anomaly sources correspond to different control constraints, and the generation basis of the control instruction is expanded from environmental parameter deviation to the joint constraint of residual source and crop response. The mapping between anomaly source labels and control constraints is shown in Table 2.

[0039] Table 2 Mapping of Anomaly Source Labels and Regulatory Constraints Measurement drift Measurement layer residual lateral consistency marking and longitudinal continuity marking Reduce the intensity of suspected measurement data participation while retaining neighboring variable constraints Suspected Source Time Window Marker Type Environmental disturbance Measurement layer residual crop response layer residual duration Segmented control quantities are formed based on the duration of the disturbance. Direction of propagation time residual of the perturbation object Insufficient execution response Control execution layer residual action fragment identifier Adjusting the continuous motion segment recording feedback for insufficient objects Action type feedback object deviates from direction Accumulated crop health risks Crop Response Layer Residual Growth Stage Identifier Tag Track Limiting the continuous increase of control input in the same direction and introducing response layer residual priority Confidence of the trajectory status of the response object

[0040] Furthermore, the calculation of the control amount does not directly replace the existing environmental target interval. Instead, it introduces control weights corresponding to the anomaly source labels under the constraint of the target interval. These control weights are jointly determined by confidence level, residual type, and data credibility label. For data labeled as measurement drift, the control weights are reduced and corrected by neighboring variable constraints. For data labeled as environmental disturbances, the control weights change in segments according to the duration of the disturbance. For data labeled as insufficient execution response, the control weights act on the action duration segment rather than simply increasing the control intensity. For data labeled as accumulated health risks, the control weights preferentially refer to the crop response layer residuals and growth stage identifiers, limiting the continuous increase of control amounts in the same direction. When generating the final control instruction, the processing end merges the target deviation, residual source, and control constraints into the control amount, avoiding the mutual cancellation of multiple control objects within the same time window. The advantage of this embodiment is that it transforms the pre-diagnosis results into control constraints, creating a correspondence between environmental control actions and anomaly sources.

[0041] In a preferred embodiment, the persistence, transition, and decay states of anomaly source labels across multiple consecutive time windows are constructed as label trajectories. A persistence state indicates that similar anomaly source labels persist in adjacent time windows; a transition state indicates that the dominant residual node shifts from a measurement residual node to a response residual node or from an execution residual node to a response residual node; and a decay state indicates that the residual amplitude decreases and returns to the prediction interval. The label trajectory is matched with subsequent crop response layer residual change segments. If the direction of change in the label trajectory has a delayed correspondence with the subsequent response layer residuals, the confidence level of the corresponding anomaly source label is increased; otherwise, the confidence level is decreased and a verification mark is generated. The verification mark is bound to the horizontal consistency mark, vertical continuity mark, and growth stage identifier of the corresponding time window, enabling subsequent model training to distinguish between validated residual chains and residual chains requiring verification. The confidence level update is calculated using the following formula: ;

[0042] in, Indicates the first Abnormal source tags within a time window Confidence level, This indicates the confidence level after the next time window update. This represents the historical confidence retention coefficient. This represents the validation value obtained by matching the label trajectory with the residual change segment of the subsequent crop response layer. This indicates that the result within the parentheses is limited to between 0 and 1; for example, measuring the confidence level of a drift label within the current time window. Historical confidence retention coefficient The verification value was obtained because the residual of the subsequent response layer did not change with a delay according to the measurement drift direction. ,but The updated value is written to the corresponding label trajectory.

[0043] In this embodiment, the crop health status archive is used to store anomaly source labels, confidence levels, label trajectories, measurement layer residuals, crop response layer residuals, control execution layer residuals, and generated environmental regulation instructions. The archive records are organized by a dual index of time windows and growth stages. When the same type of anomaly source label appears repeatedly in different time windows, the processing end counts its recurrence location, verification mark appearance location, and residual decay order, and updates the training sample weights and regulatory constraint strategy invocation order of the environmental-driven response baseline model accordingly. When a residual chain is verified as a real environmental disturbance by subsequent crop responses, the corresponding sample of that chain is obtained in subsequent training. Higher participation intensity; when a residual chain is marked as an isolated measurement jump by the verification mark, the suspected measurement values ​​in that chain will have reduced participation intensity in subsequent training; when a certain type of insufficient execution response label repeatedly appears on the same action object, the regulation constraint strategy will prioritize calling the action continuous segment adjustment rule in subsequent sampling cycles; the training sample weights and calling order are associated with the continuous state sequence of the next sampling cycle, so that the pre-diagnosis results are not only saved as historical records, but are written back to the next round of measurement processing and regulation calculation. The advantage of this embodiment is that it forms a closed loop between measurement processing, anomaly source identification, environmental regulation and health status archive.

[0044] Preferably, the residual decay order in the health status file is used to determine whether a certain regulatory constraint matches the source of anomaly. If the control execution layer residual decays first, the measurement layer residual decays subsequently, and the crop response layer residual decays late after the regulatory instruction is executed, the file records the anomaly source label and the corresponding regulatory constraint strategy as a verified chain. If the measurement layer residual decays after the regulatory instruction is executed but the crop response layer residual continues to expand, the file records the corresponding regulatory constraint strategy as a verification chain and increases the priority of the crop response layer residual in the next sampling period. If the control execution layer residual persists, the file binds the action feedback object to the insufficient execution response label and prioritizes adjusting the action persistence segment in subsequent regulation. Through this recording method, the training sample weight is not directly determined by the number of anomalies, but by whether the residual chain is verified by crop response changes. The calling order of the regulatory constraint strategy is not a fixed rule table, but is updated by the historical verification status and decay order of similar anomaly sources. The advantage of this embodiment is that it reduces isolated noise entering the model learning process and allows the verified pre-diagnosis chain to be used preferentially in subsequent control.

[0045] In a preferred embodiment, for measurement drift scenarios, if the processing end finds a large measurement layer residual for a single environmental measurement object in a continuous state sequence, while the associated environmental variables do not undergo compatible changes within the same time window, and the crop response layer residual does not appear in the direction of change of the single environmental measurement object, then the causal attention network takes this measurement residual node as a candidate dominant residual node and generates a measurement drift label by combining the horizontal consistency label and the vertical continuity label. When the environmental control command is generated, the suspected drift measurement value is downweighted, while the neighboring variable constraint and the crop response layer residual are retained. The control end does not directly increase irrigation, ventilation, or supplemental lighting actions based on a single jump value. If the crop response layer residual still does not show a corresponding delayed change in subsequent time windows, the label trajectory is determined to be mismatched with the crop response, the confidence of the measurement drift label is increased, and the verification label is written to the measurement source. The advantage of this embodiment is that it limits single-point measurement jumps to the data processing level and does not directly convert measurement deviations that have not been verified by environmental association and crop response into control actions.

[0046] In a preferred embodiment, for environmental disturbance scenarios, the processing end detects multiple environmental measurement objects forming directional compatible changes within the same or adjacent time windows, and the environmental-driven response baseline model predicts that the crop response should produce corresponding changes in subsequent time windows. If the actual crop response subsequently deviates from the predicted interval, the residual propagation sequence shows a propagation relationship where the measurement layer residual appears first, followed by the crop response layer residual. The causal attention network selects the earliest appearing measurement residual node that spans adjacent time windows as the dominant residual node based on the time-direction weight, and generates an environmental disturbance label. The environmental control command generates segmented control quantities based on the disturbance duration and the crop response layer residual. The segmented control quantities reread the residual change direction and data credibility marker in each time window to avoid continuing to apply unidirectional control when the disturbance has begun to subside. The advantage of this embodiment is that it can distinguish between real environmental disturbances and isolated measurement deviations, and makes the control action segmented and constrained as the residual chain changes.

[0047] In a preferred embodiment, for scenarios with insufficient execution response, the processing end reads the start time, end time, and object of the action from the control action record, and calculates the control execution layer residual in the feedback time window after the end of the action duration segment. If the control action has been recorded as an execution state, but the actual environmental change of the feedback object does not reach the action response prediction range, the execution residual node in the residual propagation sequence is retained as the dominant residual node, and the anomaly source label is generated as insufficient execution response. When the control instruction is generated, this situation is not directly equated with the accumulation of crop health risks, nor is the measurement layer residual used as the sole basis for control. Instead, the action duration segment is adjusted first, and the object with insufficient feedback is recorded. If the crop response layer residual expands as the execution residual continues to exist, the label trajectory shifts from insufficient execution response to the accumulation of crop health risks, and the health status file records this shift relationship synchronously. The advantage of this embodiment is that it can retain the source of anomalies on the execution side when the action has been issued but the environmental feedback is insufficient, reducing control fluctuations caused by repeated increases in control intensity.

[0048] In a preferred embodiment, for a scenario of accumulated crop health risks, if the processing end detects that the crop response layer residuals remain or expand across multiple consecutive time windows, and the measurement layer residuals, after verification through lateral consistency and longitudinal continuity, do not form an interpretable chain of environmental disturbances, and the control execution layer residuals do not show insufficient action feedback, then the residual propagation sequence will use the response residual node as the dominant residual node, and the anomaly source label will be generated as accumulated crop health risks. When generating control instructions, the processing end restricts the continuous increase of control quantities in the same direction and writes the priority of the crop response layer residuals into the health protection control instructions, so that irrigation, supplemental lighting, ventilation, or nutrient regulation actions are not repeatedly amplified due to deviations from a single environmental target. If the crop response layer residuals recede in subsequent time windows, the health status file records the label trajectory and the corresponding control constraint strategy as a verified chain. The advantage of this embodiment is that when environmental measurements have not significantly exceeded the limits but the crop response has deviated from the reasonable range of the stage, a traceable pre-diagnosis result can still be formed, and this result can be constrained into the control instructions.

[0049] Furthermore, in implementations where multiple anomaly sources coexist, the processing end allows multiple candidate anomaly source labels to exist within the same time window. However, the final control instruction determines the primary constraint source based on confidence level, the time sequence of dominant residual nodes, and the growth stage identifier. For example, when measurement layer residuals and control execution layer residuals appear in the same time window, if the control action record shows that the feedback object does not change in the predicted direction, and the relationship between the measurement layer residual and neighboring variables is established, then the insufficient execution response label takes priority in adjusting the action duration segment. If the measurement layer residual is an isolated jump, and the execution residual does not cross adjacent time windows, then the measurement drift label takes priority in reducing the weight of measurement data. If the response layer residual continues to expand in multiple time windows, then the crop health risk accumulation label is given higher priority, and the unidirectional superposition of control quantities corresponding to other labels is restricted. The advantage of this embodiment is that when multiple residuals occur simultaneously, the control instruction is still constrained by the residual source, time direction, and response verification, avoiding mutual interference caused by multiple control actions correcting themselves independently.

[0050] In this embodiment, both model updates and regulatory strategy updates are based on health status archives. After each sampling period, the processing end reads the generated label trajectories, confidence levels, and residual decay order to form a training sample update table and a regulatory constraint strategy call table. The training sample update table is used to adjust the participation intensity of different continuous state segments during the training of the environment-driven response benchmark model. The regulatory constraint strategy call table is used to determine the calling order of strategies corresponding to measurement drift, environmental disturbance, insufficient execution response, and crop health risk accumulation in the next sampling period. When the same type of anomaly occurs repeatedly in a certain growth stage and the subsequent crop response verification is consistent, the processing end uses the same residual chain in that stage as a high-confidence sample. When an anomaly source label is repeatedly verified and negated, the processing end writes its corresponding data source, time window, and sampling category into the low-confidence sample set. This processing method makes model updates dependent on residual chain verification results rather than simply anomaly frequency. The advantage of this embodiment is that historical pre-diagnosis results can continuously influence subsequent measurement processing, model training, and regulatory constraints, forming a closed-loop data chain for crop health management.

[0051] Specifically, the entire method maintains a traceable reference relationship between the original collected data, continuous state sequences, residual records, anomaly source labels, environmental control instructions, and health status files during runtime. Any control instruction can be traced back to the corresponding time window, data credibility marker, dominant residual node, and label trajectory. Any model training sample can be traced back to the corresponding residual chain and verification marker. In terms of data storage, the original collected data is not overwritten by the predicted values, and missing measurement occupants, local deviation segments, suspected jump segments, and verification markers are saved with additional fields. In terms of computational scheduling, the generation of continuous state sequences, the generation of crop response prediction intervals, the calculation of three types of residuals, the discrimination of residual propagation sequences, the generation of control instructions, and the updating of health status files are sequentially connected, but each processing step retains the index of the previous processing step. The advantage of this embodiment is that it enables a clear data transfer relationship between environmental measurement, pre-diagnosis and health management, control instruction generation, and model updating, and enables the intelligent control method for urban agricultural crop growth environment based on AI algorithms to be fully implemented without changing the existing collection objects and control objects.

Claims

1. A method for intelligent regulation of urban agricultural crop growth environment based on AI algorithm, characterized in that, include: Acquire crop growth environment measurement data, crop response data, and environmental control action records within urban agricultural planting areas; The crop growth environment measurement data, crop response data, and environmental control action records are time-aligned according to the crop growth stage to form a continuous state sequence. An environmental-driven response baseline model is established based on the continuous state sequence to generate crop response prediction intervals; Based on the actual measurement results and the crop response prediction interval, measurement layer residuals, crop response layer residuals, and control execution layer residuals are generated. The three types of residuals are coupled and discriminated to obtain pre-diagnosis and health management results; Environmental control instructions are generated based on the pre-diagnosis and health management results.

2. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 1, characterized in that, The time alignment according to crop growth stages includes: writing a growth stage identifier, a sampling source identifier, and a control action identifier for each collection time; Air temperature and humidity, light intensity, carbon dioxide concentration, substrate moisture content, nutrient solution status, and crop canopy phenotypic response were placed in the same time window; For missing values, abrupt changes, and duplicate values ​​within the time window, separate checks are established for missing data placeholders, proximity correlation, and source consistency. The correspondence between the start and end times of the control action and the environmental measurement feedback time is marked, and multiple time windows spanned by the same control action are merged into a continuous action segment, forming a continuous state sequence containing environmental state, historical disturbances and action records.

3. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 1, characterized in that, The establishment of the environmental-driven response baseline model includes: extracting environmental variable change segments, control action duration segments, and crop response change segments from the continuous state sequence; According to the growth stage, the environmental variable change segment and the control action duration segment are fused into an environmental driving vector; The crop response change segments were used as supervised objects to train the time-series prediction model; Preserve the preceding environmental disturbance sequence, the preceding control action sequence, and the intra-stage sequence identifier in the training samples; The crop response prediction interval is constructed based on the response prediction value, response prediction slope, and prediction fluctuation boundary output by the time series prediction model.

4. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 1, characterized in that, The coupling and discrimination of the three types of residuals includes: constructing a residual propagation sequence according to the generation time of the measurement layer residual, the crop response layer residual, and the control execution layer residual; Calculate the sequential relationship, unidirectional change relationship, and persistence relationship among different residuals; Anomaly source tags are generated based on the sequential relationship, the same-direction change relationship, and the persistence relationship; The anomaly source tags within multiple consecutive time windows are merged into an associated status segment, and a start and end time index is configured for the associated status segment. The anomaly source tags include measurement drift, environmental disturbance, insufficient execution response, and accumulated crop health risk, and a confidence level is configured for each anomaly source tag.

5. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 2, characterized in that, The formation of the continuous state sequence further includes: constructing a homogeneity marker and a heterogeneity consistency marker based on the sampling source identifier; Perform horizontal consistency filtering on interconnected environmental measurement data within the same time window, and perform vertical continuity filtering on data from the same source within adjacent time windows; Data that passed the vertical continuity screening but failed the horizontal consistency screening was marked as a local deviation segment, and data that passed the horizontal consistency screening but failed the vertical continuity screening was marked as a suspected jump segment. Update the data trust markers in the continuous state sequence according to the local deviation segment and the suspected jump segment, respectively, and write the corresponding time window index for the data trust markers.

6. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 3, characterized in that, The generation of the measurement layer residual, the crop response layer residual, and the control execution layer residual includes: using the difference between the actual environmental measurement value and the inferred value of the adjacent environmental variable as the measurement layer residual; The degree of deviation between the actual crop response value and the predicted crop response range is used as the crop response layer residual; The deviation between the actual environmental changes corresponding to the control action record and the action response prediction range given by the environmental drive response benchmark model is used as the control execution layer residual. A unified time window index is used for the three types of residuals, and stage identifiers, time window identifiers, and action segment identifiers are added respectively, wherein the action segment identifier is determined by the start time and end time in the control action record.

7. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 4, characterized in that, The generation of the anomaly source label includes: inputting the residual propagation sequence into a causal attention network; In the causal attention network, measurement residual nodes, response residual nodes, and execution residual nodes are respectively established; Dominant residual nodes are selected based on the time-related weights between nodes; When the time-pointing weights of multiple residual nodes satisfy the same propagation direction, the residual node that appears earliest and continues to span adjacent time windows is retained as the dominant residual node. A horizontal consistency marker is constructed based on the correlation verification results between measurement data of different environments within the same time window, and a vertical continuity marker is constructed based on the continuous verification results of the same measurement object within adjacent time windows. The dominant residual node is jointly encoded with the horizontal consistency marker, the vertical continuity marker, and the growth stage identifier to generate an anomaly source label with confidence level.

8. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 7, characterized in that, The generation of the environmental control instruction includes: selecting a control constraint strategy based on the anomaly source label; When the source of the anomaly is labeled as measurement drift, the associated measurement data is deweighted from the calculation of the control quantity, while the neighboring variable constraints composed of associated environmental measurement data within the same time window are retained. When the abnormal source label is environmental disturbance, the segmented control amount is calculated based on the disturbance duration and the crop response layer residual. When the anomaly source label is insufficient execution response, the duration of the control action is adjusted according to the control execution layer residual; When the abnormal source label is the accumulation of crop health risks, a health protection-type regulatory instruction is generated that limits the continuous increase of the control quantity in the same direction and introduces the priority of the response layer residual.

9. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 8, characterized in that, The update of the anomaly source label with confidence level includes: constructing a label trajectory based on the maintenance state, transition state and decay state of the dominant residual node in multiple consecutive time windows; The label trajectory is matched with the residual change segments of the subsequent crop response layer; When the label trajectory and the residual change segment of the subsequent crop response layer have a corresponding delay in the same direction, the confidence level of the corresponding abnormal source label is increased. When the aforementioned delay correspondence does not exist, the confidence level of the corresponding anomaly source label is lowered and a verification mark is generated; The verification mark is bound to the horizontal consistency mark, the vertical continuity mark, and the growth stage identifier of the corresponding time window.

10. The intelligent regulation method for urban agricultural crop growth environment based on AI algorithm according to claim 9, characterized in that, The formation of the pre-diagnosis and health management results includes: writing the anomaly source label, the confidence level, the label trajectory, the measurement layer residual, the crop response layer residual, the control execution layer residual, and the generated environmental regulation instructions into the crop health status file; According to the recurrence position of the same abnormal source label, the occurrence position of the verification mark, and the residual decay order in the crop health status file, update the training sample weights of the environment-driven response benchmark model and the calling order of the regulation constraint strategy, and associate the updated training sample weights and calling order with the continuous state sequence of the next sampling period.

Images