An edible mushroom environment data monitoring method based on reinforcement learning

Abstract

The application discloses an edible mushroom environment data monitoring method based on reinforcement learning, and relates to the field of edible mushroom environment monitoring.The method comprises the following steps: collecting and preprocessing edible mushroom environment data, regulation data and culture state data; compressing a low-response section between adjacent response boundaries to generate a dynamic response interval; determining the direction, time lag and response operator of a regulation event and an environmental parameter, and constructing a time-lag coupling graph; replacing the response operator of a current regulation event with the response operator of a historical event to deduce counterfactual response; generating a response residual, an abnormal code and a candidate monitoring operation; inputting a time-lag coupling double-constraint BDQ network to generate a monitoring strategy; and updating the time-lag coupling graph and the time-lag coupling double-constraint BDQ network.The application adopts time-lag coupling reinforcement learning to realize edible mushroom environment anomaly tracing and adaptive monitoring, and has the advantages of precision, high efficiency and stability.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring of edible fungi, and in particular to a method for monitoring environmental data of edible fungi based on reinforcement learning. Background Technology

[0002] Currently, in the field of environmental monitoring for edible fungi, existing technologies typically use sensors to collect data on temperature, humidity, carbon dioxide concentration, and culture medium parameters at regular intervals, and judge environmental anomalies based on fixed thresholds. However, such methods often process the adjustment equipment status and environmental parameters separately, making it difficult to reflect the directional, time-delay, and coupling response relationship between equipment actions and environmental changes at multiple monitoring points. This also results in insufficient accuracy in identifying low-response segments and hysteresis anomalies.

[0003] Meanwhile, existing technologies mostly adopt fixed sampling periods and preset monitoring rules. After an anomaly occurs, alarms are mainly based on a single residual or limit-breaking result. They lack mechanisms for comparing expected and counterfactual responses, tracing the source of anomalies, and selecting candidate monitoring operations. Furthermore, they do not comprehensively consider sensor output capabilities, communication bandwidth, and equipment occupancy constraints, making it difficult to continuously revise response benchmarks and update monitoring strategies based on actual monitoring results. Summary of the Invention

[0004] One objective of this invention is to propose a data monitoring method for edible fungi environment based on reinforcement learning. This invention employs time-delay coupled reinforcement learning to achieve anomaly tracing and adaptive monitoring of edible fungi environment, which has the advantages of accuracy, efficiency, and stability.

[0005] A method for monitoring edible fungi environmental data based on reinforcement learning according to an embodiment of the present invention includes: Collect and preprocess environmental data, conditioning data, and cultivation status data of edible fungi to generate monitoring data; The regulation events and event response sequences are determined from the monitoring data. The response boundaries are determined based on the event response sequences. The low-response segments between adjacent boundaries are compressed to generate a dynamic response range. Based on monitoring data and dynamic response intervals, determine the direction, time delay, and response operator of the regulation event and environmental parameters, and construct a time delay coupling diagram; Based on the time-delay coupling diagram and monitoring data, the expected response is deduced, and historical events with the same initial state at the same stage are matched. The response operators of the historical events are replaced with the response operators of the current adjustment event to deduce the counterfactual response. Compare monitoring data, expected response, and counterfactual response; perform forward verification and reverse tracing; and generate response residuals, anomaly codes, and candidate monitoring operations. The time-delay coupling graph, anomaly coding, and candidate monitoring operations are input into the coupling layer of the time-delay coupled double-constraint BDQ network to generate the coupling state. The coupling state is then input into the branch decision layer to generate the branch action values ​​for each candidate monitoring operation. Path coverage and timing consistency correction are performed. The corrected branch action values ​​are then input into the constraint layer to exclude unexecutable candidate monitoring operations. Lagrange penalties are applied to the branch action values ​​based on the constraint violations of the remaining candidate monitoring operations to generate the monitoring strategy. The monitoring strategy is executed and monitoring results are generated. Based on the monitoring results, the state drift and residual change are generated and the expected response baseline value is corrected. The time-delay coupling diagram and the time-delay coupling double-constraint BDQ network are updated accordingly.

[0006] Optionally, the environmental data includes air temperature, air humidity, carbon dioxide concentration, light intensity, and culture medium temperature, moisture content, and pH; the adjustment data includes control commands, start and stop times, set values, and actual feedback values ​​of the adjustment equipment; the cultivation status data includes the edible fungus growth stage, inoculation time, fruiting cycle, growth images, and culture medium configuration information; and the preprocessing includes time alignment, outlier removal, missing value completion, and data scaling.

[0007] Optionally, the generation of the dynamic response range includes: The control status and feedback status of the regulating equipment are read from the monitoring data. The start and end of the regulating event are determined based on the corresponding changes of the two. The equipment operations with the same area of ​​action, the same control direction and the same time are merged into the same regulating event. Based on the start time of the adjustment event, environmental parameter records before, during, and after the event are extracted, and the stable state before the event is used as the starting point for the response to generate an event response sequence. The response rate of each environmental parameter in the event response sequence is calculated according to the effective sampling time. Based on the change of the response rate relative to the historical stable fluctuation range, the effective response segment and the low response segment are divided. The first and last records, direction change records, and spatial difference records of the low response segment are retained. The remaining records are subjected to hierarchical sampling and compression, and a correspondence between the retained records and the original sampling time is established. The effective response segments and the compressed low response segments are connected according to the original time sequence, and the dynamic response interval is generated by dividing the response boundary.

[0008] Optionally, the generation of the time-delay coupling diagram includes: Generate the actual feedback sequence of the control equipment and the response sequence of environmental parameters based on monitoring data and dynamic response range; The direction of response between the regulation event and the environmental parameter is determined based on the cumulative change of the environmental parameter response sequence relative to the response starting point and the sign consistency of the response rate. Based on the actual action time of the regulating equipment, the actual feedback sequence of the regulating equipment is shifted according to the event sampling period, and the time shift correlation is calculated in combination with the response sequence of environmental parameters. The response time delay is generated according to the time offset corresponding to the peak correlation. Align the actual feedback sequence of the adjustment equipment and the response sequence of environmental parameters according to the response time delay, fit the discrete response kernel according to the change in adjustment input and the offset of environmental parameters, and generate the response operator. Using the regulation event as the event node and the environmental parameters and corresponding monitoring points as the response nodes, a directed edge is established when the response direction, response delay, and response operator meet the validity requirements. The response direction, response delay, response operator, and response confidence are written into the directed edge to construct a time-delay coupling graph.

[0009] Optionally, the deduction of the counterfactual response includes: The current adjustment input change is shifted according to the response time delay of each directed edge in the time-delay coupling diagram, and the one-sided response of each response node is calculated using the corresponding response operator. The expected response is generated based on the one-sided response. Historical events are selected from those with consistent growth stages, adjusted event types, and response nodes, and whose initial environmental parameter differences do not exceed the corresponding matching limits. Historical events with the same initial state at the same stage are also identified. Replace the response operators of each directed edge of the current adjustment event with the response operators of the corresponding directed edges in the historical events of the same stage and initial state, keeping the current adjustment input change, response nodes and graph connectivity unchanged, recalculate the one-sided response of each response node, and generate counterfactual responses based on the recalculated one-sided responses.

[0010] Optionally, the generation of the response residual, anomaly encoding, and candidate monitoring operations includes: Based on monitoring data, expected response, and counterfactual response, the differences between actual environmental parameter values ​​and expected environmental parameter values ​​and counterfactual environmental parameter values ​​are calculated respectively, generating expected residual sequences and counterfactual residual sequences; Perform amplitude verification, direction verification, and response time verification on the expected residual sequence to generate a positive verification result; For a response node that is a positive verification anomaly, remove the corresponding single-edge response one by one along the directed edge pointing to the response node in the time-delay coupling graph, compare the changes in the residual before and after removal, determine the source edge and source event of the anomaly, and generate the reverse tracing result. Based on the expected residual sequence, counterfactual residual sequence, forward verification results, and reverse tracing results, response residuals, anomaly codes, and candidate monitoring operations are generated.

[0011] Optionally, the generation of the monitoring strategy includes: The time-delay coupling graph, anomaly coding, and candidate monitoring operations are input into the coupling layer of the time-delay coupled dual-constraint BDQ network; The time-delay coupled dual-constraint BDQ network includes a coupling layer, a branch decision layer, and a constraint layer. In the coupling layer, graph relationship aggregation and exception operation fusion are performed to generate a coupled state; The coupling state is input into the branch decision layer, and action branches are established for each candidate monitoring operation. Branch action values ​​are generated for each action branch, and path coverage and timing consistency correction are performed on the branch action values. The corrected branch action value is input into the constraint layer to exclude unexecutable candidate monitoring operations. The constraint violation of the remaining candidate monitoring operations is calculated. Lagrange multipliers are used to apply penalties to the corresponding branch action value. Candidate monitoring operations are selected according to the penalized branch action value to generate a monitoring strategy.

[0012] Optionally, the updating of the time-delay coupling graph and the time-delay coupled double-constraint BDQ network includes: Each monitoring operation is scheduled according to the execution order, monitoring cycle, and duration in the monitoring strategy, and monitoring results are generated. The environmental parameter values, equipment feedback values, and calibration measurement values ​​before and after the monitoring operation are statistically analyzed, and the state drift is generated based on the change range of similar data. Based on the monitoring results, the residuals between the actual environmental parameter values ​​and the expected response are recalculated, and the residuals before and after the monitoring operation are compared to generate the residual change. When the state drift continues and the residuals do not converge, a baseline correction amount is generated based on the deviation between the monitoring results and the expected response, and the baseline correction amount is used to correct the baseline value of the expected response. The corrected expected response baseline, state drift, and monitoring results are written into the time-delay coupling graph. Training samples are constructed based on the coupling state before and after the monitoring strategy is executed, the monitoring strategy, and the monitoring results. The time-delay coupling dual-constraint BDQ network is then updated.

[0013] The beneficial effects of this invention are: This invention proposes a reinforcement learning-based environmental data monitoring method for edible fungi. This method collects environmental data, regulation data, and cultivation status data, and establishes an event response sequence around regulation events. It distinguishes between effective and low-response segments of environmental parameters, and performs hierarchical compression to retain key records in low-response segments. This reduces redundant monitoring data and computational load while maintaining key dynamic information such as response boundaries, direction changes, and spatial differences. Furthermore, based on actual feedback from regulation equipment and environmental parameter responses, it determines the response direction, response delay, and response operator between regulation events and each environmental parameter, constructing a time-delay coupling graph. This transforms the originally dispersed relationships between equipment actions, monitoring points, and environmental parameters into a graph structure with a clear propagation direction and time delay. The expected response is deduced through this graph structure, and counterfactual responses are obtained by replacing the response operators of historical events with the same initial state at the same stage. This allows for comparison of different response patterns under the same regulation input and connectivity, thereby improving the ability to distinguish response amplitude deviation, response direction reversal, response advance or delay, and response operator drift, and reducing misjudgments caused by relying solely on fixed thresholds or single residuals.

[0014] This invention also combines forward verification and reverse tracing to not only identify abnormal response nodes but also remove unilateral responses edge by edge along the time-delay coupled graph. Based on the degree of residual improvement, it determines the anomaly source edge and anomaly source event, enabling the monitoring results to extend from anomaly discovery to anomaly source localization. The time-delay coupled dual-constraint BDQ network couples graph relationships, anomaly encoding, and candidate monitoring operations, evaluating operations such as sampling period adjustment, synchronous monitoring of adjacent monitoring points, sensor cross-verification, and adjustment equipment feedback verification in parallel in a branching manner. It also utilizes path coverage and temporal consistency to correct the value of actions. Simultaneously, the constraint layer eliminates operations that do not meet sensor output capabilities, equipment status, and communication bandwidth requirements, and applies Lagrange penalties to resource consumption and operation switching. Therefore, it can generate monitoring strategies that balance anomaly coverage, execution continuity, and engineering resource constraints. After the monitoring strategy is implemented, the expected response baseline value is corrected based on the state drift and residual change, and the time delay coupling diagram and network parameters are updated. This enables the monitoring method to adapt to the continuous changes in the growth stage of edible fungi, the cultivation batch, and the response characteristics of the equipment, forming a closed-loop monitoring process from response modeling, anomaly tracing, strategy selection to feedback updates. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a reinforcement learning-based method for monitoring environmental data of edible fungi proposed in this invention. Figure 2This is a flowchart illustrating the time-delay coupled dual-constraint BDQ network used to generate the monitoring strategy for an edible fungus environmental data monitoring method based on reinforcement learning proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-2 A reinforcement learning-based method for monitoring environmental data of edible fungi includes: Collect and preprocess environmental data, conditioning data, and cultivation status data of edible fungi to generate monitoring data; The regulation events and event response sequences are determined from the monitoring data. The response boundaries are determined based on the event response sequences. The low-response segments between adjacent boundaries are compressed to generate a dynamic response range. Based on monitoring data and dynamic response intervals, determine the direction, time delay, and response operator of the regulation event and environmental parameters, and construct a time delay coupling diagram; Based on the time-delay coupling diagram and monitoring data, the expected response is deduced, and historical events with the same initial state at the same stage are matched. The response operators of the historical events are replaced with the response operators of the current adjustment event to deduce the counterfactual response. Compare monitoring data, expected response, and counterfactual response; perform forward verification and reverse tracing; and generate response residuals, anomaly codes, and candidate monitoring operations. The time-delay coupling graph, anomaly coding, and candidate monitoring operations are input into the coupling layer of the time-delay coupled double-constraint BDQ network to generate the coupling state. The coupling state is then input into the branch decision layer to generate the branch action values ​​for each candidate monitoring operation. Path coverage and timing consistency correction are performed. The corrected branch action values ​​are then input into the constraint layer to exclude unexecutable candidate monitoring operations. Lagrange penalties are applied to the branch action values ​​based on the constraint violations of the remaining candidate monitoring operations to generate the monitoring strategy. The monitoring strategy is executed and monitoring results are generated. Based on the monitoring results, the state drift and residual change are generated and the expected response baseline value is corrected. The time-delay coupling diagram and the time-delay coupling double-constraint BDQ network are updated accordingly.

[0018] In this embodiment, environmental data includes air temperature, air humidity, carbon dioxide concentration, light intensity, and culture medium temperature, moisture content, and pH. Adjustment data includes control commands, start and stop times, set values, and actual feedback values ​​of the adjustment equipment. Culture status data includes edible fungus growth stage, inoculation time, fruiting cycle, growth images, and culture medium configuration information. Preprocessing includes time alignment, outlier removal, missing value completion, and data scaling. The oyster mushroom cultivation area was selected as the monitoring object. The cultivation area was divided into an air intake zone, a middle zone and a return air zone along the air intake direction. An air monitoring point was set at the top, middle and bottom of the cultivation rack in each zone, forming a total of 9 air monitoring points. Two representative cultivation bags were selected in each zone to set up culture medium monitoring points, forming a total of 6 culture medium monitoring points. For environmental data, the sampling period was determined based on the response time of each sensor, the fastest rate of change of environmental parameters after the adjustment equipment was started and stopped, and the air exchange status of the culture area. The sampling period for air temperature, air humidity, carbon dioxide concentration, and culture medium temperature was set to 20 seconds, the sampling period for light intensity and culture medium moisture content was set to 60 seconds, and the sampling period for culture medium pH was set to 30 minutes. Using 20 seconds as a common time interval, air temperature, air humidity, carbon dioxide concentration, and culture medium temperature were directly mapped to the corresponding common time point. Environmental parameters with sampling periods longer than 20 seconds were associated with the corresponding common time point. The common time interval was determined based on the sampling period of carbon dioxide concentration, which has the fastest effective response speed and the longest sensor response time among the collected environmental parameters. At the same time, the 1st percentile and 99th percentile of the historical stable records of the same variety and the same growth stage were used as the scale boundaries of each environmental parameter. For adjustment data, the controller reads the control commands and actual feedback status of each adjustment device at a 1-second recording cycle; For the cultivation status data, the inoculation date, culture medium raw material ratio, growth stage and fruiting round are recorded according to the cultivation batch. A fixed camera device collects a growth image every 30 minutes with an image resolution of 1920×1080 pixels. The growth stages are divided into mycelial growth stage, primordium formation stage and fruiting body growth stage, and the time of stage change is written into the cultivation status record.

[0019] In this embodiment, the generation of the dynamic response range includes: The control status and feedback status of the regulating equipment are read from the monitoring data. The start and end of the regulating event are determined based on the corresponding changes of the two. The equipment operations with the same area of ​​action, the same control direction and the same time are merged into the same regulating event. Read the control commands, actual feedback values, and equipment operating status of each regulating device. When the control command starts or stops, switches gears, or changes the set value, and the actual feedback value changes accordingly within two common time intervals, the moment when the control command changes is determined as the start time of the regulating event. When the control command stops changing, and the change in the actual feedback value within three consecutive common time intervals is within the stable fluctuation range of the equipment, the third common time point is determined as the end time of the regulating event. The stable fluctuation range of the equipment is determined based on the normal operation records of the corresponding equipment. When two adjacent equipment operations act on the same culture area, have the same control direction, and the time interval between the two operations does not exceed two common time intervals, the two equipment operations are merged into the same regulating event. When the control directions are opposite, the areas of action are different, or the equipment fault status changes, regulating events are established separately. Based on the start time of the adjustment event, environmental parameter records before, during, and after the event are extracted, and the stable state before the event is used as the starting point for the response to generate an event response sequence. The system reads the control commands and actual feedback status of the regulating equipment from the monitoring data. The moment when the regulating equipment actually starts to execute is determined as the start time of the regulating event, and the moment when the regulating equipment stops executing and the actual feedback status remains stable is determined as the end time of the regulating event. When no regulating event occurs, the system collects various environmental parameters according to the normal sampling cycle and retains the monitoring records before the regulating event occurs in a cyclic buffering manner. When the regulating equipment starts to execute, the event sampling mode is triggered. The original sampling cycle is shortened according to the rate of change of environmental parameters and the sensor response capability, and the event sampling cycle is reset. After the regulating event ends, the event sampling mode is maintained until the corresponding environmental parameters return to stability or the maximum observation time is reached. The end time of the regulating event is determined by the cultivation process records corresponding to the current edible fungus variety and growth stage, as well as the operating mode of each regulating equipment. Based on the start time of the adjustment event, the system reads monitoring records from the circular buffer where no other adjustment events have occurred and all environmental parameters have remained within their historical stable fluctuation ranges. This generates pre-event data, and the stable state of each environmental parameter in the pre-event data is used as the response starting point. From the start time to the end time of the adjustment event, the system reads records of each environmental parameter, control commands, actual feedback values, and operating status of the adjustment equipment according to the event sampling period, generating event execution data. Starting from the end time of the adjustment event, the system continues to read records of each environmental parameter. When an environmental parameter re-enters the pre-event stable fluctuation range, and its response rate remains stable for three consecutive valid sampling times, the first valid sampling time that meets the condition is determined as the recovery time for that environmental parameter. The data between the end time of the adjustment event and the recovery time is then processed. The monitoring records are identified as post-event data. When a new adjustment event occurs during the post-event data reading period, an event overlay marker is set at the start time of the new adjustment event. The independent response records before the event overlay and the common response records after the event overlay are retained respectively. Taking the start time of the adjustment event as the zero time, the time interval between the effective sampling time of each environmental parameter and the start time of the adjustment event is recorded to establish an event relative time axis. Different environmental parameters are written into the event relative time axis according to the original effective sampling time. It is not required that each parameter has a sample value at every relative time point. The pre-event data, event execution data and post-event data corresponding to the same adjustment event are connected sequentially according to the event relative time axis. The response channels are divided according to the environmental parameter identifier and the monitoring point identifier to generate the event response sequence corresponding to the adjustment event. The event response sequence includes the event identifier, event type, control equipment data, event start time, event end time, environmental parameter identifier, monitoring point identifier, environmental parameter value, original sampling time, event relative time, response start point, recovery time, sampling mode, valid sampling identifier, event overlay marker, and unrecovered identifier. Among them, the event relative time refers to the time difference between the sampling time of a certain environmental parameter and the start time of the event. The response rate of each environmental parameter in the event response sequence is calculated according to the effective sampling time. Based on the change of the response rate relative to the historical stable fluctuation range, the effective response segment and the low response segment are divided. For each event response sequence, response channels are divided according to environmental parameter identifiers and monitoring point identifiers. Valid sampling records corresponding to the same environmental parameter and the same monitoring point are arranged according to the original sampling time. For two adjacent valid sampling records, the environmental parameter value of the latter valid sampling record is subtracted from the environmental parameter value of the former valid sampling record, and then divided by the actual time interval between the two original sampling times to generate the response rate. The response rate is calculated only based on the actual valid sampling records. When the time interval between adjacent valid sampling records exceeds twice the event sampling period, an interruption identifier is added to the corresponding response rate. The segment transition position is not determined separately using this response rate. The historical stable fluctuation range is determined based on the historical stable records of the same variety, the same growth stage, the same environmental parameters, and the same monitoring point. Valid records from the past seven days in which no adjustment events, equipment failures, sensor anomalies, or human interventions have occurred are read. The historical response rate is calculated according to the actual sampling interval, and the upper and lower boundaries of the stable fluctuation range are determined by the 95th percentile of the positive and negative response rates, respectively. When the response rate of a certain environmental parameter exceeds the historical stable fluctuation range for three consecutive valid sampling times, the first valid sampling time that exceeds the historical stable fluctuation range is determined as the starting position of the effective response segment, which is the response start boundary. When the response rate re-enters the historical stable fluctuation range for three consecutive valid sampling times, the first valid sampling time that re-enters the historical stable fluctuation range is determined as the ending position of the effective response segment, which is the response end boundary. The subsequent continuous stable records are determined as the low response segment. When the response rate in the effective response segment changes from a positive value to a negative value or from a negative value to a positive value, and the response direction after the change is maintained for at least two consecutive valid sampling times, the first valid sampling time in which the positive and negative states change is determined as the additional response boundary. The first and last records, direction change records, and spatial difference records of the low response segment are retained. The remaining records are subjected to hierarchical sampling and compression, and a correspondence between the retained records and the original sampling time is established. The first and last records of the low response segment are retained, as well as the positions where the response rate changes from positive to negative or from negative to positive and the records before and after them. For multiple monitoring points corresponding to the same environmental parameter, the difference between the maximum and minimum environmental parameter values ​​at each monitoring point at the same sampling time is calculated. When the difference exceeds the spatial difference boundary determined by the historical stable records, or when the monitoring point corresponding to the maximum or minimum environmental parameter value changes continuously, the records of each monitoring point at the corresponding sampling time are determined as spatial difference records and retained. The first-level candidate compression interval is determined by twice the event sampling period. The remaining records are sampled and retained according to the candidate compression interval. The environmental parameter values ​​of the corresponding times of the compressed records are estimated based on the adjacent retained records and their original sampling times. The difference between the estimated environmental parameter values ​​and the actual environmental parameter values ​​of the compressed records is determined as the compression reconstruction deviation. The allowable compression deviation is determined based on the sensor calibration error and the reconstruction deviation of historical stable records of the same variety and growth stage. The reconstruction deviation of historical stable records is obtained by estimating the environmental parameter values ​​of intermediate historical records from adjacent historical records and comparing them with the corresponding actual values. When the compression reconstruction deviation of each compressed record does not exceed the allowable compression deviation, the candidate compression interval is expanded step by step. When any compression reconstruction deviation exceeds the allowable compression deviation, the expansion is stopped, and the previous-level candidate compression interval is determined as the final compression interval, generating the compressed low-response segment. Record the original sampling time and retention type for each retained record. The retention types include first and last records, response rate direction change records, spatial difference records, and sampling retention records. For compressed records, record their original sampling time and the previous and next retained records that are adjacent in time, thereby establishing the correspondence between the retained records and the original sampling time. The effective response segments and the compressed low response segments are connected according to the original time sequence, and the dynamic response interval is generated by dividing the response boundary. For the same regulation event, the same environmental parameter, and the same monitoring point, the effective response segment and the compressed low response segment are connected sequentially according to the original sampling time. Multiple dynamic response intervals are generated by using each response boundary as the interval endpoint. Dynamic response intervals are generated for each environmental parameter response channel in an event response sequence and arranged according to the regulation event identifier, environmental parameter identifier, monitoring point identifier, and original time order to form a set of dynamic response intervals corresponding to the regulation event. Each dynamic response interval includes an adjustment event identifier, an environmental parameter identifier, a monitoring point identifier, an interval start time, an interval end time, a response direction, an interval type, a response boundary type, the number of original records, the number of compressed records, the original sampling time of the retained records, the time range of the compressed records, an event overlay marker, and a data discontinuity identifier.

[0020] In this embodiment, the generation of the time-delay coupling diagram includes: Generate the actual feedback sequence of the control equipment and the response sequence of environmental parameters based on monitoring data and dynamic response range; For each regulation event, the corresponding dynamic response range is read according to the regulation event identifier, and the response channel is divided according to the environmental parameter identifier and the monitoring point identifier. The actual feedback value, operating status and original sampling time of the regulation equipment are read to form the actual feedback sequence of the regulation equipment. The environmental parameter value, response rate and original sampling time in the same response channel are read to form the environmental parameter response sequence. The actual feedback sequence of the regulating equipment is normalized to a benchmark. For switching equipment, the stop state is recorded as 0 and the running state as 1. For gear-type equipment, the ratio of the actual gear to the maximum gear is used as the feedback value. For continuous regulating equipment, the actual feedback value is subtracted from the stable feedback value before the event starts and divided by the effective regulating range of the equipment to generate the normalized regulating input. The direction of response between the regulation event and the environmental parameter is determined based on the cumulative change of the environmental parameter response sequence relative to the response starting point and the sign consistency of the response rate. Using the response start point in the environmental parameter response sequence as a benchmark, the environmental parameter value at the end of the dynamic response interval is subtracted from the environmental parameter value corresponding to the response start point to generate a cumulative change. The direction determination band is determined based on the sensor calibration error and the parameter fluctuation amplitude recorded before the event. When the cumulative change is higher than the upper boundary of the direction determination band, the response direction is determined as positive. When the cumulative change is lower than the lower boundary of the direction determination band, the response direction is determined as negative. When the cumulative change is within the direction determination band, the corresponding response channel is marked as not having formed a stable response direction. The duration for which the sign of the response rate is consistent with the direction of the cumulative change within the effective response segment is counted, and the proportion of this duration to the total duration of the effective response segment is calculated to generate a direction consistency rate. When the direction consistency rate is greater than or equal to the direction consistency rate threshold, the determined response direction is retained. When the direction consistency rate is less than the direction consistency rate threshold, the response channel is marked as having an unstable direction. The determination of the direction determination zone specifically includes: reading the maximum absolute measurement error in the most recent calibration record of the corresponding environmental parameter sensor; since the cumulative change is obtained by subtracting the response start value from the end value of the dynamic response interval, both environmental parameter values ​​may contain measurement errors; therefore, twice the maximum absolute measurement error is determined as the measurement error limit; from the historical stable records of the same variety, the same growth stage, the same environmental parameter, and the same monitoring point, a stable time window with the same duration as the current dynamic response interval is extracted; the historical stable change obtained by subtracting the initial environmental parameter value from the end value of each stable time window is calculated; and the values ​​are arranged in ascending order, with the 95th percentile taken as the historical positive fluctuation limit; the absolute value of the negative values ​​in the historical stable change is taken and arranged in ascending order, with the 95th percentile taken as the historical negative fluctuation limit; the larger value between the measurement error limit and the historical positive fluctuation limit is determined as the upper boundary of the direction determination zone; and the negative value of the larger value between the measurement error limit and the historical negative fluctuation limit is determined as the lower boundary of the direction determination zone. The determination of the directional consistency rate threshold specifically includes: reading historical valid events that did not involve event superposition, equipment failure, or sampling interruption in the same variety, same growth stage, and same type of regulation events; calculating the directional consistency rate of each historical valid event; and determining the 10th percentile of the directional consistency rate as the directional consistency rate threshold, with the initial engineering value set at 0.7. Based on the actual action time of the regulating equipment, the actual feedback sequence of the regulating equipment is shifted according to the event sampling period, and the time shift correlation is calculated in combination with the response sequence of environmental parameters. The response time delay is generated according to the time offset corresponding to the peak correlation. The actual action time of the regulating equipment is determined according to the equipment type. For switch-type or gear-type regulating equipment, when the equipment's operating state changes from a stopped state to an operating state or the actual feedback gear changes, and the actual feedback state remains unchanged for two consecutive feedback cycles, the first feedback time after the state change is determined as the actual action time. For continuous regulating equipment, five consecutive normal feedback values ​​are read before the event starts, the difference between the maximum and minimum feedback values ​​is calculated, the feedback fluctuation amplitude is generated, the equipment's feedback resolution is read, and the larger value between the feedback resolution and the feedback fluctuation amplitude is determined as the action judgment limit. When the change in the actual feedback value relative to the average feedback value before the event starts exceeds the action judgment limit, and the change direction remains the same for two consecutive feedback cycles, the first feedback time that meets the condition is determined as the actual action time. According to the event sampling period corresponding to the environmental parameters, the actual feedback sequence of the regulating equipment is time-aligned. When the equipment feedback period is less than the event sampling period, the most recent actual feedback value before each event sampling moment is retained. When the equipment feedback period is greater than the event sampling period, only the sampling moment with a real feedback value is used, and missing feedback values ​​are not repeatedly filled. The actual feedback value of the next actual feedback moment is subtracted from the previous actual feedback value, and divided by the actual time interval between the two to generate the equipment feedback change rate sequence. With the actual action moment of the regulating equipment as the zero time shift position, the equipment feedback change rate sequence is successively shifted backward according to the event sampling period corresponding to the environmental parameters, and matched according to the original sampling moment. Records in which both sequences have valid values ​​at the same sampling moment are retained. The normalized cross-correlation method is used to calculate the equipment feedback change rate sequence and the environmental parameter response rate under the corresponding time offset. The normalized time-shift correlation between sequences is calculated as follows: when the determined response direction is positive, the time offset with the largest correlation is selected from the time offsets with positive correlation; when the determined response direction is negative, the time offset with the largest absolute value of correlation is selected from the time offsets with negative correlation. The selected time offset is then determined as the candidate response delay. When the positive or negative state of the correlation is inconsistent with the determined response direction, the corresponding response channel is marked as having an inconsistent delay direction, and no determined response delay is output. The time difference between the response start boundary of the dynamic response interval and the actual action time of the regulating device is calculated to generate the boundary response delay. When the difference between the candidate response delay and the boundary response delay does not exceed two event sampling periods, the candidate response delay is determined as the response delay; otherwise, the boundary response delay is determined as the response delay, and a low-confidence delay flag is added to the corresponding response relationship. Align the actual feedback sequence of the adjustment equipment and the response sequence of environmental parameters according to the response time delay, fit the discrete response kernel according to the change in adjustment input and the offset of environmental parameters, and generate the response operator. Add the corresponding response time delay to the original sampling time of each feedback record in the actual feedback sequence of the regulating device to generate the feedback alignment time. Match the feedback alignment time with the environmental parameter response sequence, retain data with a time difference of no more than half an event sampling period and all of which are valid records, and generate time delay alignment records. The actual feedback values ​​of the devices in the time-delay alignment record are normalized. The current feedback value is subtracted from the previous feedback value to generate the adjustment input change. The current environmental parameter value is subtracted from the environmental parameter value corresponding to the response start point to generate the environmental parameter offset. The duration between the response start boundary and the response end boundary is divided by the event sampling period and rounded up to generate the response kernel length. According to the original time order, the adjustment input changes within the current moment and the aforementioned response kernel length are arranged sequentially to form the adjustment input vector. The actual environmental parameter offset at the current moment is used as the corresponding output. Only when each adjustment input change and the corresponding actual environmental parameter offset in the adjustment input vector have valid records, a set of response kernel valid fitting samples is generated. The number of all valid fitting samples is counted to generate the number of valid fitting samples. Least squares with regularization constraints is used. The fitting method involves determining the response coefficients corresponding to each time position in the adjustment input vector, arranging the response coefficients in chronological order to generate a discrete response kernel, weighting and accumulating the adjustment input change with the discrete response kernel to generate the fitting environment parameter offset, subtracting the fitting environment parameter offset from the actual environment parameter offset to generate the fitting residual, and generating a normalized fitting error based on the ratio of the root mean square value of each fitting residual to the maximum offset of the actual environment parameter. The response coefficients in the discrete response kernel are accumulated to generate the cumulative response gain corresponding to a unit adjustment input. The response direction, response time delay, discrete response kernel, cumulative response gain, normalized fitting error, and number of effective fitted samples are combined to generate a response operator. The actual environment parameter offset is obtained by subtracting the actual environment parameter value corresponding to the response start point from the actual environment parameter value in the corresponding time delay alignment record. Using the regulation event as the event node and the environmental parameters and corresponding monitoring points as the response nodes, a directed edge is established when the response direction, response delay, and response operator meet the validity requirements. The response direction, response delay, response operator, and response confidence are written into the directed edge to construct a time-delay coupling graph. When the response direction, response delay, and response operator are all valid, a directed edge is established from the event node to the response node. Specifically, when the corresponding response channel is marked as not having formed a stable response direction or with an unstable direction, a directed edge is not established. Otherwise, the determined positive or negative response direction is used as the direction attribute. When the response delay is determined, the positive or negative state of the time shift correlation is consistent with the response direction, and the absolute value of the time shift correlation is not lower than the 10th percentile of the historical effective responses of the same variety, the same growth stage, and the same type of regulatory event, the response delay is deemed valid, and the response delay and time shift correlation are used as the delay attribute. When the actual maximum offset is not 0, the number of effective fitted samples is greater than the number of response coefficients in the discrete response kernel, and the normalized fitting error is not higher than the 90th percentile of the normalized fitting error of the same type of historical effective response, the response operator is deemed valid. When there are fewer than 20 historical effective responses, the lower and upper limits are determined by supplementing the same type of response records formed during the equipment debugging phase. For response confidence, the number of effective fitted samples is divided by the theoretical number of samples within the dynamic response interval to generate the sample coverage rate. The result obtained by subtracting the normalized fitting error from 1 is limited to between 0 and 1 to generate the operator fitting confidence. The arithmetic mean of the direction consistency rate, the absolute value of the time shift correlation, the operator fitting confidence, and the sample coverage rate is calculated to generate the response confidence with a value between 0 and 1.

[0021] In this embodiment, the deduction of the counterfactual response includes: The current adjustment input change is shifted according to the response time delay of each directed edge in the time-delay coupling diagram, and the one-sided response of each response node is calculated using the corresponding response operator. The expected response is generated based on the one-sided response. Arrange the actual feedback values ​​of the current regulation event according to the sampling time, subtract the previous actual feedback value from the next actual feedback value to generate the current regulation input change sequence. For any directed edge, divide the response time delay of the directed edge by the event sampling period of the corresponding environmental parameter, and take the nearest integer of the result to generate the time delay step number. Move the current regulation input change sequence backward according to the time delay step number, and multiply each of the moved regulation input changes by the response coefficient at the corresponding time position in the discrete response kernel and accumulate them to generate the single-sided response corresponding to the directed edge. When multiple directed edges point to the same response node, the one-sided responses generated by each directed edge at the same sampling time are added together to generate the total response offset of the response node. The total response offset is added to the response starting point value of the response node to generate the expected environmental parameter value at the sampling time. The expected environmental parameter values ​​are arranged in chronological order to generate the expected response of the corresponding response node. Historical events are selected from those with consistent growth stages, adjusted event types, and response nodes, and whose initial environmental parameter differences do not exceed the corresponding matching limits. Historical events with the same initial state at the same stage are also identified. The average values ​​of environmental parameters within the stable segment before the actual action of the current regulation event and each historical event are read respectively to generate the current initial environmental state and the historical initial environmental state. For each environmental parameter, the larger value between twice the maximum calibration error of the sensor and the 95th percentile of the historical stable fluctuation of the environmental parameter under the same variety and growth stage is determined as the initial state matching limit. The difference between the same environmental parameter in the current initial environmental state and the historical initial environmental state is calculated respectively. When the absolute value of the difference of any environmental parameter is greater than the corresponding initial state matching limit, the corresponding historical event is excluded. Otherwise, the corresponding historical event is retained. The difference of each environmental parameter is divided by the corresponding initial state matching limit to generate standardized parameter difference. The sum of the squares of all standardized parameter differences is divided by the number of environmental parameters involved in the matching and the square root is taken to generate the standardized initial state distance. The historical event with the smallest value is determined as the historical event with the same initial state in the same stage. Replace the response operators of each directed edge of the current adjustment event with the response operators of the corresponding directed edges in the historical events of the same stage and initial state, keep the current adjustment input change, response nodes and graph connection unchanged, recalculate the one-sided response of each response node, and generate counterfactual response based on the recalculated one-sided response. The historical time delay steps are generated based on the historical response time delays. The one-sided response is recalculated using the historical discrete response kernel and the current adjustment input change sequence. The recalculated one-sided responses are summarized according to the response node and sampling time, and added to the response starting point value of the current response node to generate counterfactual environmental parameter values. The counterfactual environmental parameter values ​​are arranged in chronological order to generate the counterfactual response.

[0022] In this embodiment, the generation of response residuals, anomaly codes, and candidate monitoring operations includes: Based on monitoring data, expected response, and counterfactual response, the differences between actual environmental parameter values ​​and expected environmental parameter values ​​and counterfactual environmental parameter values ​​are calculated respectively, generating expected residual sequences and counterfactual residual sequences; According to the response node identifier and the original sampling time, the actual environmental parameter values ​​in the monitoring data, the expected environmental parameter values ​​in the expected response, and the counterfactual environmental parameter values ​​in the counterfactual response are matched. Only records in which the actual environmental parameter values, expected environmental parameter values, and counterfactual environmental parameter values ​​are all valid at the same sampling time are retained. No missing values ​​are supplemented by interpolation. When there is a low response segment compression in the dynamic response interval, each retained record is mapped to the original sampling time according to the original time mapping relationship saved during the compression process. For each valid matching record, the expected environmental parameter value at the same sampling time is subtracted from the actual environmental parameter value to generate the expected residual. The counterfactual environmental parameter value at the same sampling time is subtracted from the actual environmental parameter value to generate the counterfactual residual. The expected residuals are arranged in order of sampling time to generate the expected residual sequence. The counterfactual residuals are arranged in order of sampling time to generate the counterfactual residual sequence. Perform amplitude verification, direction verification, and response time verification on the expected residual sequence to generate a positive verification result; Read historical response records for the same variety, growth stage, type of regulation event, environmental parameter, and monitoring point, and arrange the expected residuals in ascending order. Determine the 5th percentile of the historical expected residuals as the lower boundary of the initial residuals, and the 95th percentile of the historical expected residuals as the upper boundary of the initial residuals. When the maximum calibration error is greater than the upper boundary of the initial residuals, the maximum calibration error is determined as the upper boundary of the normal residuals. When the maximum calibration error is greater than the absolute value of the lower boundary of the initial residuals, the maximum calibration error is negative and determined as the lower boundary of the normal residuals. For amplitude verification, when the expected residual is higher than the upper boundary of the normal residual for three consecutive valid sampling times, a positive amplitude abnormality segment is determined; when the expected residual is lower than the lower boundary of the normal residual for three consecutive valid sampling times, a negative amplitude abnormality segment is determined. For each amplitude abnormality segment, the maximum value among the absolute values ​​of the expected residual is selected as the maximum residual. The absolute values ​​of the expected residuals in the abnormal segment are added together and divided by the number of valid sampling records to generate the average absolute residual. The expected residuals in the abnormal segment are squared respectively. The squared results are added together and divided by the number of valid sampling records. The square root of the result is then taken to generate the root mean square value of the expected residual. For direction verification, the actual environmental parameter values ​​within the actual response segment are subtracted from the response start value to generate the actual environmental parameter offset. When the response direction of the corresponding directed edge in the time-delay coupling graph is positive, and the actual environmental parameter offset is lower than the lower boundary of the direction decision band, and this state lasts for two consecutive event sampling periods, a negative reversal anomaly is generated. When the response direction of the directed edge is negative, and the actual environmental parameter offset is higher than the upper boundary of the direction decision band, and this state lasts for two consecutive event sampling periods, a positive reversal anomaly is generated. For response time verification, the original sampling time corresponding to the actual response start boundary in the dynamic response interval is determined as the actual response start time. The expected environmental parameter values ​​are read sequentially from the expected response, and their expected offset relative to the response start value is calculated. When the expected offset exceeds the boundary of the direction determination zone for two consecutive sampling times, the first sampling time that exceeds the boundary is determined as the expected response start time. The actual response start time is subtracted from the expected response start time to generate the response time deviation. When the absolute value of the current response time deviation is greater than the allowable time deviation, a response time anomaly is generated. The allowable time deviation is determined by the larger value of the 90th percentile of the absolute values ​​of the historical effective response time deviations of the same variety, the same growth stage, the same type of regulation event, and the same response node arranged from smallest to largest, and the larger value in an event sampling period. The positive verification result is obtained by combining the amplitude anomaly section, maximum residual, mean absolute residual, expected root mean square value of residual, direction anomaly indicator and response time deviation according to the response node. For a response node that is a positive verification anomaly, remove the corresponding single-edge response one by one along the directed edge pointing to the response node in the time-delay coupling graph, compare the changes in the residual before and after removal, determine the source edge and source event of the anomaly, and generate the reverse tracing result. When a response node has at least one of the following: amplitude anomaly segment, direction anomaly identifier, or response time anomaly identifier, the response node is identified as an abnormal response node. The verification anomaly interval corresponding to amplitude anomaly is the time interval during which the expected residual continuously exceeds the normal residual boundary. The verification anomaly interval corresponding to direction anomaly starts from the sampling time when the actual environmental parameter offset deviates from the expected response direction for the first two consecutive event sampling periods, and ends when the actual environmental parameter offset recovers to the expected response direction for two consecutive event sampling periods, or reaches the response end boundary of the dynamic response interval. The verification anomaly interval corresponding to response time anomaly starts from the earlier of the actual response start time and the expected response start time, and continues for one response kernel duration. When the end time after the extension exceeds the response end boundary of the dynamic response interval, the response end boundary is taken as the end time of the verification anomaly interval. The response kernel duration is the product of the number of response coefficients in the discrete response kernel and the event sampling period. When multiple positive verification anomalies exist simultaneously in the same response node, the verification anomaly intervals are merged to form the reverse tracing interval of the abnormal response node. Read all directed edges pointing to anomalous response nodes in the time-delay coupling graph. Identify directed edges with at least one non-zero one-sided response value within the reverse tracing interval as candidate source edges. Subtract the original expected environmental parameter value at the same sampling time from the actual environmental parameter value of the anomalous response node within the reverse tracing interval to generate the original residual. Square each original residual, sum the squared results, divide by the number of valid records, and then take the square root of the result to generate the root mean square value of the original residual. For any candidate source edge, subtract the corresponding one-sided response value from the original expected environmental parameter value according to the original sampling time to generate the expected environmental parameter value after removal. Subtract the removal value at the same sampling time from the actual environmental parameter value. After the expected environmental parameter values ​​are obtained, the residuals after removal are generated. Each residual after removal is squared, the squared results are added together and divided by the number of valid records, and the square root of the result is taken to generate the root mean square value of the residuals after removal. The root mean square value of the original residuals is subtracted from the root mean square value of the residuals after removal, and the difference is divided by the root mean square value of the original residuals to generate the source improvement rate. When the source improvement rate of a candidate source edge is higher than the source improvement rate threshold, the candidate source edge is identified as an abnormal source edge, and the adjustment event corresponding to the starting point of the abnormal source edge is identified as an abnormal source event. The source improvement rate threshold is determined by performing the same edge-by-edge removal process on the historical normal response records. The initial value of the project is set to the 95th percentile of the historical source improvement rate. The reverse tracing result is generated by combining the abnormal response node, the reverse tracing interval, the candidate source edge, the abnormal source edge, the abnormal source event, the original residual root mean square value, the residual root mean square value after removal, and the source improvement rate. Based on the expected residual sequence, the counterfactual residual sequence, the forward verification result, and the reverse tracing result, generate the response residual, anomaly code, and candidate monitoring operations; Calculate the root mean square (RMS) values ​​of the expected residuals and counterfactual residuals within the reverse source tracing interval. Select the expected residual with the largest absolute value from the expected residual sequence to generate the maximum expected residual. Sum the absolute values ​​of all expected residuals and divide by the number of valid sampled records to generate the average absolute expected residual. Select the counterfactual residual with the largest absolute value from the counterfactual residual sequence to generate the maximum counterfactual residual. Sum the absolute values ​​of all counterfactual residuals and divide by the number of valid sampled records to generate the average absolute counterfactual residual. Subtract the RMS value of the counterfactual residual from the RMS value of the expected residuals and divide the difference by the RMS value of the expected residuals to generate the counterfactual improvement rate. Combine the expected residual sequence, counterfactual residual sequence, maximum expected residual, maximum counterfactual residual, average absolute expected residual, average absolute counterfactual residual, RMS value of expected residuals, RMS value of counterfactual residuals, counterfactual improvement rate, response time deviation, anomaly source edge and corresponding source improvement rate to generate the response residual. When a directional anomaly is detected in the forward verification result, the main anomaly type is identified as response direction anomaly, coded as A3. When there is no directional anomaly but a response time anomaly is detected, the main anomaly type is identified as response time anomaly, coded as A2. When there are no response direction anomalies or response time anomalies, the counterfactual improvement rate is higher than the counterfactual improvement rate threshold, and there is an anomaly source edge in the reverse tracing result, the main anomaly type is identified as response operator drift, coded as A4. When the expected residual continuously exceeds the normal residual boundary, but the counterfactual improvement rate is not higher than the counterfactual improvement rate threshold, and there is an anomaly source edge in the reverse tracing result, the main anomaly type is identified as response amplitude anomaly, coded as A1. When both the expected residual and the counterfactual residual continuously exceed the normal residual boundary, and no anomaly source edge is identified in the reverse tracing result, the main anomaly type is identified as monitoring anomaly pending verification, coded as A5. The counterfactual improvement rate threshold is determined by the 95th percentile of the historical counterfactual improvement rate derived from historical normal response records. Candidate monitoring operations include at least one of sampling period adjustment, synchronous monitoring of adjacent monitoring points, sensor cross-verification, and adjustment equipment feedback verification. The target object, monitoring period, and duration are configured for the candidate monitoring operation. The duration is determined based on the response time delay and the response kernel duration. The response time delay is multiplied by 3 to generate the first duration. The number of response coefficients in the discrete response kernel is multiplied by the event sampling period to generate the second duration. The maximum value among the first duration, the second duration, and ten event sampling periods is determined as the duration of the candidate monitoring operation.

[0023] In this embodiment, the generation of the monitoring strategy includes: The time-delay coupling graph, anomaly coding, and candidate monitoring operations are input into the coupling layer of the time-delay coupled dual-constraint BDQ network; The time-delay coupled dual-constraint BDQ network includes a coupling layer, a branch decision layer, and a constraint layer. The model training process specifically includes: acquiring historical monitoring records of completed monitoring strategy execution; dividing the valid records into training sample sets, validation sample sets, and test sample sets according to training batches, with proportions of 70%, 15%, and 15%, respectively; constructing a set of training samples based on the generation and execution process of a single monitoring strategy; using the coupling state at the time of monitoring strategy generation as the pre-execution coupling state; using the actual executed candidate monitoring operations and their action levels as action labels; using the coupling state regenerated based on the monitoring results and the updated time-delay coupling diagram as the post-execution coupling state; generating a task completion label based on whether the monitoring task has ended; limiting the residual change to between -1 and 1 to generate a monitoring effect label; limiting the monitoring completion rate to between 0 and 1 to generate an execution completion label; calculating the average values ​​of transmission bandwidth occupancy, monitoring channel occupancy, and verification equipment occupancy to generate a resource occupancy label; and assigning the monitoring effect label to the target value. The result label is added to the execution completion label, and then the resource consumption label and the Lagrange constraint penalty are subtracted to generate the feedback value label. The squared difference between the predicted branch action value and the target branch action value is calculated and averaged over all samples in the training batch to generate the temporal difference loss. The constraint loss is generated by multiplying each soft constraint violation by the corresponding Lagrange multiplier and summing the results. The temporal difference loss, constraint loss, and network weight regularization loss are added to generate the total model loss. The Adam optimizer is used to update the coupling layer, shared value path, dominant path of each action branch, path coverage correction coefficient, and temporal consistency correction coefficient based on the total model loss. The initial learning rate is set to 0.001 and the training batch size is set to 64. When the total model loss of the validation sample set changes by no more than 1% for five consecutive times, the average feedback value changes by no more than 2%, and the hard constraint violation rate is 0, the time-delay coupled double-constraint BDQ network is considered to have converged. The improvements to the model specifically include: 1. This invention adds a coupling layer before the original BDQ network to aggregate and fuse the relationships between the time-delay coupling graph, anomaly coding, and candidate monitoring operations, generating a coupling state that enables action decisions to distinguish between abnormal response nodes, anomaly source edges, and different response time delays; 2. This invention adds path coverage correction and timing consistency correction, adjusting branch action values ​​based on the degree of abnormal path coverage and the continuity of the target object, operation type, and monitoring cycle; 3. This invention adds a constraint layer to directly exclude actions that violate sensor output cycle, device online status, and bandwidth limits, and uses Lagrange penalties to constrain resource usage and operation switching, ensuring that the generated monitoring strategy has both monitoring effectiveness and engineering feasibility. In the coupling layer, graph relationship aggregation and exception operation fusion are performed to generate a coupled state; The adjustment event type, adjustment input change, environmental parameter value, response delay, cumulative response gain, response confidence, and source improvement rate in the time-delay coupling graph are combined into graph structure features. The anomaly type, anomaly response node, expected residual root mean square value, counterfactual residual root mean square value, and response time deviation are combined into anomaly features. The target object, operation type, monitoring cycle, duration, and required number of monitoring channels for each candidate monitoring operation are combined into candidate operation features. The event node features, response node features, and directed edge features are aggregated to generate a graph relationship representation. The graph relationship representation is fused with the anomaly features to generate an anomaly coupling representation. The anomaly coupling representation is fused with the candidate operation features to generate a candidate operation association representation. The graph relationship representation, anomaly coupling representation, and candidate operation association representation are concatenated to generate a joint feature vector. The joint feature vector is input into a fully connected layer and linearly mapped. ReLU activation processing is performed on each element in the mapped feature vector to generate a coupling state. The coupling state is input into the branch decision layer, and action branches are established for each candidate monitoring operation. Branch action values ​​are generated for each action branch, and path coverage and timing consistency correction are performed on the branch action values. Action branches are established according to the operation identifiers of candidate monitoring operations. Each action branch has two action levels: "not selected" and "selected". When a candidate monitoring operation has multiple available parameter levels, corresponding action levels are set respectively. The action levels for the sampling period adjustment operation include maintaining the original sampling period, shortening it to half of the original sampling period, shortening it to one-third of the original sampling period, and using the minimum output period of the sensor. The action levels for the synchronous monitoring operation of adjacent monitoring points include selecting the nearest monitoring point, selecting the nearest two monitoring points, and selecting the nearest three monitoring points. The action levels for the feedback verification operation of the adjustment equipment include verifying the operating status, verifying the operating status and actual feedback value, and synchronously verifying the operating status, actual feedback value, and feedback timestamp. The coupling state is input into the dominant path corresponding to each action branch. Each dominant path includes a first fully connected layer, a ReLU activation layer, and a second fully connected layer. The first fully connected layer multiplies the coupling state with the branch weight matrix and adds it to the branch bias vector to generate the branch intermediate vector. ReLU activation is then performed on the branch intermediate vector. The second fully connected layer maps the activated branch intermediate vector to the same number of floating-point numbers according to the number of action levels in the corresponding action branch, generating the dominant value of each action level in that action branch, and calculating the average of the dominant values ​​of all action levels in the same action branch. The mean is used to input the coupled state into the shared value path, which includes a first fully connected layer, a ReLU activation layer, and a second fully connected layer. The first fully connected layer multiplies the coupled state with the value weight matrix and adds it to the value bias vector to generate a value intermediate vector. The value intermediate vector is then subjected to ReLU activation processing, setting elements less than 0 to 0. The second fully connected layer maps the activated value intermediate vector to a floating-point number to generate a shared state value. For any action level, the shared state value is added to the advantage value of that action level, and then the mean is subtracted to generate the branch action value of that action level. For path coverage correction, a set of objects to be covered, including abnormal response nodes, is established. When there is an abnormal source edge in the reverse tracing result, the abnormal source edge and the corresponding abnormal source event are added to the set of objects to be covered. When the abnormal source event is also connected to other abnormal response nodes, the other abnormal response nodes are added to the set of objects to be covered. Each candidate monitoring operation and the set of objects to be covered are compared. The number of objects to be covered directly covered by the candidate monitoring operation is divided by the total number of objects to be covered to generate the direct coverage rate. When the target object of the candidate monitoring operation is connected to the object to be covered by a directed edge, but does not directly point to the object to be covered, the object to be covered is recorded as a half-covered object. The number of half-covered objects is multiplied by one-half and added to the number of directly covered objects, and then divided by the total number of objects to be covered to generate the path coverage rate. The average path coverage rate of all candidate monitoring operations is calculated. The path coverage rate of the current candidate monitoring operation is subtracted from the average path coverage rate to generate the relative path coverage. The relative path coverage is multiplied by the path coverage correction coefficients trained in the branch decision layer to generate the path coverage correction value. For time-series consistency correction, the monitoring strategies generated but not yet completed in the three consecutive decision cycles before the current decision cycle are read. The current candidate monitoring operation is compared with the operation in the historical monitoring strategy. When the target response node is the same, the target consistency flag is recorded as 1, otherwise it is recorded as 0. When the operation type is the same, the type consistency flag is recorded as 1, otherwise it is recorded as 0. When the difference between the current monitoring cycle and the historical monitoring cycle does not exceed 20% of the larger of the two, the cycle consistency flag is recorded as 1, otherwise it is recorded as 0. The target consistency flag, type consistency flag and cycle consistency flag are added together and divided by three to generate the time-series consistency rate. When there is no operation with the same target object in the historical monitoring strategy, the time-series consistency rate is recorded as half. When the historical operation has not reached half of its duration, and the current candidate monitoring operation changes its target object or operation type, the remaining duration of the historical operation is divided by the original duration to generate the operation switching amount. The time-series consistency rate is multiplied by the time-series consistency correction coefficient trained in the branch decision layer and the operation switching amount is subtracted to generate the time-series consistency correction value. For each candidate monitoring operation, its branch action value is added to the path coverage correction value and the timing consistency correction value to generate the corrected branch action value; The corrected branch action value is input into the constraint layer to exclude unexecutable candidate monitoring operations. The constraint violation of the remaining candidate monitoring operations is calculated. Lagrange multipliers are used to apply penalties to the corresponding branch action value. Candidate monitoring operations are selected according to the penalized branch action value to generate a monitoring strategy. When the monitoring cycle of a candidate monitoring operation is shorter than the minimum output cycle of the target sensor, or when a candidate monitoring operation requires a backup sensor or portable detector but the corresponding device is not online or available, the corresponding candidate monitoring operation is considered an unexecutable operation. When the feedback field required for the adjustment device feedback verification operation is not output by the adjustment device controller, the action level containing the missing feedback field is excluded, and the action level that can be actually output by the controller is retained. The single sampling data volume of the candidate monitoring operation is multiplied by the number of monitoring channels and then divided by the monitoring cycle to generate a new data transmission rate. The current gateway data transmission rate is added to the new data transmission rate. When the result is higher than 80% of the current available transmission bandwidth of the gateway, the corresponding candidate monitoring operation or corresponding parameter level is considered an unexecutable operation. When two candidate monitoring operations are configured with different sampling cycles for the same sensor, the shorter sampling cycle is retained, and the repeated operation corresponding to the longer sampling cycle is excluded. For candidate monitoring operations that have not been excluded, the soft constraint violation amount is calculated. The soft constraints include bandwidth usage, sensor workload, number of simultaneous monitoring channels, and operation switching frequency. Specifically, the 90th percentile of the resource usage during the historical normal monitoring period is read as the soft constraint boundary, and the rated upper limit of the device or the allowed upper limit of the gateway is read as the hard constraint boundary. When the resource usage after the candidate monitoring operation is executed is not higher than the soft constraint boundary, the corresponding constraint violation amount is recorded as 0. When the resource usage is higher than the soft constraint boundary but not higher than the hard constraint boundary, the resource usage is subtracted from the soft constraint boundary and then divided by the difference between the hard constraint boundary and the soft constraint boundary to generate the constraint violation amount. The bandwidth constraint violation amount, sensor load constraint violation amount, channel number constraint violation amount, and operation switching constraint violation amount are added together and divided by the number of constraints involved in the calculation to generate the comprehensive constraint violation amount of the candidate monitoring operation. Each constraint violation is multiplied by the corresponding Lagrange multiplier of each soft constraint, and the products are summed to generate a Lagrange penalty. The corrected branch action value is subtracted from the Lagrange penalty to generate a penalized branch action value. For each action branch, if the penalized branch action value of the "Select" action level is higher than that of "Do Not Select", the candidate monitoring operation corresponding to "Select" is retained; otherwise, the candidate monitoring operation is not selected. Candidate monitoring operations are added one by one in descending order of penalized branch action values. After each addition, the transmission bandwidth, monitoring channel, and equipment occupancy status are re-verified, and candidate monitoring operations that satisfy hard constraints are retained. The execution order is determined according to the penalized branch action values. The finally selected candidate monitoring operations, execution order, monitoring cycle, duration, target response node, target directed edge, and target adjustment device are combined to generate a monitoring strategy.

[0024] In this embodiment, the updating of the time-delay coupling graph and the time-delay coupling double-constraint BDQ network includes: Each monitoring operation is scheduled according to the execution order, monitoring cycle, and duration in the monitoring strategy, and monitoring results are generated. Monitoring tasks are issued to sensors, regulating equipment controllers, monitoring gateways, or calibration equipment in the execution order, and the task issuance time, equipment reception time, and actual start time are recorded. The duration is divided by the monitoring cycle and rounded up to generate the theoretical sampling quantity. The number of records with valid timestamps and valid measurements is counted, and the number of valid records is divided by the theoretical sampling quantity to generate the monitoring completion rate. The monitoring completion rate of similar historical normal monitoring tasks is read, and its 10th percentile is determined as the lower limit of the completion rate. When the current monitoring completion rate is not lower than the lower limit of the completion rate, the corresponding monitoring task is determined as a valid monitoring task. The operation identifier, anomaly code, target object, planned monitoring cycle, actual monitoring cycle, planned duration, actual duration, original sampling time, actual measurement value, equipment feedback value, calibration measurement value, number of valid records, and monitoring completion rate are combined to generate the monitoring result. The environmental parameter values, equipment feedback values, and calibration measurement values ​​before and after the monitoring operation are statistically analyzed, and the state drift is generated based on the change range of similar data. The average of the five most recent valid environmental parameter records before the actual start time of the monitoring operation is calculated to generate the pre-operation state value. The average of the five most recent valid environmental parameter records before the end of the monitoring operation is calculated to generate the post-operation state value. The post-operation state value is subtracted from the pre-operation state value to generate the state change value. The difference between the 95th percentile and the 5th percentile of the historical stable environmental parameter values ​​for the same variety, growth stage, and monitoring point is calculated to generate the historical stable fluctuation width. The larger of twice the maximum calibration error of the corresponding sensor and the historical stable fluctuation width is determined as the drift scale. The state change value is divided by the drift scale to generate the state drift component of the corresponding environmental parameter. The same method is used to calculate the corresponding state drift component for the feedback value of the adjustment equipment and the calibration measurement value. Arrange the state drift components corresponding to each environmental parameter, equipment feedback and verification measurement in order to generate a state drift vector, and determine the component with the largest absolute value as the state drift quantity. Based on the monitoring results, the residuals between the actual environmental parameter values ​​and the expected response are recalculated, and the residuals before and after the monitoring operation are compared to generate the residual change. Read the five most recent environmental parameter records that can be aligned with the expected response before the actual start of the monitoring operation. Subtract the expected environmental parameter values ​​at the same sampling time from the actual environmental parameter values ​​to generate the pre-operation residuals. Square each pre-operation residual separately. Add the squared results and divide by the number of effective pre-operation residuals. Take the square root of the result to generate the root mean square value of the pre-operation residuals. Read the five most recent environmental parameter records that can be aligned with the expected response before the end of the monitoring operation. Generate the post-operation residual root mean square value in the same way. Subtract the post-operation residual root mean square value from the pre-operation residual root value. Divide the difference by the pre-operation residual root value to generate the residual change. When the pre-operation residual root mean square value is 0, record the residual change as 0. When the state drift continues and the residuals do not converge, a baseline correction amount is generated based on the deviation between the monitoring results and the expected response, and the baseline correction amount is used to correct the baseline value of the expected response. When the monitoring task is effective, the state drift is continuous, and the residual change is less than or equal to 0, the expected response benchmark correction is initiated. The five most recent effective records before the end of monitoring are read, and the expected environmental parameter value at the same sampling time is subtracted from each actual environmental parameter value to generate a benchmark deviation sample. The median of the five benchmark deviation samples is selected as the candidate benchmark correction amount. When the absolute value of the candidate benchmark correction amount does not exceed the historical stable fluctuation width of the corresponding environmental parameter, the candidate benchmark correction amount is retained. Otherwise, the historical stable fluctuation width is used as the benchmark correction amount according to the positive or negative direction of the candidate benchmark correction amount. The original expected response benchmark value is added to the benchmark correction amount to generate the candidate expected response benchmark value. The expected environmental parameter values ​​corresponding to the five records before the end of monitoring are regenerated using the candidate expected response benchmark value, and the root mean square value of the residual is recalculated. When the recalculated root mean square value of the residual is lower than the root mean square value of the residual after the operation, the candidate expected response benchmark value is determined as the corrected expected response benchmark value. Otherwise, the correction is canceled, and the original expected response benchmark value is used again. The corrected expected response baseline, state drift, and monitoring results are written into the time-delay coupling graph. Training samples are constructed based on the coupling state before and after the monitoring strategy is executed, the monitoring strategy, and the monitoring results. The time-delay coupling double-constraint BDQ network is then updated. For response nodes that have completed baseline correction, the corrected expected response baseline value, state drift, post-operation state value, monitoring completion rate, and update time are written into the response node attributes. For response nodes that have not completed baseline correction but have generated valid monitoring results, the post-operation state value, state drift, and monitoring completion rate are written into the response node attributes, the original expected response baseline value is retained, the target directed edge corresponding to the monitoring result is read, and the actual response direction is compared with the response direction recorded by the directed edge. When the actual response direction is consistent and the difference between the actual response start time and the time corresponding to the directed edge response delay does not exceed one event sampling cycle, the correct response is considered complete. If the residual after the operation does not exceed the normal residual boundary, the monitoring is recorded as a response confirmation. If any of the above conditions are not met, the monitoring is recorded as a response contradiction. The number of response confirmations for the same directed edge is divided by the sum of the number of response confirmations and the number of response contradictions to generate the updated response credibility. When the same directed edge forms a response contradiction three times in a row, the directed edge is marked as pending review, but its original response direction, response delay and response operator are retained. The directed edge is not directly deleted. The updated time delay coupling graph is generated from the updated response node attributes, directed edge response credibility and directed edge status. The coupling state corresponding to the generation of the monitoring strategy is taken as the pre-execution coupling state. The actual monitoring operations and their action levels are taken as the actual action set. The coupling state regenerated based on the monitoring results and the updated time-delay coupling diagram is taken as the post-execution coupling state. The residual change is limited to between -1 and 1 to generate the monitoring effect value. The monitoring completion rate is limited to between 0 and 1 to generate the execution completion value. The transmission bandwidth utilization rate, monitoring channel utilization rate, and verification device utilization rate are calculated separately during the actual monitoring process. The average of the three utilization rates is taken as the resource utilization value. The monitoring effect value is added to the execution completion value, and then the resource utilization value is subtracted to generate the feedback value corresponding to this monitoring strategy. The pre-execution coupling state, actual action set, feedback value, post-execution coupling state, constraint violation amount, and task completion identifier are combined to generate network training samples.

[0025] Example 1: To verify the feasibility of this invention in practice, it was applied to a mushroom cultivation room for *Pleurotus ostreatus* at a mushroom production base in East China. The cultivation room, approximately 360 square meters in area, was divided into an air intake zone, a central zone, and a return air zone along the air intake direction. It was equipped with a variable frequency blower, an exhaust fan, an ultrasonic humidifier, a chilled water coil, and a supplemental lighting device. Two problems were observed: after the respiration intensity of the culture bags increased, the decrease in carbon dioxide concentration in the central and return air zones lagged significantly behind the equipment's response; and during simultaneous humidification and exhaust operation, humidity disturbances and localized airflow changes easily caused repeated alarms. Nine air monitoring points and six culture medium monitoring points were set up in the cultivation room. The air monitoring points collected temperature, humidity, carbon dioxide concentration, and light intensity, while the culture medium monitoring points collected temperature, moisture content, and pH. Air temperature, air humidity, carbon dioxide concentration, and culture medium temperature were collected at 20-second intervals, and the feedback from the adjustment equipment was recorded at 1-second intervals; after an adjustment event occurred, the carbon dioxide concentration sampling cycle was shortened to 10 seconds. A total of 192 effective regulation events were obtained over 30 consecutive production days, including 78 ventilation events, 64 humidification events, 32 cooling events, and 18 supplementary lighting events. After manual review, 46 abnormal response samples were identified, including 12 abnormal response amplitudes, 9 abnormal response times, 7 abnormal response directions, 10 response operator drifts, and 8 monitoring anomalies pending verification.

[0026] The system establishes an event response sequence around each adjustment event, retaining records of response boundaries, direction changes, and spatial differences, while compressing low-response segments. Subsequently, based on actual equipment feedback and environmental changes at each monitoring point, it calculates the response direction, response time delay, and discrete response kernel, constructing a time-delay coupling diagram. In a single return air zone carbon dioxide anomaly, the initial concentration at monitoring point P09 was 1326 mg / m³. After the exhaust fan was upgraded from level two to level four, the actual response start time was 50 seconds later than expected. The system identified the exhaust fan feedback edge as the primary source of the anomaly by removing unilateral responses edge by edge, and generated P09 sampling cycle adjustment, P08 and P09 synchronous monitoring, and exhaust fan feedback verification operations, ultimately monitoring continuously for 220 seconds at a 10-second cycle. The monitoring completion rate was 96.7%, and the expected residual root mean square value decreased from 62.4 mg / m³ to 28.7 mg / m³. The corrected baseline value was written into the response node and used for the time-delay coupling diagram and network updates. To facilitate the evaluation of the implementation effect of the present invention, the fixed threshold monitoring method, LSTM time series prediction method and the method of the present invention were compared with each other using the same culture area, the same monitoring point, the same regulatory event sample and the same manual verification benchmark. All methods were evaluated using the same criteria. The comprehensive results are shown in Table 1.

[0027] Table 1. Comparison of the overall performance of edible fungi regulation response anomaly tracing and adaptive monitoring

[0028] Table 1 shows that the anomaly identification accuracy of the method of the present invention is 94.7%, which is 16.1 percentage points higher than the fixed threshold method and 6.8 percentage points higher than the LSTM method; the anomaly source localization accuracy reaches 88.5%, which is 34.2 and 16.1 percentage points higher, respectively. The main reason is that the present invention organizes the adjustment device, response node, response direction, response delay, and response operator into a time-delay coupled graph, and then determines the anomaly source by the expected response, counterfactual response, and edge-by-edge removal of one-sided response, which can distinguish between device execution deviation, response hysteresis, and sensor anomaly.

[0029] In response modeling, the mean absolute error of the response delay in this invention is 14.2 seconds, a decrease of 67.6% compared to the fixed threshold method and a decrease of 52.0% compared to the LSTM method; the root mean square error of the expected carbon dioxide response is 27.9 mg / m³, a decrease of 58.6% and 39.1%, respectively. This is because the dynamic response interval retains records of rapid changes and direction reversals, while compressing redundant low-response data, allowing the discrete response kernel to fit the true and effective response. The false alarm rate decreased from 12.8% with the fixed threshold method to 3.4%, and the average anomaly confirmation time was shortened to 3.7 minutes, indicating that forward verification and reverse tracing can reduce repeated alarms caused by single threshold fluctuations.

[0030] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring environmental data of edible fungi based on reinforcement learning, characterized in that, include: Collect and preprocess environmental data, conditioning data, and cultivation status data of edible fungi to generate monitoring data; The regulation events and event response sequences are determined from the monitoring data. The response boundaries are determined based on the event response sequences. The low-response segments between adjacent boundaries are compressed to generate a dynamic response range. Based on monitoring data and dynamic response intervals, determine the direction, time delay, and response operator of the regulation event and environmental parameters, and construct a time delay coupling diagram; Based on the time-delay coupling diagram and monitoring data, the expected response is deduced, and historical events with the same initial state at the same stage are matched. The response operators of the historical events are replaced with the response operators of the current adjustment event to deduce the counterfactual response. Compare monitoring data, expected response, and counterfactual response; perform forward verification and reverse tracing; and generate response residuals, anomaly codes, and candidate monitoring operations. The time-delay coupling graph, anomaly coding, and candidate monitoring operations are input into the coupling layer of the time-delay coupled double-constraint BDQ network to generate the coupling state. The coupling state is then input into the branch decision layer to generate the branch action values ​​for each candidate monitoring operation. Path coverage and timing consistency correction are performed. The corrected branch action values ​​are then input into the constraint layer to exclude unexecutable candidate monitoring operations. Lagrange penalties are applied to the branch action values ​​based on the constraint violations of the remaining candidate monitoring operations to generate the monitoring strategy. The monitoring strategy is executed and monitoring results are generated. Based on the monitoring results, the state drift and residual change are generated and the expected response baseline value is corrected. The time-delay coupling diagram and the time-delay coupling double-constraint BDQ network are updated accordingly.

2. The method for monitoring edible fungi environmental data based on reinforcement learning according to claim 1, characterized in that, The environmental data includes air temperature, air humidity, carbon dioxide concentration, light intensity, and culture medium temperature, moisture content, and pH. The adjustment data includes control commands, start and stop times, set values, and actual feedback values ​​of the adjustment equipment. The cultivation status data includes the edible fungus growth stage, inoculation time, fruiting cycle, growth images, and culture medium configuration information. The preprocessing includes time alignment, outlier removal, missing value completion, and data scaling.

3. The method for monitoring edible fungi environmental data based on reinforcement learning according to claim 1, characterized in that, The generation of the dynamic response range includes: The control status and feedback status of the regulating equipment are read from the monitoring data. The start and end of the regulating event are determined based on the corresponding changes of the two. The equipment operations with the same area of ​​action, the same control direction and the same time are merged into the same regulating event. Based on the start time of the adjustment event, environmental parameter records before, during, and after the event are extracted, and the stable state before the event is used as the starting point for the response to generate an event response sequence. The response rate of each environmental parameter in the event response sequence is calculated according to the effective sampling time. Based on the change of the response rate relative to the historical stable fluctuation range, the effective response segment and the low response segment are divided. The first and last records, direction change records, and spatial difference records of the low response segment are retained. The remaining records are subjected to hierarchical sampling and compression, and a correspondence between the retained records and the original sampling time is established. The effective response segments and the compressed low response segments are connected according to the original time sequence, and the dynamic response interval is generated by dividing the response boundary.

4. The method for monitoring edible fungi environmental data based on reinforcement learning according to claim 1, characterized in that, The generation of the time-delay coupling diagram includes: Generate the actual feedback sequence of the control equipment and the response sequence of environmental parameters based on monitoring data and dynamic response range; The direction of response between the regulation event and the environmental parameter is determined based on the cumulative change of the environmental parameter response sequence relative to the response starting point and the sign consistency of the response rate. Based on the actual action time of the regulating equipment, the actual feedback sequence of the regulating equipment is shifted according to the event sampling period, and the time shift correlation is calculated in combination with the response sequence of environmental parameters. The response time delay is generated according to the time offset corresponding to the peak correlation. Align the actual feedback sequence of the adjustment equipment and the response sequence of environmental parameters according to the response time delay, fit the discrete response kernel according to the change in adjustment input and the offset of environmental parameters, and generate the response operator. Using the regulation event as the event node and the environmental parameters and corresponding monitoring points as the response nodes, a directed edge is established when the response direction, response delay, and response operator meet the validity requirements. The response direction, response delay, response operator, and response confidence are written into the directed edge to construct a time-delay coupling graph.

5. The method for monitoring edible fungi environmental data based on reinforcement learning according to claim 1, characterized in that, The deduction of the counterfactual response includes: The current adjustment input change is shifted according to the response time delay of each directed edge in the time-delay coupling diagram, and the one-sided response of each response node is calculated using the corresponding response operator. The expected response is generated based on the one-sided response. Historical events are selected from those with consistent growth stages, adjusted event types, and response nodes, and whose initial environmental parameter differences do not exceed the corresponding matching limits. Historical events with the same initial state at the same stage are also identified. Replace the response operators of each directed edge of the current adjustment event with the response operators of the corresponding directed edges in the historical events of the same stage and initial state, keeping the current adjustment input change, response nodes and graph connectivity unchanged, recalculate the one-sided response of each response node, and generate counterfactual responses based on the recalculated one-sided responses.

6. The method for monitoring edible fungi environmental data based on reinforcement learning according to claim 1, characterized in that, The generation of the response residual, anomaly code, and candidate monitoring operation includes: Based on monitoring data, expected response, and counterfactual response, the differences between actual environmental parameter values ​​and expected environmental parameter values ​​and counterfactual environmental parameter values ​​are calculated respectively, generating expected residual sequences and counterfactual residual sequences; Perform amplitude verification, direction verification, and response time verification on the expected residual sequence to generate a positive verification result; For a response node that is a positive verification anomaly, remove the corresponding single-edge response one by one along the directed edge pointing to the response node in the time-delay coupling graph, compare the changes in the residual before and after removal, determine the source edge and source event of the anomaly, and generate the reverse tracing result. Based on the expected residual sequence, counterfactual residual sequence, forward verification results, and reverse tracing results, response residuals, anomaly codes, and candidate monitoring operations are generated.

7. The method for monitoring edible fungi environmental data based on reinforcement learning according to claim 1, characterized in that, The generation of the monitoring strategy includes: The time-delay coupling graph, anomaly coding, and candidate monitoring operations are input into the coupling layer of the time-delay coupled dual-constraint BDQ network; The time-delay coupled dual-constraint BDQ network includes a coupling layer, a branch decision layer, and a constraint layer. In the coupling layer, graph relationship aggregation and exception operation fusion are performed to generate a coupled state; The coupling state is input into the branch decision layer, and action branches are established for each candidate monitoring operation. Branch action values ​​are generated for each action branch, and path coverage and timing consistency correction are performed on the branch action values. The corrected branch action value is input into the constraint layer to exclude unexecutable candidate monitoring operations. The constraint violation of the remaining candidate monitoring operations is calculated. Lagrange multipliers are used to apply penalties to the corresponding branch action values. Candidate monitoring operations are selected according to the penalized branch action values ​​to generate a monitoring strategy.

8. The method for monitoring edible fungi environmental data based on reinforcement learning according to claim 1, characterized in that, The updates to the time-delay coupling graph and the time-delay coupling double-constraint BDQ network include: Each monitoring operation is scheduled according to the execution order, monitoring cycle, and duration in the monitoring strategy, and monitoring results are generated. The environmental parameter values, equipment feedback values, and calibration measurement values ​​before and after the monitoring operation are statistically analyzed, and the state drift is generated based on the change range of similar data. Based on the monitoring results, the residuals between the actual environmental parameter values ​​and the expected response are recalculated, and the residuals before and after the monitoring operation are compared to generate the residual change. When the state drift continues and the residuals do not converge, a baseline correction amount is generated based on the deviation between the monitoring results and the expected response, and the baseline correction amount is used to correct the baseline value of the expected response. The corrected expected response baseline, state drift, and monitoring results are written into the time-delay coupling graph. Training samples are constructed based on the coupling state before and after the monitoring strategy is executed, the monitoring strategy, and the monitoring results. The time-delay coupling dual-constraint BDQ network is then updated.

Images