An artificial intelligence monitoring method for an edible mushroom substrate fermentation process

Abstract

The application discloses an artificial intelligence monitoring method for a substrate fermentation process of edible fungi, and belongs to the technical field of intelligent monitoring of edible fungi fermentation. In order to solve the problem that multi-source variable responses are not synchronized and manual inspection is difficult to timely locate abnormalities, the application takes the actuator control time as a common driving event, constructs a dynamic response graph with propagation delay, attenuation and direction based on the multi-source response of the region, extracts the environment, metabolism and phenotype representation and phase consistency residual on the common stage clock, outputs the abnormal region, type and trend through the stage conditional graph time sequence model, generates partition correction suggestions or alarms under the stage correlation constraint, and updates the dynamic response graph by using the post-action response and mycelium phenotype, so that the technical effects of propagation relationship correction, abnormal positioning and closed-loop verification are achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring of edible fungi fermentation, and in particular to an artificial intelligence monitoring method for the fermentation process of edible fungi substrate. Background Technology

[0002] The fermentation process of edible fungi substrates typically requires continuous regulation of temperature, humidity, and aeration conditions, and the fermentation progress is judged by combining carbon dioxide release and mycelial growth status. Existing monitoring methods mostly use a single acquisition channel or a fixed time point as the basis for judgment. There are time differences in airflow propagation, heat and moisture exchange, and metabolic response in different locations, which means that data obtained at the same time may not be in the same response stage.

[0003] Manual inspections cannot continuously cover all substrate areas, and changes in air volume, carbon dioxide, humidity, and temperature caused by fan start-up and shutdown, valve adjustment, and humidification have propagation delays and amplitude attenuation. If this propagation relationship is not identified, environmental changes, metabolic changes, and mycelial phenotypic changes can easily be misaligned, making it difficult to distinguish the location and cause of insufficient ventilation, water content deviations, localized overheating, or stagnation of mycelial expansion.

[0004] Existing anomaly assessments may confuse local ventilation path interruptions and single-channel sensor failures with actual fermentation anomalies. Corrective actions lack safety boundary constraints related to microbial species and fermentation stages, and the propagation response and mycelial phenotypic changes after the actions are not used to continuously update monitoring relationships.

[0005] Therefore, there is a need for an artificial intelligence monitoring method for the fermentation process of edible fungi substrate that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose an artificial intelligence monitoring method for the fermentation process of edible fungi substrate. Addressing the problems of asynchronous responses from multiple variable sources and the inability of manual inspections to promptly identify abnormal locations and causes in existing technologies, this invention proposes a method that uses actuator control timestamps as common driving events. It detects response events from location-registered airflow, carbon dioxide, humidity, and temperature data, calculating propagation delay, amplitude attenuation, and direction to establish a dynamic response map. Based on the fermentation stage and a common stage clock after delay compensation, it generates environmental, metabolic, and phenotypic characterizations and phase consistency residuals, and outputs anomaly diagnosis results and status trends through a stage conditional graph time-series model. Furthermore, it generates zoning suggestions or alarms based on stage-related corrective constraints, and updates the dynamic response map with post-action responses and mycelial phenotypic changes. This invention possesses the technical effects of correcting time misalignments in multi-source data, distinguishing between genuine fermentation anomalies and ventilation or data acquisition failures, and forming a safe closed-loop verification mechanism.

[0007] This invention provides an artificial intelligence monitoring method for the fermentation process of edible fungi substrate, comprising: S1. Acquire environmental data, internal temperature data, and mycelial image sequence that are registered with the substrate within the fermentation space of the edible fungi substrate. Determine the control timestamp of the actuator as a common driving event. Detect the response event corresponding to the common driving event from the environmental data and the internal temperature data of the substrate to form a multi-source response sequence. S2. Based on the location-based multi-source response sequence, determine the response delay, response amplitude attenuation, and response direction from the common driving event to each response event, and construct a dynamic response graph with matrix location as nodes and the response delay, response amplitude attenuation, and response direction as edge attributes. S3. Determine the fermentation stage based on the locational multi-source response sequence and the mycelial image sequence, perform time compensation on the data of each node according to the response delay of the dynamic response map, generate environmental driving characterization, metabolic characterization and phenotypic characterization on the common stage clock, and calculate the phase consistency residual between environment, metabolism and phenotypic on the same node and adjacent response paths to form stage synchronization node features. S4. Input the phase synchronization node features and the edge attributes into the phase condition graph time series model to obtain the anomaly diagnosis results and state trends, including anomaly location, anomaly type and anomaly confidence. S5. Based on the abnormal diagnosis results and status trends, and combined with the phase-related correction constraints, generate partition correction suggestions or alarm information. When the suggestion is executed, obtain the response events and mycelial phenotypic changes after the action, and update the dynamic response map accordingly.

[0008] Optionally, S1 includes: setting location markers according to the spatial location of the substrate layer, substrate pile or culture unit, associating temperature acquisition points, humidity acquisition points, air volume acquisition points or air valve opening acquisition points, carbon dioxide concentration acquisition points, substrate internal temperature acquisition points and mycelium image acquisition areas to the location markers, and associating the spatial location of the fan or air valve and its airflow outlet to the drive source location markers. Record the start time, end time, and amplitude of the fan start / stop, valve opening change, and humidification action, and associate the action with the corresponding drive source location identifier. Within the response detection window after the start of each action, the response threshold is determined based on the baseline before the action and the noise statistics. The response event is determined when the air volume, carbon dioxide change rate, humidity change rate or temperature change rate exceeds the corresponding response threshold and continues to reach the preset number of samplings.

[0009] Optionally, S2 includes: using the difference between the start time of the action of the common driving event and the first time the response event crosses the response threshold as the response delay; According to the airflow reachability relationship and the order of response events, starting from the drive source location associated with the common drive event, the location that has responded before each response location and is reachable by airflow to that response location is determined as the upstream reference node. The response amplitude attenuation is determined based on the ratio of the peak or integral response of the response location within the response detection window to the upstream reference node under the same common drive event. The response direction is determined by the direction from the upstream reference node to the response location. The response delay and response amplitude decay are updated according to the time sliding window. The edge confidence weight is determined based on the ratio of the number of times a corresponding node pair forms a valid response event to the total number of common driving events within the time sliding window. The update time and the edge confidence weight are written into the corresponding edge attribute.

[0010] Optionally, S3 includes: dividing the fermentation stage after the inoculation time according to the turning point when the carbon dioxide release rate changes from rising to stabilizing or falling and the time when the mycelial coverage growth rate changes. The data from each node is shifted to the common stage clock according to the corresponding response delay; The environmental driving characterization is generated by air volume or valve opening, humidity, ambient temperature and carbon dioxide concentration; the metabolic characterization is generated by carbon dioxide release rate and substrate temperature rise rate; and the phenotypic characterization is generated by hyphal coverage, hyphal expansion rate and texture continuity. The phase consistency residual is calculated based on the differences in the direction of change, the time difference of the turning point, and the amplitude of the standardized change within the same stage. The residual of the node itself and the residual of the adjacent nodes aggregated along the directed response path are written together into the stage synchronization node feature.

[0011] Optionally, S4 includes: dividing the fermentation history batches into training samples according to strain and fermentation stage; using the node's environmental driving characterization, metabolic characterization, phenotypic characterization and phase consistency residual as node input; using the response delay, response amplitude decay, response direction and edge confidence weight as edge input; performing time series encoding on the graph message passing results and outputting the anomaly classification probability and state prediction sequence. The anomaly type and anomaly confidence level are determined based on the anomaly classification probability. The anomaly type includes at least insufficient ventilation, water content deviation, local overheating, stagnation of mycelial expansion, and collection anomaly. The state prediction sequence includes at least the predicted temperature, humidity, and carbon dioxide concentration values ​​of each node within a predetermined time window, and the state trend is determined based on the state prediction sequence.

[0012] Optionally, S5 includes: reading the allowable temperature range, allowable humidity range, allowable carbon dioxide concentration range, and automatic correction threshold from the process configuration record associated with the current strain and fermentation stage. The automatic correction threshold is taken when there is an intersection between the abnormal classification probability distributions of the labeled normal batch and abnormal batch; when there are multiple intersections, the first intersection is taken after sorting in ascending order by the sum of the number of false positives and the number of false negatives in the validation set; when there is no intersection, the first abnormal classification probability value is taken after sorting in ascending order by the sum of the number of false positives and the number of false negatives in the validation set. Candidate actuators and candidate action directions are matched according to the anomaly type. The amplitude of candidate actions is determined according to the state trend. Candidate actions are screened out based on the correction constraints of not exceeding the allowable range of the corresponding environment, the rated amplitude of the actuator, and the minimum action interval. When there are no candidate actions that satisfy the correction constraints or the abnormal confidence level does not reach the automatic correction threshold, an alarm message and a manual review request are output; otherwise, the suggestions for zonal ventilation, humidification or cooling are output. Furthermore, after executing the partition ventilation, humidification or cooling recommendations, the execution time is registered as a new common driving event, the response event after the action is detected in step S1 and the corresponding edge attribute is updated in step S2. At the end of the preset phenotypic verification window, the verification results of the action effect are determined based on the changes in hyphal coverage, hyphal expansion speed, and texture continuity. When the action effect verification result indicates that the phenotypic change direction is consistent with the suggested target, the updated edge attribute is retained; when the two are inconsistent, the edge confidence weight of the corresponding response edge is multiplied by the preset weight decay coefficient in the process configuration record and a model verification mark is generated. Furthermore, when the phase consistency residual exceeds the anomaly threshold associated with the current strain and fermentation stage and the anomaly confidence does not reach the automatic correction threshold, the alarm information is retained, and the safety margins of the current temperature, humidity and carbon dioxide concentration relative to their respective environmental allowable range boundaries are read from the process configuration record before entering manual review. The diagnostic action amplitude and duration are determined based on the safety margin and the rated amplitude of the actuator. The predicted values ​​of temperature, humidity and carbon dioxide concentration in the state prediction sequence are determined as the predicted environmental quantities during the diagnostic action. When the predicted environmental quantity does not exceed the boundary of the corresponding environmental allowable range and the minimum action interval is reached from the last action, a limited diagnostic ventilation pulse is applied to the fan or valve that has an airflow reachability relationship with the suspected abnormal location, and a correction or manual review is determined based on the pulse response. If any of the aforementioned conditions are not met, the manual review request is executed directly. Furthermore, taking the start time of the limited diagnostic ventilation pulse as the common time zero point, the arrival time, peak value, integral response, and recovery time of the air volume, carbon dioxide change rate, humidity change rate, and temperature change rate of each matrix location are extracted. The effective attenuation interval is determined based on the statistical interval of the response order and response amplitude ratio of upstream and downstream nodes in the unblocked historical pulse sample. Node pairs that satisfy the condition that the upstream node responds first, the downstream node responds later, and the response amplitude attenuation is within the effective attenuation interval are established or updated as directed edges. For directed edges whose consecutive response missing counts reach the weight reduction threshold, their edge confidence weight is multiplied by the pulse weight reduction coefficient in the device calibration record. For directed edges whose consecutive response missing counts reach the deletion threshold, they are deleted. The weight reduction threshold and the deletion threshold are given by the device calibration record, and the deletion threshold is greater than the weight reduction threshold. Furthermore, anomaly verification is performed using the updated dynamic response map: when the environmental response of a suspected abnormal location reaches the corresponding response threshold, the response sequence matches the response direction of the corresponding directed edge, and the response amplitude decay is within the effective decay range, and at least one component of the metabolic characterization or phenotypic characterization continuously exceeds the preset quantile range of the historical normal samples in the current fermentation stage for a preset number of samplings, it is determined to be a real fermentation anomaly and a partition correction is selected. When the number of times that the air volume channel and at least one of the channels of carbon dioxide concentration, humidity or temperature in the suspected abnormal area fails to reach the corresponding response threshold reaches the weighting threshold, and the downstream node does not generate a corresponding response event, it is determined to be a local ventilation blockage and ventilation path maintenance information is output. When the adjacent nodes of a suspected abnormal location meet the directed edge establishment conditions, the number of times a single acquisition channel of the suspected abnormal location fails to reach the corresponding response threshold reaches the weighting threshold, and the other acquisition channels form corresponding response events, the sensor is determined to be abnormal and sensor maintenance information is output. The judgment conditions that are met simultaneously are processed in the order of sensor abnormality, local ventilation blockage and actual fermentation abnormality. For suspected abnormalities that do not meet the aforementioned judgment conditions, a manual review request is output. Acquire changes in mycelial coverage, mycelial expansion speed, and texture continuity after zonal correction, ventilation path repair, or sensor maintenance; determine the verification results of the review effect; and verify the state changes after processing.

[0013] The beneficial effects of this invention are: 1. By registering actuator actions as common driving events and constructing dynamic response maps based on the delay, amplitude attenuation, and direction of response events in each region, time compensation can be performed on multi-source collected data using the propagation relationships in the actual fermentation space, reducing the misalignment of response stages caused by direct comparison at fixed times.

[0014] 2. By combining environmental-driven characterization, metabolic characterization, hyphal phenotype characterization and phase consistency residuals on a common phase clock, and combining dynamic response diagrams for graph time series inference, the system can simultaneously provide the location of anomalies, anomaly type, anomaly confidence level and state trend, supporting the differentiation of insufficient ventilation, water content deviation, local overheating, hyphal expansion stagnation and collection anomalies.

[0015] 3. Through phase-related corrective constraints and amplitude-limited diagnostic ventilation pulses, the propagation path can be verified within the environmentally permissible range and actuator conditions. The edge attributes are updated based on the response events and hyphal phenotypic changes after the action, thus providing a traceable closed-loop basis for zonal correction, ventilation path inspection, sensor maintenance, or manual verification. Attached Figure Description

[0016] 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 the artificial intelligence monitoring method for the fermentation process of edible fungi substrate according to the present invention.

[0017] Figure 2 This is a flowchart of constructing the dynamic response graph in S2 of the present invention. Detailed Implementation

[0018] 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.

[0019] refer to Figures 1-2 An artificial intelligence monitoring method for the fermentation process of edible fungi substrate, comprising: S1. Acquire environmental data, internal temperature data, and mycelial image sequence that are registered with the substrate within the fermentation space of the edible fungi substrate. Determine the control timestamp of the actuator as a common driving event. Detect the response event corresponding to the common driving event from the environmental data and the internal temperature data of the substrate to form a multi-source response sequence. S2. Based on the location-based multi-source response sequence, determine the response delay, response amplitude attenuation, and response direction from the common driving event to each response event, and construct a dynamic response graph with matrix location as nodes and the response delay, response amplitude attenuation, and response direction as edge attributes. S3. Determine the fermentation stage based on the locational multi-source response sequence and the mycelial image sequence, perform time compensation on the data of each node according to the response delay of the dynamic response map, generate environmental driving characterization, metabolic characterization and phenotypic characterization on the common stage clock, and calculate the phase consistency residual between environment, metabolism and phenotypic on the same node and adjacent response paths to form stage synchronization node features. S4. Input the phase synchronization node features and the edge attributes into the phase condition graph time series model to obtain the anomaly diagnosis results and state trends, including anomaly location, anomaly type and anomaly confidence. S5. Based on the abnormal diagnosis results and status trends, and combined with the phase-related correction constraints, generate partition correction suggestions or alarm information. When the suggestion is executed, obtain the response events and mycelial phenotypic changes after the action, and update the dynamic response map accordingly.

[0020] In this specific embodiment, S1 includes: This specific implementation takes a fermentation room for edible fungi substrate with four ventilation branches, three layers of culture racks and forty-eight culture units as the monitoring object. The data processing server sets location identifiers according to the spatial location of the substrate layer, substrate pile or culture unit, and establishes a location master record for each culture unit. The location master record includes batch identifier, strain identifier, culture rack number, layer number, unit number, three-dimensional installation coordinates, substrate loading quality and a list of associated equipment. The list of associated equipment records the equipment identifiers of environmental temperature sensor, relative humidity sensor, carbon dioxide sensor, wind speed sensor, temperature probe inserted into the center of the substrate and top-view mycelium camera, and registers the fan, air valve, humidification valve and its airflow outlet as the drive source location. The server saves the original record with the location identifier, equipment identifier and collection timestamp as a joint key. Ambient temperature, relative humidity, carbon dioxide concentration, wind speed, and substrate internal temperature are collected at a 30-second sampling cycle, while mycelial images are acquired at a 30-minute sampling cycle. The controller, acquisition gateway, and camera are all synchronized with the same network clock. The server verifies the device time using the gateway's received time. When the device time deviates from the network clock by more than two seconds, the device timestamp is corrected according to the median of the time deviations of the most recent 20 synchronization messages. The time before correction, the time after correction, and the time quality flag are written into the acquisition record, thereby enabling environmental data, substrate internal temperature data, and image sequences to be registered with batch identifiers and location identifiers. For each continuous acquisition channel, the server first removes readings that exceed the sensor's range, then replaces isolated pulses with the median filtering result of five sampling points. When two consecutive sampling points are missing, linear interpolation of the effective readings before and after is used. When more than two consecutive sampling points are missing, the channel is marked as unusable in the current response detection window. The image acquisition record saves the image file address, location identifier, shooting time, exposure parameters, and sharpness score. When the sharpness score is lower than the 0.75 threshold obtained by camera calibration, the original image is retained and a re-acquisition flag is written. The image is not used to update the mycelial phenotype. The controller generates common drive event records for fan start / stop, valve opening changes, and humidifier valve actions. Record fields include event identifier, drive source location identifier, actuator identifier, action type, action start time, action end time, pre-action setpoint, post-action setpoint, and action amplitude. Fan action amplitude is expressed as a percentage of rated speed, valve action amplitude as an opening percentage, and humidifier action amplitude as the valve opening duty cycle. The server uses the action start time as the zero point of the common drive event time and reads the expected response polarity from the device calibration record according to the actuator identifier, drive source location, action direction, and acquisition channel type. The expected response polarity specifies whether airflow, carbon dioxide concentration, relative humidity, ambient temperature, and substrate internal temperature should increase, decrease, or not be determined under the corresponding action. When the same control command generates multiple actuator actions within ten seconds of each other, a composite common event record with actuator sub-records is formed. For common-driven events, the server first assigns responses based on the reachable airflow locations and expected response polarity of each sub-action. Only when the affected location sets do not overlap or the arrival times are separable will the response be assigned to the corresponding sub-action. Synchronous multi-source responses that cannot be separated are written to the multi-source inseparable flag and are not used for automatic edge building, control gain, or dynamic response template updates, but are transferred to manual review. For all common-driven events, the server also checks for window conflicts of different control commands when establishing response detection windows. If the response detection windows of two events overlap and the reachable airflow locations of the actuators intersect, the response is assigned only when the responses of each event to each channel can be uniquely separated based on the calibrated arrival times or the corresponding version of the multi-action superimposed response model. Otherwise, the related events and overlapping time periods are written to the event pollution flag. Responses during overlapping time periods cannot be used for automatic edge building, edge attribute updates, control gain updates, or dynamic response template updates, and are transferred to manual review. For each common driving event, the server uses data from 20 minutes prior to the action's start time to form a pre-action baseline window, and data from 60 minutes after the action's start time to form a response detection window. Raw data is used for airflow, while the first-order rate of change of adjacent sample values ​​is used for carbon dioxide concentration, relative humidity, ambient temperature, and matrix internal temperature. At sampling time rate of change calculate, Indicates channel At sampling time rate of change and unit is channel The data collection unit is per second. Indicates channel At sampling time The collected values ​​are in channels. The collection unit, This represents the value collected at the previous sampling time, and the unit is also channel. The data collection units are as follows: different channels use their own units for airflow, concentration, relative humidity, or temperature. This represents the time interval between adjacent sampling moments, and in this specific embodiment, it is thirty seconds. A dimensionless channel index representing airflow, carbon dioxide concentration, relative humidity, ambient temperature, or internal substrate temperature. Represents a dimensionless integer sampling index that increments over time; The server determines the response threshold based on the number of similar detections within the baseline window prior to the action, and then... Calculate the positive response threshold and press Check for negative responses. Indicates channel The response threshold relative to the baseline before the action, with units consistent with the corresponding detection quantity. This represents the median of the absolute deviation of the detections within the baseline window before the action, with units equal to the number of channels. The corresponding detection volume is consistent. This represents a robust noise scale obtained by multiplying the median absolute deviation by 1.4826, with units equal to the channel. The corresponding detection volume is consistent. Indicates channel The dimensionless threshold multiples for air volume, carbon dioxide change rate, humidity change rate and temperature change rate are respectively calibrated by the 99th percentile of the noise quantile of 30 batches of historical data without abnormal fermentation and stored in the threshold configuration table. In this specific embodiment, the values ​​are 3.2, 3.0, 3.0 and 3.5 respectively. Starting from the zero point of the time of the common driving event, the response detection window is checked point by point. When the deviation of the air volume relative to the median before the action, the rate of change of carbon dioxide, the rate of change of humidity, or the rate of change of temperature of a certain location exceeds the expected response polarity of the corresponding channel along the actuator sub-record and the three consecutive sampling points are established, the first threshold sampling time is determined as the first arrival time, the maximum absolute deviation of the relative baseline within the continuous threshold interval is determined as the response peak value, and the integral response is calculated by trapezoidal integral. If the threshold direction is opposite to the expected response polarity, the reverse response direction is recorded. If the expected response polarity is not determined by direction, the positive and negative thresholds are checked respectively and the measured direction is saved. If multiple channels meet the conditions at the same time, channel response sub-records are formed and associated with the same response event identifier. When the airflow channel corresponding to the operation of the fan or damper is unavailable, the server does not replace the evidence of airflow arrival with the threshold exceeding the threshold of carbon dioxide, humidity or temperature. Instead, it marks the event of that location as pending verification. When only one environmental channel is unavailable in the non-driving source location, the other valid channels are used to detect the response and the missing channel mask is saved in the record. When all environmental channels do not exceed the threshold within the response detection window, a response missing record is generated. The response missing record still retains the common driving event identifier, location identifier and window end time for subsequent edge weighting and edge deletion judgment. Before each batch of fermentation begins, the location registration relationship is checked by the maintenance terminal. The maintenance personnel trigger a short-term opening change of 5% for each air valve in sequence and check the location of the drive source and the first air volume to reach the location. Then, a frame image is taken using a calibration board with location code to check the camera area. If any of the equipment association, spatial location or camera area is inconsistent, the automatic response detection for that batch is stopped and the conflicting equipment is written into the registration verification list. After the verification is completed, a new location configuration version is generated and the historical acquisition records continue to retain the original version. Each record in the threshold configuration table uses the strain identifier, fermentation stage, channel type, and sensor model as the lookup key. It stores the baseline window length, response detection window length, threshold multiplier, number of consecutive samples, noise scale unit, calibration batch range, and version number. The server matches the threshold configuration according to the current record key. If there is no completely matching record, only alarm-level response candidates are generated, but the candidate events are not used to automatically update the dynamic response map. After the same type of sensor has accumulated no less than ten batches of normal data, the threshold multiplier is recalibrated and verified by reserved batches before it is enabled. The server sorts response events according to common driving event identifier, location identifier, and channel type. It writes the action start time, first arrival time, response direction, response peak, integral response, time of recovery to within the threshold, data quality flag, and missing channel mask into the location multi-source response sequence. The response events after the action are used as new batch-in-batch measured records for S2 to update the dynamic response map. Among them, the actions triggered by the correction suggestion are marked as closed-loop correction in the event source field, and the actions triggered by the amplitude-limited diagnostic ventilation pulse are marked as diagnostic pulse. The resulting location multi-source response sequence, together with the registered hyphal image sequence, is read by S2 and S3. After executing the suggested zonal ventilation, humidification, or cooling, the controller writes the suggested execution time, actual actuator feedback time, and actual action amplitude back to the original suggestion record. The server registers a new common drive event with the actual actuator feedback time and restarts the response detection window. When the feedback time is missing or the actual amplitude deviates from the suggested amplitude by more than 5% of the rated amplitude, the event is not merged with the normal production action. Instead, an actuator feedback abnormality flag is output. S2 only uses the action response with a valid feedback time to update the corresponding edge attribute.

[0021] In this specific embodiment, S2 includes: The graph construction module reads the location multi-source response sequence formed by S1 and reads the reachability relationship between the fan outlet, air valve, air duct branch and culture unit from the airflow topology configuration of the fermentation chamber. It establishes a dynamic graph record with the location identifier as the node key. The node fields include location coordinates, sensor availability mask, recent common driving event identifier and recent response time. Candidate directed edges are generated only when the upstream location is reachable from the downstream location airflow and the two form the same type of response under the same common driving event. For common driving events and response location Response delay calculate, Indicates the response location In public driving events The response delay is shown below, and the unit is seconds. This indicates the moment when the location first consecutively crosses the response threshold, using the server timestamp after S1 time synchronization, employing the same time base as the event time, and measured in seconds. Indicates common driving events The action starts at the time specified by the server timestamp after S1 synchronization, using the same time base as the first arrival time and measured in seconds. Dimensionless index representing matrix location nodes. The dimensionless index represents the common driving event. The missing response record that is missing the first arrival time is not included in the delay statistics but is counted in the number of missing responses of the corresponding candidate edge. The graph construction module performs a breadth-first search, starting from the location of the driving source associated with the common driving event, according to the order of airflow reachability relationships and response events, and then searches for the response locations. The node with the latest first arrival time among the nodes that responded earlier than the location and can directly reach the location by airflow is selected as the upstream reference node. If the arrival times are the same, the node with the higher historical edge confidence weight is selected. If they are still the same, the node with the shorter wind tunnel geometric distance is selected. Candidate directed edges are established from the upstream reference node to the response location, so that the edge direction simultaneously satisfies the airflow reachability relationship and the response sequence relationship. For similar response quantities, a robust fusion value of the peak and integral responses is used, according to... calculate, Indicates the response location In public driving events The fusion response quantity is given below, and the unit is consistent with the acquisition quantity of this channel. This represents the peak response relative to the baseline before the action, and the unit is the acquisition unit of the corresponding channel. It represents the time integral of the absolute deviation from the baseline within the response detection window, and the unit is the acquisition unit of the corresponding channel multiplied by seconds. This indicates the duration of the response detection window, in seconds. The peak fusion weight is a dimensionless quantity. In this specific implementation, it is calibrated to 0.6 by the repeatability verification of non-blocking historical actions. The peak value and integral response are only compared within the same acquisition channel and the same common driving event. From the upstream reference node Pointing to the response area The response amplitude attenuation is based on calculate, This represents the dimensionless index of the upstream reference node of the candidate edge. Indicates candidate edges in common driving events The dimensionless response amplitude ratio under the following conditions This represents the fused response of the upstream reference node under the same event, and the unit is . and same, Indicates channel The same unit of zero protection quantity is taken as one percent of the median of the historical fusion response quantity of the channel without blockage. The response amplitude ratio is allowed to be greater than one and is marked as local amplification when it is greater than one. The device calibration record saves the effective amplitude ratio range by channel. Records outside the range are written with an amplitude over-limit flag and do not participate in the representative response amplitude attenuation and automatic correction update. When the upstream fusion response quantity is not greater than the zero protection quantity or the upstream and downstream channel types are inconsistent, the amplitude ratio record is discarded but the response sequence record is retained. The dynamic graph uses the most recent twenty common driving events as a time sliding window. The edge represents the response delay, which is the median of the effective delay samples, and the edge represents the response amplitude decay, which is the mean after removing the upper and lower 10 percent extreme values. Only when When the value is greater than zero, the edge confidence weight is then applied. calculate, Indicates from the upstream node to downstream nodes The dimensionless edge confidence weights range from zero to one. This indicates the number of common driving events that, within the sliding window, satisfy the requirements of airflow reachability, correct response sequence, and formation of valid, similar responses for that node. This represents the total number of common driving events that can reach this node pair within the sliding window; when When the weight is zero, the division operation is blocked before evaluation. The edge state is recorded as to be accumulated and the previous valid weight is retained. If there is no previous valid weight, the weight field is set to empty, and the reason for the unreachable event is recorded. When the weight is greater than zero but less than five, the weight is still calculated and saved for traceability, but the edge state is recorded as to be accumulated and is not used for automatic correction. Each edge record saves the upstream location identifier, downstream location identifier, channel type, representative response delay, representative response amplitude attenuation, response direction, edge confidence weight, number of valid samples, number of missing responses, last update time, and parameter version. The response direction is fixed as upstream location to downstream location. If different channels of the same node pair produce different representative delays, the channel edges are saved separately and the corresponding delay is selected according to the channel used for representation during S3 time compensation. When the channel edge confidence weight is less than 0.4, it is only used for manual viewing and does not participate in path residual aggregation. For the limited diagnostic ventilation pulse executed by S5, the graph construction module recalculates the arrival time, peak value, integral response and recovery time of the air volume, carbon dioxide change rate, humidity change rate and temperature change rate of each location at the pulse start time. It also reads the effective attenuation intervals layered by bacterial species, loading density and air valve opening from the equipment calibration record. Only nodes that respond first upstream and then downstream and whose amplitude ratio falls into the effective attenuation interval are established or updated with directed edges. No new edges are created and calibration supplement flags are generated when the calibration layer is not matched. The equipment calibration record also saves the threshold for the number of times of weight reduction, the threshold for the number of times of edge deletion, and the pulse weight reduction coefficient. In this specific implementation, these are three, five, and 0.7, respectively. The threshold for the number of times of edge deletion is greater than the threshold for the number of times of weight reduction. When the number of consecutive missing responses reaches three, the confidence weight of the original edge is multiplied by 0.7 and written into the reason for weight reduction. When it reaches five, the directed edge is deleted and the attributes before deletion are transferred to the edge history table. When a new valid response arrives, the number of consecutive missing responses is cleared to zero before the sliding window update is performed. After executing the partitioned ventilation, humidification, or cooling recommendations, the graph construction module filters the response events after the action based on the target location and actuator feedback identifier in the recommendation record. It then links the graph version before the action, the first response after the action, and subsequent phenotypic verification records in series using the common driving event association key. If there is another unmerged actuator action during a recommendation period, the interfered response event is excluded from the current edge update sample while retaining the interference event flag, thus avoiding attributing the propagation of concurrent actions to the current recommendation. Before submitting a dynamic graph version, edge consistency checks are performed. Candidate edges are deduplicated according to upstream location, downstream location, and channel type. If the representative delay of an edge in the same channel is not positive, the edge confidence weight exceeds zero to one, or the response direction violates the airflow reachability relationship, the submission of the candidate version is stopped and the previous valid graph snapshot is restored. At the same time, the failed edges and their common driving event ranges are written into the graph construction review table. Only versions that pass the check are available for time compensation and anomaly review. After each normal control action or diagnostic pulse processing is completed, the graph construction module generates a dynamic response graph snapshot with a version number. The snapshot includes a node table, a channel edge table, edge history references, and the current sliding window event range. The version number is written back to the corresponding common driving event record. S3 reads the latest dynamic graph snapshot with valid edge status in the same batch for time compensation. When the action effect is inconsistent with the target, S5 continues to perform confidence weight decay and model verification marking for the edges involved in this action.

[0022] In this specific embodiment, S3 includes: The phase synchronization processing module reads the locational multi-source response sequence of S1, the registered mycelial image sequence, and the dynamic response map snapshot of S2. Using the inoculation completion time as the batch relative time starting point, it aggregates environmental and substrate temperature data every ten minutes and image phenotypes every thirty minutes. Each ten-minute window is associated with only one common driving event at the end of the window. Candidate events must meet the following conditions: their actual execution feedback time is no later than the end of the window; their response detection window covers the current window; and they are not set with a multi-source inseparable flag or an event contamination flag. If multiple candidate events exist, the event with the latest actual execution feedback time, a true quality validity flag, and the smallest lexicographical order of the event identifier is selected sequentially. If no candidate events exist, an unassociated event flag is written, and a compensation edge is selected from the same channel edge of the current valid dynamic map version. When the current ten-minute window intersects with any contamination period recorded in S1, the window is not compensated for a single event delay. All nodes affected by the overlapping response are written into the event contamination hard mask, the automatic diagnosis qualification flag is set to false, and the corresponding window... Excluded from stage transition statistics, model training, online graph message passing, effective node ratio calculation, abnormal location selection, and S5 automatic diagnosis, until entering a subsequent window that does not intersect with the contaminated period; when aligning the image phenotype to the ten-minute common stage clock, only the phenotype value of the most recent effective image before the end of the current ten-minute window is used, and the original image time and phenotype age are recorded, without using future images for interpolation; if the phenotype age is no more than thirty minutes, the value is maintained in the subsequent ten-minute window; if it exceeds thirty minutes, the image is of low confidence, or there is no effective image, a phenotype missing mask is set and the value is not used for stage transition determination; training and online inference use the same resampling rules, and aggregate records save batch identifier, location identifier, window start and end time, effective sample count, original image time, phenotype age, resampling method, missing channel mask, event contamination hard mask, associated event identifier, event selection basis, unassociated event flag, and dynamic graph version number; windows with an effective sampling ratio of less than 80% retain the missing flag and are not used for stage transition determination; Carbon dioxide release rate is calculated based on the location-specific exhaust volume and the difference in carbon dioxide concentration between the intake and exhaust air, according to... The carbon dioxide release rate per unit matrix volume was obtained. Indicates location In the aggregation window The carbon dioxide release rate, expressed in milligrams per cubic meter per minute. This represents the converted exhaust volumetric flow rate for that location, expressed in cubic meters per minute. It is primarily obtained by multiplying the median effective wind speed of the exhaust outlet in that location within the same window of S1 by the effective flow area in the equipment calibration record and then converting it to per minute. If the location wind speed is unavailable, it is obtained by multiplying the total exhaust volumetric flow rate measured by the fan flow meter in the same window by the flow distribution coefficient obtained from the duct balance calibration record based on the location identifier and damper opening. The calibration record stores the equipment identifier, location identifier, damper opening level, effective flow area, flow distribution coefficient, version number, and validity period. This indicates the carbon dioxide concentration in the exhaust air at a given location, expressed in milligrams per cubic meter. This indicates the carbon dioxide concentration in the air entering through the same window, expressed in milligrams per cubic meter. Indicates location The effective substrate volume, in cubic meters, is calculated by multiplying the effective bottom area of ​​the culture rack in that location by the average height of the batch loading. This value is then entered into the batch location configuration record and must be greater than zero. This represents the stage aggregation window index in S3, which increments by ten minutes. It is a dimensionless integer derived from the current batch's stage aggregation window sequence and is independent of the sampling index in S1, which increments by thirty-second sampling times. When both exhaust volume acquisition paths are unavailable, or the calibration version is invalid... Missing If the value is zero or its size source fails the batch verification, it will not be calculated. ,Will If the carbon dioxide release rate field is left empty and a release rate missing mask is set, this window will not enter the stage transition judgment and automatic diagnosis. The image processing module scales the mycelial image to 1024 x 1024 pixels and uses a U-Net segmentation network trained with mycelial regions manually annotated in historical batches of the same species to output a mycelial mask. The training set, validation set, and test set are divided into batches in a ratio of 7:2:1. The loss function uses a weighted combination of cross-entropy and Dice loss. When the Dice coefficient of the validation set is not lower than 0.9, the model version is fixed. During inference, pixels with a segmentation probability of not lower than 0.5 are identified as mycelial pixels. When the image sharpness score is lower than 0.75 or the sudden change in the segmentation mask area exceeds 20% of the median of two adjacent frames, the previous valid phenotype is retained and a low confidence flag is written. The hyphal coverage is determined by the ratio of the number of hyphal pixels to the number of effective matrix region pixels, ranging from zero to one. The hyphal expansion rate is determined by dividing the coverage difference between adjacent effective images by the shooting time interval, with the unit being one hour. The texture continuity is determined by the ratio of the length of the maximum connected component of the hyphal mask skeleton to the total skeleton length, ranging from zero to one. The phenotypic record saves the coverage, expansion rate, texture continuity, image model version, and low confidence markers. Five frames of Savitzky-Golay smoothed results are used for stage transition detection while retaining the original results for traceability. The phase synchronization processing module performs piecewise linear fitting on the median CO2 release rate and the median mycelial coverage growth rate of the effective locations in the entire chamber. After inoculation, the interval where the slope of the CO2 release rate is positive for six consecutive 10-minute windows is marked as the warming metabolism phase. When the absolute value of its slope drops to within 20% of the aforementioned positive slope median and the coverage growth rate is still positive, it is marked as the mycelial expansion phase. When the slope of the CO2 release rate is negative for six consecutive windows or the coverage growth rate drops to below 30% of the peak value of this batch, it is marked as the mature and stable phase. If the candidate inflection points of the two signals are no more than six hours apart, the midpoint between the two is taken; otherwise, the batch is marked as a phase pending verification. The phase determination is only performed when the number of effective locations in the same aggregation window reaches 80% of the number of locations configured in the entire chamber. The number of locations corresponding to 80% is determined by rounding up. If the effective location set is empty or the number of effective locations has not reached the lower limit, the phase determination is performed. If, after inoculation, six consecutive windows with the median effective carbon dioxide release rate and the median effective mycelial coverage growth rate have not yet formed, or if the initial stage conditions have not been met, undefined median or piecewise fitting will not be performed. The stage number will be set to "Stage to be determined," and the data missing mask and automatic diagnosis qualification flag for the stage features will be set to true and false, respectively. The reason for waiting and the channel to be supplemented will be recorded, and the data will not be submitted to the S4 automatic diagnosis queue. If a stage that has passed quality verification already exists, and if the number of effective regions in the current window is insufficient or there are fewer than six consecutive effective windows for stage fitting, only the most recent effective stage number will be retained for manual review. Old slopes, old inflection points, or old residuals will not be reused. The corresponding numerical field for the current stage feature will be set to empty, and the missing mask and automatic diagnosis qualification flag will be set to false. The data will be recalculated in chronological order after the effective windows are supplemented. If there is no most recent effective stage or feature snapshot, a data quality alarm and a manual review request will be generated directly. For nodes in the dynamic response graph and channels The stage synchronization processing module first selects compensation edges from the valid incoming edges of the same channel on the reachable path of the event when there are related events; when there are no related events, it selects compensation edges from all valid incoming edges of the same channel in the current valid dynamic graph version. In both cases, the edge with the highest confidence weight is selected first. If the confidence weights are the same, the edge with the last update time, the shorter air duct geometric distance, and the edge with the smaller lexicographical order of the edge identifier are selected in that order to uniquely determine the compensation edge and its representative response delay. The acquisition sequence is shifted to the common stage clock according to the representative response delay of the selected channel edge, and the compensation value is calculated according to... Sure, Represents a node passage During the common phase The compensation value and the channel The collection unit, This represents the original timeline acquisition sequence of the same node and channel at time [time value missing]. The value is taken and the channel is used. The data collection unit, with the superscript "comp" indicating the identity of the compensated sequence and the superscript "raw" indicating the identity of the original sequence. This indicates the common phase moment, measured in seconds, based on the associated common driving event or the start of the batch phase. This represents the representative response delay of the selected channel edge, in seconds. The compensation waiting deadline for each stage window is the sum of the window end time, the selected representative response delay, and one sampling period. The window feature is delayed before the deadline, and only when the deadline is reached... Linear interpolation is performed when the sample is located between the earliest and latest valid times in the existing acquisition sequence and when two adjacent valid sample values ​​exist before and after it. If the value is earlier than the earliest valid time, later than the latest valid time, or if any adjacent valid sample value is missing, extrapolation is not performed and the old value is not maintained. Instead, the compensation value is set to empty and written into the compensation data missing mask and the reason for going out of bounds. At the same time, the automatic diagnosis qualification flag is set to false and the aforementioned hard mask is applied. If there is no unique valid edge, the translation is also not performed and the uncompensated flag and the reason for no valid edge are written into the mask. At the same time, the automatic diagnosis qualification flag is set to false. The node sequence is excluded from model training, online graph message passing, valid node ratio calculation, abnormal location selection and S5 automatic diagnosis through hard mask. The selected edge identifier or the reason for no valid edge is written into the stage feature record. Environmental-driven characterization consists of compensated airflow or damper opening, relative humidity, ambient temperature, and carbon dioxide concentration. Metabolic characterization consists of carbon dioxide release rate and substrate temperature rise rate. The substrate temperature rise rate is calculated by dividing the change in the difference between the internal substrate temperature and the ambient temperature over a ten-minute window by the window duration, with units in degrees Celsius per hour. Phenotypic characterization consists of hyphal coverage, hyphal expansion rate, and texture continuity. Each component is robustly standardized using the median and interquartile range of normal historical samples for the current strain and fermentation stage, and truncated to -5 to +5. When the interquartile range is zero, the smallest scale in the same component calibration table is used. Normal historical statistics are only used when there are at least five historical batches that have completed quality verification and been judged to be normal for the same strain and the same stage, and when each component to be calibrated, slope, or inflection point has at least thirty valid stage window samples. This unified quality threshold applies simultaneously to the median and interquartile range, the 75th percentile of the absolute value of the slope, and the 95th percentile of the inflection point time difference required for robust standardization. When the corresponding historical sample set is empty or any quantity threshold is not met, the above statistics are not calculated or reused across species or stages, and the unverified default scale is not used. The relevant standardized characterization and phase consistency residual numerical fields are written to zero and a normal historical insufficiency mask is set so that they do not participate in statistics, loss, or graph message contribution. At the same time, the automatic diagnosis qualification flag is set to false, and no data is submitted to the S4 automatic diagnosis queue. A data quality alarm and manual review request are generated, which include missing statistics, actual sample number, required sample number, and current calibration version. Only after subsequent normal batches confirmed by manual review meet all thresholds are new statistical versions of the same species and stage fixed and calculation resumed. For nodes Characterization and stage window , This represents the stage aggregation window index, which increments by ten minutes within S3 to S4. It is a dimensionless integer derived from the current batch's stage aggregation window sequence and increments by time starting from the first valid stage window. It is independent of the sampling index, which increments by sampling time in S1. The stage synchronization processing module first takes the median of the components for the three dimensionless characterization vectors: environment, metabolism, and phenotype. It then performs a linear fit using the five most recent valid values ​​up to the current window. For each characterization, the 75th percentile of the absolute value of the slope of normal historical samples of the same species at the same stage is recorded as the non-negative dead zone amplitude. A positive slope dead zone explicitly uses this non-negative dead zone amplitude, and a negative slope dead zone explicitly uses the opposite of this non-negative dead zone amplitude. When the fitted slope is higher than the positive dead zone for three consecutive windows, the direction of change is determined to be upward. When the window is below the negative dead zone, it is determined to be a descent; otherwise, it is determined to be stationary. When the direction first switches from one of the three—ascending, descending, or stationary—and the new direction is maintained for three consecutive windows, the start time of the first window in these three windows is determined as the turning point. If there are multiple candidate turning points in the current stage, the one closest to the current window is selected. If the candidate turning points are the same, they are selected according to a fixed order of environment, metabolism, and phenotype. If any representation has no effective turning point in the current stage, a turning point missing mask is set, the phase consistency residual value field containing that representation is written to zero, and the residual of the previous window is prohibited from being used. Both training and online inference use this mask to prevent the placeholder value from participating in residual statistics, loss calculation, effective node proportion, and graph message contribution. For representation pairs without a turning point missing mask, the phase consistency residual is calculated as follows: calculate, Representation With representation The dimensionless phase consistency residual between them, with larger values ​​indicating greater asynchrony. The value is obtained from the previously determined direction of change. When one of the two characteristics is rising and the other is falling, the value is 1; otherwise, the value is zero. The difference is a dimensionless direction difference with a set of values ​​of zero or one. and The values ​​are obtained from the aforementioned transition detection, representing the most recent transition time difference between the two representations relative to the start of the current stage, in hours. The normalized scale representing the transition time difference is defined as six hours based on the 95th percentile of the normal historical sample. and These represent the dimensionless standardized amplitude changes of the two characteristics between adjacent stage windows. This indicates a dimensionless normalized scale for the magnitude difference, and its value is two. , and The dimensionless weights representing the direction difference, turning point time difference, and amplitude difference are minimized by the validation set. The variances of the normal batch residuals are calibrated as 0.3, 0.4, and 0.3. These represent the dimensionless category identifiers for environmental, metabolic, and phenotypic characterizations, respectively. If any environmental, metabolic, or phenotypic characterization has fewer than five accumulated valid values ​​in the current stage, linear fitting for that category is not performed; the direction of change and the turning point for that category are written as missing, and a turning point missing mask is set. If a stage number already exists, only that stage number is retained; the old slope, old direction, old turning point, or old phase consistency residuals for that category are not used. The residual value field containing that category of characterization is set to zero and excluded by the mask. The automatic diagnosis eligibility flag is set to false, and the number of valid values ​​waiting to be supplemented is recorded. Only after all three categories of characterizations have reached five valid values ​​and the subsequent direction continuity condition is met are the residuals recalculated and automatic diagnosis eligibility restored. If the first stage has not yet been formed or there is no most recent valid stage, data quality alarms and manual review requests are directly generated according to the aforementioned stage-to-be-determined path. When aggregating the residuals of adjacent paths along the dynamic response graph, only incoming edges with an edge confidence weight of not less than 0.4 are used. For the three sets of representations of a node itself, the maximum value of the residual is taken as its own residual. The residuals of the upstream nodes of the first hop and the second hop are weighted by the product of the edge confidence weights on the path and divided by the weight sum. When the weight sum is zero, the path aggregation residual is set to zero and a no-valid-path flag is set. This allows the environmental, metabolic and phenotypic phases that continuously appear on the propagation path to enter the current node features without treating low-confidence edges as deterministic relationships. When a stage transition judgment conflict occurs, the dynamic graph version verification fails, or any core representation is unavailable for three consecutive aggregation windows, the stage synchronization processing module stops submitting the current node sequence to the automatic diagnosis queue, restores and retains the stage number and feature snapshot that passed the quality verification most recently, and writes the failure reason, missing window range and the channel to be supplemented into the stage review record. After the supplemented data arrives and is recalculated, it switches to the new stage feature. The phase synchronization node feature record ultimately saves the phase number, phase relative time, environment-driven representation vector, metabolic representation vector, phenotypic representation vector, original phenotypic image time, phenotypic age, phenotypic resampling method, phase consistency residuals of the three sets of nodes, path aggregation residuals, data missing mask, transition missing mask, low confidence image flag, uncompensated flag, automatic diagnosis qualification flag, associated event flag, event selection basis, compensation selected edge flag, and dynamic graph version number. Features are written into the phase feature library according to the location flag and phase window order to form a sequence of twelve fixed-length windows. If there are less than twelve windows, zeros are padded at the beginning of the sequence and a mask is set. The sequence and the edge attributes of the same graph version are jointly processed by the phase condition graph time series model of S4.

[0023] In this specific embodiment, S4 includes: The model training module divides the fermentation history batches into training samples according to strain and fermentation stage. It constructs stratified samples from the historical fermentation batches that have completed quality verification, and divides them into 70% training set, 15% validation set and 15% test set by batch to avoid cross-set leakage between adjacent windows of the same batch. The node input is the environment-driven characterization, metabolic characterization, phenotypic characterization, phase consistency residual and missing mask of twelve consecutive stage windows. The edge input is the response delay, response amplitude attenuation, response direction, edge confidence weight and edge channel type. The abnormal labels of the training samples are determined by a combination of fermentation records, manual verification forms, ventilation maintenance forms, sensor maintenance forms, and post-action phenotypic verification results. The categories include normal, insufficient ventilation, water content deviation, local overheating, mycelial expansion stagnation, and collection abnormality. The abnormal location is indicated by the verified location identifier. The state prediction monitoring measures the ambient temperature, relative humidity, and carbon dioxide concentration for the next six ten-minute windows. When there is a conflict in the records, the labels are determined in the order of sensor maintenance evidence, ventilation maintenance evidence, and fermentation abnormality evidence. Samples with insufficient evidence are not included in the monitoring loss calculation. The stage-conditional graph temporal model consists of a two-layer edge attribute attention graph network, a 64-dimensional stage embedding layer, and two layers of gated recurrent units. Node features are first encoded by a 64-dimensional linear layer, and edge attributes are encoded by a 32-dimensional linear layer. The stage number is mapped to a 64-dimensional stage vector and added to the node encoding. The graph network transmits messages along the directed edges of S2 in each stage window. Edges with a confidence weight lower than 0.4 are excluded by a mask. The message aggregation results are then input into a gated recurrent unit with a hidden dimension of 128 in chronological order. Image message transmission implement, Represents a node In the stage window The 64-dimensional hidden vector after graph message passing is complete. This represents the 64-dimensional node encoding before message passing. Indicates upstream node In the stage window Sixty-four-dimensional node encoding is performed before message passing at this layer. This represents the parameter matrix of a linear transformation from 64-dimensional to 64-dimensional nodes. Represents the nodes pointed to in the dynamic response graph And through the upstream node set of the confidence mask, This indicates that the dimensionless attention weights obtained by Softmax from the upstream and downstream node encodings and edge encodings, and the weights within the same non-empty set of incoming edges sum to one; when If the value is empty, Softmax is not executed; the neighbor summation term is determined as a 64-dimensional zero vector and then... Only retain the transformation of its own node, and write the mask with no valid incoming edges; This represents the neighbor message linear transformation parameter matrix from a 96-dimensional concatenated vector to a 64-dimensional message vector. Indicates from node To the node The 32-dimensional normalized edge attribute vector, This represents a 96-dimensional concatenated vector formed by combining 64-dimensional node encoding and 32-dimensional edge encoding. This represents vector concatenation. Represents the ReLU activation function. This indicates the upstream node index. The aforementioned standardized encoding vector and parameter matrix are all internal numerical values ​​of the model, and the units are not applicable. The model's final stage includes an anomaly classification head, a location output head, and a state prediction head. The anomaly classification head outputs six probabilities for each valid node. The state prediction head outputs the temperature, relative humidity, and carbon dioxide concentration for the next six windows for each valid node. The location output head selects only the node with the highest probability from the node-by-node anomaly category probabilities as anomaly location candidates. The joint training objective function is set according to... calculate, Indicates the dimensionless joint training loss. This represents the dimensionless cross-entropy classification loss with weights equal to the inverse of class frequency. The dimensionless cross-entropy loss represents the location of anomalies. This represents the dimensionless Huber prediction loss, obtained by standardizing the three environmental quantities according to their respective training set scales. , and This represents the dimensionless weights of the three loss terms, and the grid search on the validation set yields weights of 0.5, 0.2, and 0.3. Specifically refers to the loss weight of abnormal locations and the fusion weight with the peak value of S2. Distinguishing between samples: Only samples with anomaly classification labels of "abnormal" and unique anomaly location identifiers provided by human evidence participate in the anomaly location loss calculation. Normal category samples, samples with insufficient evidence, or samples with multiple undisambiguated anomaly locations have their anomaly location loss mask set to zero, so that the anomaly location loss contribution of the sample is recorded as zero and is not included in the batch average denominator of the loss term. However, the anomaly classification loss and the state prediction loss when there are effective future supervision quantities are still calculated according to their respective masks. When a batch has no anomaly location samples that meet the above conditions, the anomaly location loss term of the batch is directly determined to be zero and its back gradient is skipped. Training employs an Adam optimizer with a batch size of 16, an initial learning rate of 10^-3, and a maximum of 200 epochs. Training stops and the optimal model is retained when the joint loss on the validation set does not decrease for 15 consecutive epochs. The macro-average classification performance value is fixed as the arithmetic mean of the harmonic performance values ​​of the six outlier categories. The harmonic performance value for each category is the harmonic mean of its precision and recall. When the denominator of the harmonic mean is zero, the harmonic performance value for that category is recorded as zero. On the test set, the model must satisfy the following conditions: the macro-average classification performance value for the six categories is not less than 0.85, and the mean absolute errors of the three environmental quantities after denormalization are not higher than 1.5 degrees Celsius, 5% relative humidity, and 300 milligrams per cubic meter, respectively. The arithmetic mean of the mean absolute errors of the three categories is obtained by dividing them by the corresponding quality upper limits and taking the result. The indicator name, calculation order, category order, normalization scale, and rule version are all recorded in the model evaluation record. Model versions that do not meet the conditions are not included in the online evaluation. The model library is used to generate training and verification tasks. Before each online inference cycle begins, the model management module checks whether the online model exists and has passed the aforementioned test thresholds. It also verifies the model file hash, input feature architecture version, dynamic graph edge attribute version, category order, and standardized configuration version. If the current version is unavailable, fails to load, or is incompatible with the input features, the verification is rolled back and re-executed only if the previously verified version actually exists, the model file is complete, and it is compatible with the current input feature architecture. If there is no qualified online version in the model library, no rollback version exists for the first online deployment, or the rollback version still fails to load or is incompatible, model inference and automatic correction for this cycle are prohibited. A model unavailability alarm and manual verification request are generated. The alarm record saves the unavailability reason, S3 input feature version, dynamic graph version, candidate version, current version, and rollback version. The S3 features for this cycle are retained for traceability. No fictitious abnormal probabilities, abnormal locations, or state predictions are generated. The online inference module reads the stage synchronization node features and their dynamic graph versions written by S3 every ten minutes. First, it applies a double hard mask based on the automatic diagnosis qualification flag and the event contamination hard mask. Uncompensated nodes, nodes with false automatic diagnosis qualification flags, and nodes with event contamination hard masks within the twelve stage windows of the current sequence do not participate in graph message transmission, effective node ratio calculation, abnormal location selection, and output, and cannot submit automatic diagnosis records to S5. The missing windows of other nodes block message contributions through masking instead of using zero values ​​as the true measurement. If the number of effective nodes after hard masking is less than 80% of the total number of nodes configured in the graph, an abnormal data relationship alarm and a manual review request are output, and no automatic diagnosis result is generated. When the ratio requirement is met, the model outputs the abnormal classification probability vector and the state prediction sequence for the next sixty minutes for each effective location. The location corresponding to the maximum abnormal category probability among the effective nodes is taken as the abnormal location, and the abnormal category with the highest probability at that location is taken as the abnormal type. The maximum probability is used as the abnormal confidence level. If the probabilities of all abnormal categories are lower than the alarm thresholds for the current species and stage, the normal state is output, but the state prediction sequence is still saved. The status trend is fitted with a least-squares straight line based on six predicted points in the future. The slopes of temperature, humidity and carbon dioxide concentration are saved every ten minutes. At the same time, it is recorded whether the predicted sequence will cross the allowable interval boundary recorded by the S5 process configuration within the prediction window. When the absolute value of the slope is less than or equal to the 75th percentile of the normal historical slope of the same species at the same stage, it is marked as stable. When the slope is positive and the absolute value is greater than the quantile, it is marked as rising. When the slope is negative and the absolute value is greater than the quantile, it is marked as falling. The priority of the interval boundary crossing mark is higher than the slope direction. Before going live, the anomaly classification probability is calibrated by temperature scaling on the validation set. A positive temperature coefficient is selected based on the minimum negative log-likelihood of the validation set and fixed in the model configuration. During online inference, the scores of each unnormalized class in the model are first divided by the positive temperature coefficient, and then Softmax is performed in the same class order as training and validation to obtain the calibration probability. The anomaly confidence and class interval are calculated based on the calibrated abnormal class probabilities. When the difference between the calibrated maximum abnormal class probability and the second largest probability is less than 0.1, the diagnostic record is marked as a class conflict and two candidate anomaly types are retained. S5 only outputs an alarm for class conflict records without automatic correction. After denormalization, the output of the state prediction head retains the unpruned original prediction values ​​and performs physical range checks point by point according to temperature -10 to 60 degrees Celsius, relative humidity 0 to 100%, and carbon dioxide concentration 0 to 20,000 mg / m³. If any original prediction value exceeds the corresponding physical range, the inference is deemed to have failed. The out-of-range original value, the corresponding range, and the reason for failure are saved. The most recent valid state prediction sequence is restored and a prediction failure flag is written. The restored sequence is only used for trend display and does not pass the automatic correction qualification check of S5. Only when all original prediction values ​​are within the corresponding physical range are they written into the current valid state prediction sequence without pruning. The model management module sorts candidate model versions by the average classification performance of the six macro categories in the test set, the dimensionless normalized prediction error, and the inference time. It first filters out versions that do not meet the quality lower limit or whose single batch inference time exceeds ten seconds. Then, it selects the online version by descending order of the average classification performance of the six macro categories, ascending order of the dimensionless normalized prediction error, and ascending order of the inference time. Training, testing, and model management all read the same versioning index rules. When two versions have the same index, the version that updated the training data cutoff time is selected. When three consecutive batches of online models show inconsistent action effects, the current version is stopped and the previous verified version is switched to. Before model execution, the online inference entry checks the proportion of effective nodes, the consistency between the dynamic graph version and node feature records, and the number of consecutive missing windows for key channels in each location. If the proportion of effective nodes is less than 80%, the versions are inconsistent, or any key channel is missing for more than three consecutive windows, the processing for this cycle is blocked before model inference. The original probability vector, full node prediction tensor, abnormal location candidates, abnormal confidence or state trend for this cycle are not generated, nor is the old model output restored as the result for this cycle. The online inference module only generates quality failure records, saving the model version, input feature version, dynamic graph version, affected node set, missing channel mask, failure conditions and quality flags. Each model output field is explicitly written as empty and marked as not generated. The anomaly type is set to collection anomaly or data relationship anomaly according to the failure conditions. The quality failure records are handed over to S5 only for identifying the first priority blocking, output alarm and manual review request and ending the cycle. Candidate action screening, stage-related constraint screening or diagnostic pulses are not performed.

[0024] In this specific embodiment, S5 includes: This specific embodiment of an artificial intelligence monitoring method for edible fungi substrate fermentation process involves a correction decision module reading the anomaly diagnosis record of S4. Using the strain identifier and current fermentation stage as a joint key, it retrieves from the process configuration library the allowable temperature range, allowable relative humidity range, allowable carbon dioxide concentration range, intolerable hard safety boundary, anomaly threshold, automatic correction threshold, actuator rated amplitude, minimum action interval, maximum recovery time, routine action duration parameters, weight attenuation coefficient, and phenotypic verification window. The method also includes the process configuration record storage unit, effective range, sample batch range, configuration version, and effective time. The stages, anomaly types, and treatment types are used as joint keys to store phenotypic validation targets. The phenotypic validation targets specify the expected increase, decrease, or maintenance direction, minimum effective change, allowable tolerance, and validation inadequacy conditions for mycelial coverage, mycelial expansion speed, and texture continuity, respectively. In this specific implementation, a record of the mycelial expansion stage of Pleurotus ostreatus is set with the allowable temperature range of 22 to 26 degrees Celsius, the allowable relative humidity range of 65% to 75%, the allowable carbon dioxide concentration range of 1,200 to 3,000 mg / m³, the minimum action interval of 20 minutes, and the phenotypic validation window of 12 hours. The automatic correction threshold is determined by the empirical distribution of the abnormal classification probability of the labeled normal batch and abnormal batch on the validation set. First, the intersection of the two empirical frequency curves is calculated at intervals of 0.01 to 1. If there is only one intersection, the probability value is directly saved. If there are multiple intersections, the sum of the number of misjudgments and the number of missed judgments generated by each intersection as the threshold is calculated and sorted in ascending order of total number and descending order of threshold. Then, the first intersection is taken. If there is no intersection, all abnormal classification probability values ​​in the validation set are traversed and sorted in ascending order of the sum of the number of misjudgments and the number of missed judgments and descending order of threshold. Then, the first probability value is taken. The obtained threshold, along with the validation set version and calculation date, is written into the process configuration record. The error correction decision module generates candidate actions from the actuator mapping records according to the anomaly type. For example, insufficient ventilation corresponds to increasing the upstream fan speed or damper opening in the suspected anomaly area; moisture content deviation corresponds to increasing the zonal humidification duty cycle when humidity prediction decreases and decreasing it when prediction increases; localized overheating corresponds to increasing ventilation or reducing the influence of nearby heat sources; mycelial expansion stagnation only generates environmental correction actions when the predicted environmental quantity deviates from the allowable range; data collection anomalies do not generate automatic environmental actions and directly enter the maintenance or manual review branch; each candidate action records the actuator identifier, location identifier, and action details. Direction, rated amplitude, most recent action time, set of affected nodes, and action duration. The action duration is read from the process configuration or safety action calibration record according to the bacterial species, stage, anomaly type, actuator type, and candidate amplitude range. The shortest effective duration, nominal duration, and maximum allowable duration are read. The timed action takes the nominal duration that is limited between the shortest effective duration and the maximum allowable duration. The continuous set value action is maintained until the next monitoring cycle is recalculated, all affected nodes return to the allowable range, or the interlock is triggered. The action termination time is generated based on the first of these three conditions. The action is terminated unconditionally when the interlock is triggered. The magnitude of the candidate action is determined based on the deviation of the state prediction sequence from the nearest boundary of the allowable interval, according to... calculate, Indicates environmental quantity The amplitude of the corresponding candidate action and the unit are consistent with the actuator control quantity. This represents a positive amplitude gain, greater than zero, obtained by linear regression of the environmental changes and action amplitudes of the most recent thirty safety correction actions. The unit is the actuator control quantity divided by the environmental quantity. Indicates the environmental quantity within the prediction window The non-negative maximum unit deviation relative to the nearest boundary of the allowed interval along the direction of the matched candidate action and Greater than or equal to zero This indicates the rated amplitude of the corresponding actuator. This means restricting the result to between a given lower and upper limit. The system includes a dimensionless index of environmental quantities such as temperature, relative humidity, or carbon dioxide concentration; control gain records are stored hierarchically by species, stage, location group, and actuator type, and contain a multivariate time-sharing response template obtained from safety action response calibration. The template stores the action response delay, unit action gain, recovery time, number of valid samples, and template version for each affected node under different combinations of action amplitude and action duration for temperature, relative humidity, and carbon dioxide concentration in the next six ten-minute prediction steps; the maximum recovery time limit in the process configuration record must not exceed the common coverage time domain of the S4 state prediction sequence and the matched multivariate time-sharing response template. In this specific embodiment, the common coverage time domain is sixty minutes. If the maximum recovery time limit exceeds sixty minutes, the coverage time domain of the two is insufficient, or the versions are inconsistent, automatic correction is disabled and manual review is required. Similarly, no automatic action is generated if no valid amplitude gain, action duration parameter, or multivariate time-sharing response template is matched. When screening candidate actions, the correction decision module reads the S4 full-node prediction tensor according to the set of affected nodes, and superimposes the multivariate time-sharing response template corresponding to the actuator, action direction, action amplitude, action duration, and node in the control gain record step by step. Before the response delay, the action quantity is not superimposed. After the delay, the predicted values ​​of temperature, relative humidity, and carbon dioxide are updated together with the unit action gain of each prediction step, the candidate amplitude, and the effective action duration. After the recovery time is exceeded, the action response is attenuated using the recovery coefficient specified in the template. For environmental quantities that have exceeded the allowable range but have not yet exceeded the hard safety boundary before the action takes effect, the existing deviation is allowed to be retained, but the predicted deviation must not be expanded step by step. For environmental quantities that still exceed the boundary after the action takes effect, all affected nodes are required to reduce the most unfavorable deviation along the candidate correction direction step by step, and return to the allowable range within the maximum recovery time limit jointly given by the process configuration and the response template. For environmental quantities that are in the allowable range before the action, the candidate action must not cause new boundary violations. If any current value or predicted value exceeds the hard safety boundary, the candidate is directly rejected and an alarm, interlock, or... Manual review; for each affected node, each prediction step, and each type of environmental quantity, the correction decision module divides the distance below the allowable lower bound or above the allowable upper bound by the corresponding allowable interval width to obtain the dimensionless out-of-bounds amount. Out-of-bounds amounts within the interval are recorded as zero. Then, the reduction amount is obtained by subtracting the out-of-bounds amount after superposition from the out-of-bounds amount before the candidate action is superimposed. The arithmetic mean of the reduction amounts for all affected nodes, six prediction steps, and three types of environmental quantities is taken as the dimensionless out-of-bounds improvement score for the candidate action. The constraint verification results save the existing out-of-bounds amount and the gradual improvement amount before the action takes effect. The criteria for selection include: dimensionless outbound improvement score, normalization scale, aggregation method and rule version, maximum recovery time limit, most unfavorable prediction point, boundary distance, hard safety boundary, template version and rejection reason. The candidate amplitude does not exceed the rated amplitude of the actuator. The current time is at least the minimum action interval from the last action of the same actuator. The confidence weight of the inbound edge of the dynamic graph of the affected node is not less than 0.4. When multiple candidates meet the criteria, the one with the highest dimensionless outbound improvement score is selected first, followed by the one with the lower action amplitude value. If they are still the same, they are selected in ascending order according to the actuator identifier. The error correction decision module processes records within the same period with a single priority: Category conflicts, model input quality failures, model unavailability, mismatched process configurations, or hard safety boundary violations are considered first priority, immediately disabling automatic error correction and diagnostic pulses and outputting alarm information and a manual review request; if no first priority condition exists and there are no candidate actions that meet the error correction constraints, alarm information and a manual review request are directly output; if the aforementioned blocking conditions do not exist and the anomaly confidence level is lower than the automatic error correction threshold, and the phase consistency residual exceeds the current bacterial species and stage anomaly thresholds, only the alarm is retained and the handling status is set to diagnostic pulse pending judgment, temporarily suspending manual review. Upon receiving a nuclear request, the process proceeds to the diagnostic pulse safety assessment described later. Other low-confidence records directly output a manual review request. If there is no aforementioned blockage, a qualified candidate exists, and the abnormal confidence level reaches the automatic correction threshold, a zone ventilation, humidification, or cooling suggestion is output. The alarm or suggestion record saves the abnormal location, abnormal type, abnormal confidence level, category conflict flag, phase consistency residual, status trend, rejected candidates, and rejection reasons. The suggestion record also saves the target location, actuator, action direction, action amplitude, action duration, constraint verification results, and configuration version, and is executed by a controller with interlocking function. Only one handling status is allowed to take effect in the same cycle. For suspected anomalies where the phase consistency residual exceeds the current bacterial species and stage anomaly threshold but the anomaly confidence level is lower than the automatic correction threshold, the correction decision module retains the alarm and calculates the diagnostic pulse safety margin. For each current measurement value of temperature, humidity, and carbon dioxide concentration, and each S4 predicted value during the diagnostic action, the minimum of the distance obtained by subtracting the lower allowable limit from the current value and the distance obtained by subtracting the current value from the upper allowable limit is taken as the signed margin. The margin is positive when it is within the allowable range, zero when it is equal to the boundary, and negative when it crosses the boundary. For each type of environmental quantity, the minimum of its current measurement value and the signed margin of all predicted points is taken. When the minimum signed margin before the action of the three types of environmental quantities is greater than zero, the trial allowable increment of the fan speed or valve opening is first converted with the control gain, and the trial diagnostic action amplitude is taken as the allowable increment and 10% of the actuator rated amplitude. The minimum of the three values ​​(maximum diagnostic amplitude, device calibration amplitude, and action duration) is used. The action duration is the calibration duration that can form three consecutive sampling points and does not exceed five minutes. Then, a multivariate time-sharing response template that matches the target actuator, action direction, trial amplitude, action duration, and template version is reused. According to the response delay of each affected node, the influence of the pulse on temperature, relative humidity, and carbon dioxide concentration is superimposed step by step on the original S4 prediction. The signed margin of all affected nodes in all prediction steps is recalculated. If there is no valid template, the template is inconsistent with the S4 prediction version, any signed margin before the action is not greater than zero, or the signed margin of any environmental quantity of any node after the pulse is superimposed is not greater than zero, the diagnostic action amplitude is set to zero. The current value, original prediction point, prediction point after the action, allowable boundary, template version, and blocking reason are recorded, and a manual review request is directly output. The controller applies a limited diagnostic ventilation pulse only when the signed margins of the current measured value, the original predicted environmental quantity of S4, and the predicted environmental quantity after the action following the superimposed diagnostic pulse are all greater than zero during the diagnostic action, the diagnostic action amplitude is greater than zero, there is no violation of the hard safety boundary of temperature, humidity, or carbon dioxide, the minimum action interval since the last action has been reached, and the suspected abnormal location has an airflow reachable relationship with the target fan or valve. Otherwise, the manual review request is executed directly, and the pulse start time is registered as a new common drive event. Based on this, S1 detects the response events of air volume, carbon dioxide change rate, humidity change rate, and temperature change rate after the action. S2 extracts the arrival time, peak value, integral response, and recovery time, and determines the effective attenuation range according to the amplitude ratio range of 5% to 95% of the unblocked historical pulse samples. For node pairs that form diagnostic pulses, if the upstream node responds first and the downstream node responds later, and the response amplitude decays into the effective decay range, then the corresponding directed edge is established or updated. If the number of consecutive missing responses reaches the three-weighting threshold in the device calibration record, then the edge confidence weight is multiplied by the pulse weighting coefficient of 0.7. If the number of consecutive missing responses reaches the five-edge deletion threshold, then the edge is deleted. Both weighting and deletion record the pulse event identifier, the original edge weight, the updated weight, or the deletion time, thereby writing the diagnostic action evidence into the new dynamic response graph version. The anomaly verification module uses the updated dynamic response map to perform mutual exclusion judgment in the order of sensor anomaly, local ventilation blockage and actual fermentation anomaly. When the adjacent nodes of the suspected anomaly area meet the directed edge establishment condition, the number of times a single acquisition channel in the area fails to reach the response threshold three times in a row, and the other acquisition channels form response events, it is determined to be a sensor anomaly and sensor maintenance information is output. When the air volume channel and at least one of the carbon dioxide, humidity or temperature channels in the area fail to reach the response threshold three times in a row, and the downstream node does not form a corresponding response event, it is determined to be a local ventilation blockage and ventilation path inspection information is output. If neither of the aforementioned two fault conditions is met, and the environmental response of the suspected abnormal location reaches a threshold, the response sequence conforms to the directed edge direction, the response amplitude decay is within the effective decay range, and at least one component of the metabolic or phenotypic characterization exceeds the 5% to 95% quantile range of the current strain and the normal historical samples of the stage for three consecutive stages, then it is determined to be a true fermentation abnormality, and the partition correction that passes the constraint screening is selected. For suspected abnormalities that still do not meet any of the judgment conditions, a manual review request is output, and the review results and the version of the dynamic graph used for judgment are written into the handling record; the diagnostic pulse status is updated according to the unidirectional state machine: safety conditions not met, template or evidence missing, pulse execution. If a unique determination cannot be reached through failure or abnormal review, the diagnostic pulse pending determination is switched to pending manual review, and a manual review request is activated at that point. When a unique determination of sensor abnormality, local ventilation blockage, or actual fermentation abnormality is reached, the diagnostic pulse pending determination state is replaced by sensor maintenance, ventilation path repair, or actual abnormality correction state, respectively, and the manual review branch that has not yet been activated is closed. During the execution of the diagnostic pulse and the waiting for evidence, the alarm is retained and no other automatic actions or manual review tasks occur. The subsequent determined state overwrites the previous temporary state but does not delete the alarm traceability record. Only one currently effective handling state is always retained for the same batch, location, and monitoring cycle. After zonal correction, diagnostic pulse, ventilation path inspection, or sensor maintenance is completed, the server continues to acquire hyphal coverage, hyphal expansion speed, and texture continuity within a twelve-hour phenotypic verification window. Using the median of the last three effective phenotypic windows before the action as the baseline, and comparing the actual changes with the expected direction, minimum effective changes, and allowable tolerances item by item according to the phenotypic verification objectives in this configuration version, which are based on species, stage, anomaly type, and treatment type. The three types of results are mutually exclusive, with data quality considered first, followed by the direction of effect. Two consecutive images showing low confidence, and three... When there are fewer than two valid items in the phenotypic project or the change of any item to be judged is within the minimum valid change and the direction cannot be determined, the result is first recorded as insufficient verification and no longer judged as consistent or inconsistent; after the data quality threshold is reached, when all valid items meet the expected direction, minimum valid change and allowable tolerance, the result is recorded as consistent and the edge attribute after S2 update is retained; when any valid item exceeds the allowable tolerance or the direction is opposite, the result is recorded as inconsistent, and the actual value, target value, tolerance, data quality threshold and mutual exclusion judgment result of each item are written into the phenotypic verification record; The error correction decision module updates the suggested status according to the aforementioned unique priority: when there is a first-priority block or no qualified candidate, it switches to pending manual review; when there is no block, the anomaly confidence level reaches the threshold, and the candidate action passes all constraints, it switches to pending execution; when the anomaly confidence level does not reach the threshold and the phase consistency residual exceeds the corresponding anomaly threshold, it first switches to diagnostic pulse pending judgment, otherwise it directly switches to pending manual review; after the diagnostic pulse safety judgment or anomaly review is completed, it only switches to one of the following according to the aforementioned one-way state machine: pending manual review, sensor maintenance, ventilation path inspection, or real anomaly correction; when the actuator interlock feedback rejects the action, it switches to execution failure and stops subsequent automatic actions; The processing record matches the data before and after the action according to the batch identifier, location identifier, and public driving event identifier. When the environmental response exists after the action but the phenotypic verification window has not yet ended, the record is kept as pending verification. When the phenotypic verification window ends and the phenotypic results are consistent, the record is switched to verification passed. When the phenotypic results are inconsistent, the record is switched to verification failed and the model review is triggered. When the phenotypic result is insufficient verification, the record is switched to pending supplementary sampling and a 12-hour supplementary sampling deadline is set from the end of the original verification window. After the supplementary sampling data arrives, it is re-evaluated according to the aforementioned mutual exclusion rules and switched to verification passed or verification failed. If the insufficient verification is not resolved after the supplementary sampling deadline, the record is switched to manual review and the edge weights, control gains, or model recovery statistics must not be updated. When multiple alarms are received simultaneously for the same suspected abnormal area, the abnormality review module sorts them according to the predetermined priority of sensor abnormality, local ventilation blockage and actual fermentation abnormality, selects the highest priority handling type as the output of this cycle and suppresses the other automatic actions. Only when the highest priority handling is completed and there are still lower-level conditions, will the corresponding handling record be regenerated in the next monitoring cycle. When manual verification confirms that the automatic diagnosis error is correct, the model management module writes the current twelve stage windows, dynamic graph version and manual label of the location into the retraining queue. When manual verification confirms that the process configuration is not applicable, the configuration version is stopped and switched to the most recently verified version of the same strain and stage. When there is no recoverable configuration, a configuration maintenance alarm is output and the actuator is kept in its original state. The diagnostic pulse results are compared with the unobstructed historical intervals item by item according to the arrival time, amplitude attenuation, and recovery time. The states within and outside the intervals or lacking evidence are organized into the input evidence for the aforementioned mutually exclusive judgment, and the treatment type is not selected independently in this section. The anomaly review module still outputs results according to four categories of conditions and predetermined priorities: sensor anomaly, local ventilation blockage, real fermentation anomaly, and manual review. When only one acquisition channel continuously fails to meet the response threshold while the other channels are valid, it satisfies the sensor anomaly evidence. When the air volume channel and at least one other environmental channel are continuously missing and there is no corresponding response downstream, it satisfies the local ventilation blockage evidence. When the response sequence or amplitude attenuation exceeds the historical interval but the evidence is insufficient, it is transferred to manual review and is not directly equivalent to ventilation path blockage. Model verification markers are aggregated according to model version and anomaly type. When the same model version generates at least three inconsistent action effect markers in three consecutive batches, the automatic correction qualification of the candidate version is stopped and it is switched to the previous verified model. If there is no previous verified model, the manual verification mode is maintained and a model maintenance alarm is generated. The stopped candidate version only performs shadow inference, and its correction suggestions are not automatically issued. Only after manual approval or after entering controlled gray-scale interlock execution under preset amplitude, duration and hard safety boundary restrictions, will an action effect verification record with candidate version identifier be generated. The recovery statistics only include the records generated by the candidate version that have qualified data quality flags and have not triggered hard safety boundaries. The action effects generated by the current rollback model are not included. The automatic correction qualification of the candidate version is restored only when the action effect verification results of ten consecutive batches of the candidate version are consistent. When the candidate version is restored, the candidate model version, effective batch, gray-scale execution record and verification statistics are written into the model version record. When the action effect verification results are inconsistent, the edge confidence weight of each response edge on the current action propagation path is multiplied by the weight attenuation coefficient of 0.8 in the process configuration record, and a model verification mark containing batch identifier, model version, dynamic graph version, anomaly diagnosis record, action record and phenotypic difference is generated. If the verification is insufficient, the edge weight is not changed and manual supplementation is performed. If the verification is consistent, the current action and its response are added to the control gain update sample. Finally, the correction suggestion or alarm, the actual executed action, the updated dynamic graph version and the verification effect results are associated and saved to form a closed-loop record for continued use in the next monitoring cycle.

[0025] The above description is only a preferred embodiment 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.

[0026] This invention uses dynamic response diagrams to record the measured propagation delay, amplitude attenuation, and direction of actuator actions across different substrate locations. It also performs time compensation and joint analysis of environmental, metabolic, and phenotypic data under fermentation stage conditions, enabling anomaly detection to be based on the same stage clock and actual propagation path. Therefore, asynchronous responses between locations are no longer directly interpreted as fermentation deviations; abnormal locations and their causes can be distinguished based on node characteristics and edge attributes.

[0027] When the confidence level of the anomaly category is insufficient, this invention uses the safety margin corresponding to the strain and fermentation stage to limit the diagnostic ventilation pulse, and updates the directed edges and edge confidence weights from the pulse response to distinguish between true fermentation anomalies, local ventilation blockages, and sensor anomalies. Subsequent changes in mycelial phenotype are used to verify the effectiveness of corrective actions, repairs, or maintenance, enabling the monitoring model and dynamic response diagram to be continuously corrected as the equipment status and fermentation process change.

Claims

1. An artificial intelligence monitoring method for the fermentation process of edible fungi substrate, characterized in that, include: S1. Acquire environmental data, internal temperature data, and mycelial image sequences that are registered with the substrate within the fermentation space of the edible fungus substrate. Determine the control timestamp of the actuator as a common driving event, and detect the response events corresponding to the common driving events from the environmental data and internal temperature data of the substrate to form a multi-source response sequence. S2. Based on the location-based multi-source response sequence, determine the response delay, response amplitude attenuation, and response direction from the common driving event to each response event. Construct a dynamic response graph with matrix location as the node and response delay, response amplitude attenuation, and response direction as the edge attributes. S3. Determine the fermentation stage based on the locational multi-source response sequence and mycelial image sequence. Perform time compensation on the data of each node according to the response delay of the dynamic response map. Generate environmental driving characterization, metabolic characterization and phenotypic characterization on the common stage clock. Calculate the phase consistency residual between environment, metabolism and phenotypic on the same node and adjacent response paths to form stage synchronization node characteristics. S4. Input the phase synchronization node features and edge attributes into the phase condition graph time series model to obtain the anomaly diagnosis results and state trends, including anomaly location, anomaly type and anomaly confidence. S5. Based on the abnormal diagnosis results and status trends, and combined with the phase-related corrective constraints, generate partition correction suggestions or alarm information. When the suggestion is executed, obtain the response events and mycelial phenotypic changes after the action, and update the dynamic response map accordingly.

2. The artificial intelligence monitoring method for edible fungi substrate fermentation process according to claim 1, characterized in that, S1 includes: Set location markers according to the spatial location of the substrate layer, substrate pile or culture unit, and associate temperature acquisition point, humidity acquisition point, air volume acquisition point or air valve opening acquisition point, carbon dioxide concentration acquisition point, substrate internal temperature acquisition point and mycelium image acquisition area with the location markers, and associate the spatial location of the fan or air valve and its airflow outlet with the drive source location markers. Record the start time, end time, and amplitude of the fan start / stop, valve opening change, and humidification action, and associate the action with the corresponding drive source location identifier. Within the response detection window after the start of each action, the response threshold is determined based on the baseline before the action and the noise statistics. The response event is determined when the air volume, carbon dioxide change rate, humidity change rate or temperature change rate exceeds the corresponding response threshold and continues to reach the preset number of samplings.

3. The artificial intelligence monitoring method for edible fungi substrate fermentation process according to claim 1, characterized in that, S2 include: The difference between the start time of the action of the common driving event and the first time the response event crosses the response threshold is used as the response delay; According to the airflow reachability relationship and the order of response events, starting from the drive source location associated with the common drive event, the location that has responded before each response location and is reachable by airflow to that response location is determined as the upstream reference node. The response amplitude attenuation is determined based on the ratio of the peak or integral response of the response location within the response detection window to the upstream reference node under the same common drive event. The response direction is determined by the direction from the upstream reference node to the response location. The response delay and response amplitude decay are updated according to the time sliding window. The edge confidence weight is determined based on the ratio of the number of times a corresponding node pair forms a valid response event to the total number of common driving events within the time sliding window. The update time and the edge confidence weight are written into the corresponding edge attribute.

4. The artificial intelligence monitoring method for edible fungi substrate fermentation process according to claim 1, characterized in that, S3 include: After inoculation, the fermentation stage is divided according to the turning point when the carbon dioxide release rate changes from rising to stabilizing or falling, and the change in the growth rate of mycelial coverage. The data from each node is shifted to the common stage clock according to the corresponding response delay; The environmental driving characterization is generated by air volume or valve opening, humidity, ambient temperature and carbon dioxide concentration; the metabolic characterization is generated by carbon dioxide release rate and substrate temperature rise rate; and the phenotypic characterization is generated by hyphal coverage, hyphal expansion rate and texture continuity. The phase consistency residual is calculated based on the differences in the direction of change, the time difference of the turning point, and the amplitude of the standardized change within the same stage. The residual of the node itself and the residual of the adjacent nodes aggregated along the directed response path are written together into the stage synchronization node feature.

5. The artificial intelligence monitoring method for the fermentation process of edible fungi substrate according to claim 1, characterized in that, S4 includes: The fermentation history batches are divided into training samples according to strain and fermentation stage. The environmental driving characterization, metabolic characterization, phenotypic characterization and phase consistency residual of the nodes are used as node inputs, and the response delay, response amplitude decay, response direction and edge confidence weight are used as edge inputs. The graph message passing results are time series encoded and the anomaly classification probability and state prediction sequence are output. The anomaly type and anomaly confidence level are determined based on the anomaly classification probability. The anomaly type includes at least insufficient ventilation, water content deviation, local overheating, stagnation of mycelial expansion, and collection anomaly. The state prediction sequence includes at least the predicted temperature, humidity, and carbon dioxide concentration values ​​of each node within a predetermined time window, and the state trend is determined based on the state prediction sequence.

6. The artificial intelligence monitoring method for edible fungi substrate fermentation process according to claim 1, characterized in that, S5 include: The allowable temperature range, allowable humidity range, allowable carbon dioxide concentration range, and automatic correction threshold are read from the process configuration record associated with the current strain and fermentation stage. The automatic correction threshold is taken when there is an intersection between the abnormal classification probability distributions of the labeled normal batch and abnormal batch; when there are multiple intersections, the first intersection is taken after sorting in ascending order by the sum of the number of false positives and the number of false negatives in the validation set; when there is no intersection, the first abnormal classification probability value is taken after sorting in ascending order by the sum of the number of false positives and the number of false negatives in the validation set. Candidate actuators and candidate action directions are matched according to the anomaly type. The amplitude of candidate actions is determined according to the state trend. Candidate actions are screened out based on the correction constraints of not exceeding the allowable range of the corresponding environment, the rated amplitude of the actuator, and the minimum action interval. When there are no candidate actions that satisfy the correction constraints or the abnormal confidence level does not reach the automatic correction threshold, an alarm message and a manual review request are output; otherwise, suggestions for zonal ventilation, humidification, or cooling are output.

7. The artificial intelligence monitoring method for edible fungi substrate fermentation process according to claim 6, characterized in that, After implementing the suggested zonal ventilation, humidification, or cooling, the execution time is registered as a new common driving event. The response event after the action is detected in step S1, and the corresponding edge attribute is updated in step S2. At the end of the preset phenotypic verification window, the verification results of the action effect are determined based on the changes in hyphal coverage, hyphal expansion speed, and texture continuity. When the action effect verification result indicates that the phenotypic change direction is consistent with the suggested target, the updated edge attributes are retained. When the two are inconsistent, the edge confidence weight of the corresponding response edge is multiplied by the preset weight decay coefficient in the process configuration record and a model verification mark is generated.

8. The artificial intelligence monitoring method for the fermentation process of edible fungi substrate according to claim 6, characterized in that, The graph message transmission results are time-series encoded and anomaly classification probability and state prediction sequence are output. When the phase consistency residual exceeds the anomaly threshold associated with the current strain and fermentation stage and the anomaly confidence does not reach the automatic correction threshold, the alarm information is retained, and the safety margin of the current temperature, humidity and carbon dioxide concentration relative to the boundary of their respective environmental allowable ranges is read from the process configuration record before manual review. The diagnostic action amplitude and duration are determined based on the safety margin and the rated amplitude of the actuator. The predicted values ​​of temperature, humidity and carbon dioxide concentration in the state prediction sequence are determined as the predicted environmental quantities during the diagnostic action. When the predicted environmental quantity does not exceed the boundary of the corresponding environmental allowable range and the minimum action interval is reached from the last action, a limited diagnostic ventilation pulse is applied to the fan or valve that has an airflow reachability relationship with the suspected abnormal location, and a correction or manual review is determined based on the pulse response. If any of the aforementioned conditions are not met, the manual review request is executed directly.

9. The artificial intelligence monitoring method for edible fungi substrate fermentation process according to claim 8, characterized in that, Using the start time of the limited diagnostic ventilation pulse as the common time zero point, the arrival time, peak value, integral response, and recovery time of the air volume, carbon dioxide change rate, humidity change rate, and temperature change rate of each matrix location are extracted. The effective attenuation interval is determined based on the statistical interval of the response order and response amplitude ratio of upstream and downstream nodes in the unblocked historical pulse sample. Node pairs that satisfy the condition that the upstream node responds first, the downstream node responds later, and the response amplitude attenuation is within the effective attenuation interval are established or updated as directed edges. For directed edges whose consecutive response missing counts reach the weight reduction threshold, their edge confidence weight is multiplied by the pulse weight reduction coefficient in the device calibration record. Directed edges whose consecutive response missing counts reach the deletion threshold are deleted. The weight reduction threshold and the deletion threshold are given by the device calibration record, and the deletion threshold is greater than the weight reduction threshold.

10. The artificial intelligence monitoring method for the fermentation process of edible fungi substrate according to claim 9, characterized in that, Anomaly verification is performed using the updated dynamic response map: when the environmental response of a suspected abnormal location reaches the corresponding response threshold, the response sequence matches the response direction of the corresponding directed edge, and the response amplitude decay is within the effective decay range, and at least one component of the metabolic characterization or phenotypic characterization continuously exceeds the preset quantile range of the historical normal samples in the current fermentation stage for a preset number of samplings, it is determined to be a real fermentation anomaly and a partition correction is selected. When the number of times that the air volume channel and at least one of the channels of carbon dioxide concentration, humidity or temperature in the suspected abnormal area fails to reach the corresponding response threshold reaches the weighting threshold, and the downstream node does not generate a corresponding response event, it is determined to be a local ventilation blockage and ventilation path maintenance information is output. When the adjacent nodes of a suspected abnormal location meet the directed edge establishment conditions, the number of times a single acquisition channel of the suspected abnormal location fails to reach the corresponding response threshold reaches the weighting threshold, and the other acquisition channels form corresponding response events, the sensor is determined to be abnormal and sensor maintenance information is output. The judgment conditions that are met simultaneously are processed in the order of sensor abnormality, local ventilation blockage and actual fermentation abnormality. For suspected abnormalities that do not meet the aforementioned judgment conditions, a manual review request is output. Acquire changes in mycelial coverage, mycelial expansion speed, and texture continuity after zonal correction, ventilation path repair, or sensor maintenance; determine the verification results of the review effect; and verify the state changes after processing.

Images