Closed space air quality regulation method and system based on multi-sensor fusion and AI algorithm

Abstract

The application discloses a closed space air quality regulation method and system based on multi-sensor fusion and AI algorithm, and the method comprises the following steps: collecting space partition relationship, connection relationship and fusion time sequence data; converting the fusion time sequence data into a space-time state diagram according to the space partition relationship and the connection relationship; obtaining source item precursor coding through a process gas time lag causal model; obtaining migration path coding through a pollution migration reasoning model; obtaining candidate action coding and counterfactual migration coding through a counterfactual action evaluation model; performing counterargument discrimination according to a personnel position sequence and a gate access state sequence to obtain target action coding; and converting the target action coding into linkage execution coding and outputting the linkage execution coding. Through process precursor identification, pollution migration reasoning and counterfactual action screening, the application changes the air regulation of the closed space from lag response to early intervention, and reduces the risk of diffusion of pollution gas to the personnel activity area and secondary migration induced by false ventilation.

Description

Technical Field

[0001] This application relates to the field of air quality monitoring and intelligent control technology, and in particular to a method and system for air quality regulation in enclosed spaces based on multi-sensor fusion and AI algorithms. Background Technology

[0002] Wastewater treatment plant areas such as screen rooms, underground pump rooms, sludge dewatering rooms, covered tanks, pipe corridors, and maintenance wells typically suffer from poor ventilation, high humidity, strong corrosiveness, dispersed pollution sources, and complex gas composition, making them prone to producing harmful gases such as hydrogen sulfide, ammonia, methane, and volatile organic compounds. Existing methods mostly rely on fixed-point sensors, threshold alarms, timed ventilation, or single-fan linkage control. These methods usually only take action after pollutant concentrations have already increased, making it difficult to respond promptly to short-term, localized, and sudden air quality anomalies.

[0003] Meanwhile, an easily overlooked problem in existing technologies is that the wastewater treatment process status often changes before the air sensor malfunctions. For example, the start-up and shutdown of booster pumps, sudden changes in liquid level, sludge return, adjustment of aeration volume, or opening of maintenance covers can all induce instantaneous gas release. However, traditional air control systems usually do not use the above process data as a basis for air risk prediction, resulting in a control lag.

[0004] Another easily overlooked issue is that ventilation control actions themselves may alter the airflow path in enclosed spaces, causing polluted gases to migrate from the source area to areas where people are active, inspection passages, or adjacent spaces. If the start and stop of the fans are controlled solely based on the concentration at a single point, it may cause secondary diffusion of pollutants or an imbalance in the load of the deodorization branch. Summary of the Invention

[0005] To address the aforementioned problems, embodiments of the present invention provide a method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms, the method comprising:

[0006] Collect spatial zoning relationships, connectivity relationships, and fused time-series data. The fused time-series data includes gas monitoring sequences, airflow status sequences, personnel location sequences, access control status sequences, equipment status sequences, and wastewater process sequences.

[0007] Based on the spatial partitioning relationship and the connectivity relationship, the fused time series data is converted into a spatiotemporal state diagram, which includes source region nodes, active region nodes, deodorization region nodes and connecting edges;

[0008] The spatiotemporal state diagram is input into the process gas time-delay causal model, and time-delay causal calculations are performed on the wastewater process sequence, the equipment state sequence, and the gas monitoring sequence to obtain the source term precursor code;

[0009] The spatiotemporal state diagram, the source term precursor code, the airflow state sequence, the personnel location sequence, and the access control state sequence are input into the pollution migration inference model to obtain the migration path code;

[0010] The migration path code and the device state sequence are input into the counterfactual action evaluation model to obtain candidate action codes and counterfactual migration codes.

[0011] Based on the personnel location sequence and the access control status sequence, the counterfactual transition code is subjected to a counterfactual verification to obtain the target action code;

[0012] The target action code is converted into a linkage execution code, and the linkage execution code is output.

[0013] Furthermore, the gas monitoring sequence is characterized by comprising a hydrogen sulfide sequence, an ammonia sequence, a methanethiol sequence, an odor concentration sequence, and an oxygen content sequence; the airflow state sequence is characterized by comprising a wind speed sequence, a wind direction sequence, and a pressure difference sequence; the wastewater process sequence is characterized by comprising a booster pump state sequence, a liquid level state sequence, an aeration state sequence, a sludge return state sequence, a sludge dewatering state sequence, a chemical dosing state sequence, and a tank cover state sequence; and the equipment state sequence is characterized by comprising a ventilation equipment state sequence, a deodorization equipment state sequence, and an alarm equipment state sequence.

[0014] Furthermore, the transformation of the spatiotemporal state diagram includes: converting the spatial partitioning relationship into source area nodes, activity area nodes, and deodorization area nodes; converting the connectivity relationship into connectivity edges; encoding the gas monitoring sequence, personnel location sequence, and wastewater process sequence into node state codes; encoding the airflow state sequence, access control state sequence, and equipment state sequence into edge state codes; and writing the node state codes and edge state codes into the spatiotemporal state diagram along the acquisition time sequence.

[0015] Furthermore, the process gas time-delay causal model includes a disturbance segment extraction layer, a response segment alignment layer, a causal edge generation layer, and a precursor coding layer. The disturbance segment extraction layer extracts state transition segments from the wastewater process sequence and the equipment state sequence. The response segment alignment layer extracts gas response segments located after the state transition segments from the gas monitoring sequence. The causal edge generation layer generates causal edge codes based on the sequential relationship between the state transition segments and the gas response segments. The precursor coding layer writes the causal edge codes, the source region nodes, and the connected edges into the source term precursor codes.

[0016] Furthermore, the pollution migration inference model includes a source region location layer, a spatial chain graph attention layer, a flow direction constraint layer, and a path output layer; the source region location layer reads the source region nodes from the source term precursor code; the spatial chain graph attention layer assigns adjacency weights to the source region nodes, the connected edges, the active area nodes, and the deodorization area nodes according to the spatiotemporal state graph; the flow direction constraint layer writes the airflow state sequence and the access control state sequence into the adjacency weights; and the path output layer generates the migration path code based on the written adjacency weights.

[0017] Furthermore, the counterfactual action evaluation model includes an action generation layer, an action injection layer, a counterfactual reconstruction layer, and an action output layer. The action generation layer generates a first action code, a second action code, and a third action code based on the migration path code and the device state sequence. The action injection layer sequentially writes the first action code, the second action code, and the third action code into the spatiotemporal state diagram to obtain a first action diagram, a second action diagram, and a third action diagram. The counterfactual reconstruction layer generates a first counterfactual migration code, a second counterfactual migration code, and a third counterfactual migration code based on the first action diagram, the second action diagram, and the third action diagram. The action output layer writes the first action code, the second action code, and the third action code into the candidate action code, and writes the first counterfactual migration code, the second counterfactual migration code, and the third counterfactual migration code into the counterfactual migration code.

[0018] Further, the counterfactual determination includes: reading the predicted arrival node and the predicted traversed edge from the counterfactual migration code; generating a first counterfactual code when the predicted arrival node is consistent with the activity area node corresponding to the personnel location sequence; generating a second counterfactual code when the predicted traversed edge is consistent with the connected edge corresponding to the access control state sequence; and determining the candidate action code that does not contain the first counterfactual code and does not contain the second counterfactual code as the target action code.

[0019] Furthermore, the linkage execution code consists of action object code, action type code, execution edge code, and data acquisition trigger code; the execution device is determined according to the action object code; the connected edge is determined according to the execution edge code; the execution device changes the edge state of the connected edge according to the action type code; the gas monitoring sequence and the airflow state sequence after execution are collected according to the data acquisition trigger code; the gas monitoring sequence, the airflow state sequence after execution, and the counterfactual migration code are used to form a correction sample; the correction sample is written into the sample cache of the process gas time-delay causal model, the pollution migration inference model, and the counterfactual action evaluation model.

[0020] The closed space air quality control method based on multi-sensor fusion and AI algorithm also includes: making a reliable judgment on the calibration sample based on the execution receipt, access control status sequence and airflow status sequence after execution, writing the reliable calibration sample into the sample cache, and writing the unreliable calibration sample into the isolation sample area.

[0021] An air quality control system for enclosed spaces based on multi-sensor fusion and AI algorithms, comprising:

[0022] The data acquisition module collects spatial partitioning relationships, connectivity relationships, and fused time-series data. The fused time-series data includes gas monitoring sequences, airflow status sequences, personnel location sequences, access control status sequences, equipment status sequences, and wastewater process sequences.

[0023] The graph construction module converts the fused time series data into a spatiotemporal state graph based on the spatial partitioning relationship and the connectivity relationship. The spatiotemporal state graph includes source region nodes, active region nodes, deodorization region nodes, and connecting edges.

[0024] The precursor identification module inputs the spatiotemporal state diagram into the process gas time-delay causal model, performs time-delay causal calculations on the wastewater process sequence, the equipment state sequence, and the gas monitoring sequence, and obtains the source term precursor code.

[0025] The migration inference module inputs the spatiotemporal state diagram, the source term precursor code, the airflow state sequence, the personnel location sequence, and the access control state sequence into the pollution migration inference model to obtain the migration path code;

[0026] The action pre-simulation module inputs the migration path code and the device state sequence into the counterfactual action evaluation model to obtain candidate action codes and counterfactual migration codes;

[0027] The counterfactual filtering module performs counterfactual discrimination on the counterfactual migration code based on the personnel location sequence and the access control status sequence to obtain the target action code.

[0028] The instruction output module converts the target action code into a linkage execution code and outputs the linkage execution code.

[0029] The technical effects and advantages of the closed space air quality control method and system based on multi-sensor fusion and AI algorithm provided by this invention are as follows:

[0030] This invention transforms air quality control in enclosed spaces from a delayed response to proactive intervention through process precursor identification, pollution migration reasoning, and counterfactual action screening. This reduces the risk of polluted gases spreading to areas with high population density and secondary migration induced by incorrect ventilation. By integrating wastewater process sequences, equipment status sequences, and gas monitoring sequences, this invention can identify gas release precursors caused by process disturbances before significant anomalies in gas monitoring values, reducing the lag in traditional threshold alarm methods. By converting spatial zoning relationships, connectivity relationships, and fused time-series data into a spatiotemporal state diagram, it can uniformly incorporate pollution source areas, areas with high population density, deodorization areas, and connecting edges into the calculation, making air quality control no longer limited to single-point sensor judgments. By writing calibration samples into the sample cache, it can use the executed gas monitoring sequences and airflow status sequences to correct subsequent model processing, improving the system's adaptability under complex operating conditions. Attached Figure Description

[0031] Figure 1 This is a flowchart of the enclosed space air quality control method based on multi-sensor fusion and AI algorithm in Example 1;

[0032] Figure 2 This is a flowchart of the trusted resample processing method in Example 2;

[0033] Figure 3 This is a schematic diagram of the connection of the enclosed space air quality control system based on multi-sensor fusion and AI algorithm in Example 3. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0035] Please see Figure 1 As shown, embodiments of the present invention provide a method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms. The method includes:

[0036] S1. Collect spatial partitioning relationships, connectivity relationships, and fused time series data. The fused time series data includes gas monitoring sequences, airflow status sequences, personnel location sequences, access control status sequences, equipment status sequences, and wastewater process sequences.

[0037] S2. Based on the spatial partitioning relationship and the connectivity relationship, the fused time series data is converted into a spatiotemporal state diagram, which includes source region nodes, active region nodes, deodorization region nodes and connecting edges.

[0038] S3. Input the spatiotemporal state diagram into the process gas time-delay causal model, perform time-delay causal calculations on the wastewater process sequence, the equipment state sequence, and the gas monitoring sequence to obtain the source term precursor code;

[0039] S4. Input the spatiotemporal state diagram, the source term precursor code, the airflow state sequence, the personnel location sequence, and the access control state sequence into the pollution migration inference model to obtain the migration path code;

[0040] S5. Input the migration path code and the device state sequence into the counterfactual action evaluation model to obtain the candidate action code and the counterfactual migration code;

[0041] S6. Based on the personnel location sequence and the access control status sequence, perform a counterfactual verification on the counterfactual transition code to obtain the target action code;

[0042] S7. Convert the target action code into a linkage execution code and output the linkage execution code.

[0043] In this embodiment, the fused time series data is formed according to a unified time scale, which is jointly written by the acquisition device, the wastewater treatment control system, and the execution equipment controller. Under each unified time scale, the gas monitoring sequence, airflow status sequence, personnel location sequence, access control status sequence, equipment status sequence, and wastewater process sequence are obtained respectively, so that the same node in the subsequent spatiotemporal state diagram has corresponding data records under the same time scale.

[0044] The gas monitoring sequence is used to represent the changes in pollutant gases and the risk of oxygen deficiency in the enclosed space of a wastewater treatment plant. It consists of hydrogen sulfide, ammonia, methanethiol, odor concentration, and oxygen content sequences. These sequences are not individual instantaneous values, but rather a set of monitoring values ​​arranged continuously according to a unified time scale. Among them, the hydrogen sulfide, ammonia, methanethiol, and odor concentration sequences are used to characterize changes in malodorous and harmful gases, while the oxygen content sequence is used to characterize the oxygen content in the enclosed space. The names of the above gas items correspond to the hydrogen sulfide, ammonia, methanethiol, and odor concentration items in the management of malodorous pollutant emissions.

[0045] The airflow state sequence is used to represent the driving conditions for the migration of pollutants between spatial zones. It consists of a wind speed sequence, a wind direction sequence, and a pressure difference sequence. The wind speed sequence is used to record the strength of gas flow at the connecting edges, the wind direction sequence is used to record the direction of gas flow at the connecting edges, and the pressure difference sequence is used to record the pressure relationship between adjacent spatial zones. When forming the spatiotemporal state diagram, the airflow state sequence is written into the connecting edges, so that the connectivity between the source zone nodes, the active zone nodes, and the deodorization zone nodes has directionality.

[0046] The wastewater process sequence is used to represent the process causes of gas release. It consists of a booster pump status sequence, a liquid level status sequence, aeration status sequence, sludge return status sequence, sludge dewatering status sequence, chemical dosing status sequence, and tank cover status sequence. The above sequences are obtained by the wastewater treatment control system, liquid level detection device, aeration control device, sludge treatment equipment, chemical dosing equipment, and tank cover status detection device, and are arranged according to a unified time scale. The operation and management of urban wastewater treatment plants involves operation status recording and supervision and management requirements. After the above process sequence is accessed as operation-side data, it is used to establish a correlation with the gas monitoring sequence.

[0047] The device state sequence is used to represent the executable conditions of the control action. It consists of the ventilation device state sequence, the deodorization device state sequence, and the alarm device state sequence. The ventilation device state sequence records the start-up, stop, and operation status of the ventilation execution device, the deodorization device state sequence records the operating status of the deodorization access device, and the alarm device state sequence records the standby and trigger status of the alarm linkage device. The device state sequence and the airflow state sequence are written together into the connected edge, so that the counterfactual action evaluation model confirms that the corresponding execution device is in a callable state before generating candidate action codes.

[0048] For example: At a certain unified time scale, the state sequence of the booster pump changes from shutdown to operation, the state sequence of the pool liquid level changes synchronously, and then the hydrogen sulfide sequence and odor concentration sequence corresponding to the same source region node rise, and the wind direction sequence on the connected edge indicates that the airflow is from the source region node to the active region node. At this time, this set of data is represented in the spatiotemporal state diagram as a continuous record of "process change first, gas change later, and migration direction pointing to the active region", which is processed sequentially by the process gas time-delay causal model, the pollution migration inference model and the counterfactual action assessment model.

[0049] In this embodiment, the spatiotemporal state diagram is obtained by converting spatial partitioning relationships, connectivity relationships, and fused time-series data. The spatial partitioning relationships are used to indicate the affiliation of functional areas within the enclosed space of the wastewater treatment plant, including at least source areas that can generate polluting gases, activity areas where personnel stay or pass through, and deodorization areas connected to deodorization treatment. During the conversion, the control system first reads the area number, area purpose, and area location in the spatial partitioning relationships, generates source area nodes for areas that cause wastewater process disturbances, generates activity area nodes for areas where personnel inspect, operate, or pass through, and generates deodorization area nodes for areas connected to deodorization equipment.

[0050] Connectivity is used to indicate the gas accessibility between source area nodes, activity area nodes, and deodorization area nodes. During the transition, the control system reads the connection information corresponding to doors, passages, air ducts, inspection ports, pool cover gaps, and deodorization inlets, and generates connecting edges for the passable paths between adjacent nodes. Each connecting edge establishes a correspondence with its corresponding access control state sequence, airflow state sequence, and equipment state sequence, so that the connecting edge not only represents spatial connection, but also the opening state, airflow direction, and equipment callable state of the connection under the acquisition time sequence.

[0051] Node status codes are used to describe the state of a node itself during the data collection sequence. When generating a node status code, the gas monitoring sequence is written into the corresponding source area node or activity area node to indicate the changes in pollutant gas and oxygen content at that node; the personnel location sequence is written into the activity area node to indicate the node where the personnel are located; and the wastewater process sequence is written into the source area node to indicate the status of lifting, liquid level, aeration, sludge treatment, chemical dosing, or tank cover corresponding to that node. Thus, the same node can simultaneously have gas status, personnel status, and process status during the same data collection sequence.

[0052] Edge state coding is used to describe the state of connected edges during the acquisition time sequence. When generating edge state coding, the airflow state sequence is written to the connected edge to represent the wind speed, wind direction and pressure difference on the connected edge; the access control state sequence is written to the connected edge to represent whether the connected edge is in the open state; and the device state sequence is written to the connected edge to represent whether the ventilation device, deodorization device or alarm device associated with the connected edge is in the executable state.

[0053] After the above conversion is completed, the control system writes the node state code and edge state code sequentially according to the acquisition time sequence to form a continuous spatiotemporal state diagram. Each time segment in the spatiotemporal state diagram contains a node, a connected edge, a node state code, and an edge state code. This enables the subsequent process gas time-delay causal model to read the wastewater process sequence and gas monitoring sequence in the source area node, the pollution migration inference model to read the airflow state sequence and access control state sequence in the connected edge, and the counterfactual action evaluation model to read the equipment state sequence and generate the corresponding candidate action code.

[0054] In this embodiment, the process gas time-delay causal model is a time-series causal graph model established based on the spatiotemporal state graph. The time-series causal graph model includes a state transition library, a gas response library, a time-delay correlation matrix, a spatial reachability matrix, causal edge generation rules, and precursor coding rules. The state transition library is used to store state transition segments extracted from the wastewater process sequence and equipment state sequence. The gas response library is used to store gas response segments extracted from the gas monitoring sequence. The time-delay correlation matrix is ​​used to record the sequential correspondence between state transition segments and gas response segments. The spatial reachability matrix is ​​used to record the reachability relationship between source nodes and adjacent nodes via connected edges.

[0055] The perturbation fragment extraction layer reads the wastewater process sequence and equipment status sequence, and identifies the status changes of nodes in the same source area during the acquisition time sequence. When the status of the booster pump, liquid level, aeration, sludge return, sludge dewatering, dosing, tank cover, or associated equipment changes, the perturbation fragment extraction layer generates the continuous data before and after the change as a status transition fragment and writes it into the status transition library. The status transition fragment carries the source area node, acquisition time sequence, status transition category, and associated connected edges.

[0056] The response segment alignment layer reads the gas monitoring sequence and searches for gas changes after the state transition segment. During the search, the sequential correspondence between the gas response segment and the state transition segment is first determined based on the time delay correlation matrix. Then, the spatial reachability matrix is ​​used to determine whether the node where the gas response segment is located belongs to a node that can be influenced by the corresponding source region node through the connected edge. Gas changes that satisfy the above correspondence are generated as gas response segments and written into the gas response library. The gas response segment carries the node where the gas response is located, the acquisition time sequence, the gas change category, and the corresponding connected edge.

[0057] To quantify the correlation between state transition segments and gas response segments, in this embodiment, the causal edge generation layer calculates the source term precursor correlation degree according to the following formula. :

[0058] ;

[0059] When the source term precursor correlation degree satisfies the causal edge generation condition, the causal edge generation layer generates the corresponding causal edge code, wherein... This indicates the state transition segment number read from the state transition library. This indicates the gas response fragment number read from the gas response library. This indicates the acquisition timing interval between the state transition segment and the gas response segment. This indicates a time-delay matching term, used to characterize whether a gas response segment is located within an associable time series following a state transition segment; Indicates the state transition category in the state transition fragment. Indicates the type of gas change in the gas response segment. This indicates the category match between the state transition category and the gas change category; This represents the source region node corresponding to the state transition segment. Indicates the node where the gas response segment is located. This represents the reachability relationship from the source region node to the gas response node in the spatial reachability matrix; This represents the connected edges between the region node and the node where the gas response is located. This represents the airflow support term on the connected edge, used to characterize whether the airflow direction, gate state, and connectivity state of the connected edge support gas migration; This represents the interference suppression term, used to indicate the impact of access control changes, equipment changes, or external disturbances unrelated to this state transition segment on the association judgment; , , , , The weight coefficients for the corresponding items can be determined from the historical sample table or sample cache. Through the above calculations, the causal edge generation layer does not generate causal edge codes solely based on whether the gas monitoring sequence changes. Instead, it combines the temporal position of the state transition segment, the change category of the gas response segment, the spatial reachability relationship between the source node and the node where the gas response is located, and the airflow support relationship on the connected edges to determine whether changes in the wastewater process constitute a precursor to gas release.

[0060] The causal edge generation layer reads the state transition library, gas response library, time-delay correlation matrix, and spatial reachability matrix. When a state transition segment is located before a gas response segment in time sequence, and the node where the gas response segment is located is within the spatial reachability range of the corresponding source node, the causal edge generation layer generates a causal edge code. The causal edge code is used to represent the causal link from the process state transition of the source node to the node where the gas response is located. The causal edge code includes the source node, the node where the gas response is located, the connected edge, the state transition category, and the gas change category.

[0061] According to the precursor coding rules, the precursor coding layer combines the causal edge coding, source region nodes, and connected edges to form the source term precursor coding. The source term precursor coding is used to indicate that the corresponding source region node has shown a precursor to gas release and to indicate the connected edge corresponding to the precursor to gas release. The source term precursor coding is used as the input of the pollution migration inference model and is used to generate the subsequent migration path coding.

[0062] For example: A covered pool is a source node, and an inspection channel is an active node. There is a connecting edge between the two. When the state of the pool cover changes from closed to open, the change is written into the state transition library. Subsequently, if a hydrogen sulfide sequence change occurs in the source node or an active node reachable via the connecting edge, the change is written into the gas response library. The time-delay correlation matrix records the correspondence between the pool cover state change and the hydrogen sulfide sequence change. The spatial reachability matrix records how the source node can influence the active node via the connecting edge. The causal edge generation layer generates causal edge codes based on this, and the precursor coding layer generates source term precursor codes.

[0063] In this embodiment, the pollution migration inference model is a weighted directed graph inference model constructed based on a spatiotemporal state graph. It includes a source area location layer, a spatial chain graph attention layer, a flow direction constraint layer, and a path output layer. The calculation object of this model is not a single gas concentration value, but a spatial chain relationship composed of source area nodes, activity area nodes, deodorization area nodes, and connecting edges. In the spatial chain relationship, the nodes represent the location of pollution release, personnel activity, or deodorization access, and the connecting edges represent the doorways, ducts, passages, maintenance ports, pool cover gaps, or exhaust branches through which the gas can pass.

[0064] The source region localization layer reads the source term precursor code and obtains the source region node from the source term precursor code. The source region node serves as the starting point for pollution migration calculation. If the source term precursor code also contains a corresponding connected edge, the source region localization layer uses this connected edge as the initial connected edge that will be given priority in migration inference, so as to avoid blindly searching from all spatial paths.

[0065] The spatial chain graph attention layer is used to generate adjacency weights based on the adjacency relationships between nodes and connected edges in the spatiotemporal state graph. The adjacency weights represent the probability of polluted gas migrating from one node to another via the corresponding connected edge. When generating adjacency weights, it is first determined whether there are connected edges between the source area nodes and the activity area nodes and deodorization area nodes based on the spatiotemporal state graph. If there are connected edges, the corresponding adjacency relationship is established; if there are no connected edges, they are not included in the current migration path calculation. Thus, adjacency weights are only generated between nodes with actual spatial connectivity.

[0066] The flow constraint layer is used to write the airflow state sequence and access control state sequence into the adjacency weight. Specifically, the airflow state sequence is used to correct the migration direction on the connected edges, so that the adjacency weight has a directional attribute from the upstream node to the downstream node; the access control state sequence is used to correct the on / off state of the connected edges. When the door, inspection port, cover plate or passage is in the closed state, the corresponding connected edge does not participate in the subsequent output as a passable path; when it is in the open state, the corresponding connected edge participates in the path calculation. After this processing, the adjacency weight no longer only represents spatial adjacency, but also reflects spatial connectivity, airflow direction and access control on / off state.

[0067] The path output layer reads the adjacency weights after they have been written, and generates the migration path code starting from the source node and along the connected edges that have connectivity and satisfy the flow direction constraints. The migration path code includes at least the starting node, the connected edges traversed, the destination node, and the path pointing type. The path pointing type is used to distinguish whether the polluted gas tends to the activity zone node or the deodorization zone node, thereby providing a computable path basis for the subsequent counterfactual action evaluation model.

[0068] For example: A grid compartment is marked as a source area node, an inspection channel is marked as an activity area node, and an odor control branch pipe inlet is marked as an odor control area node. If the source term precursor code indicates that there is a gas release precursor in the grid compartment, and the grid compartments in the spatiotemporal state diagram are connected to the inspection channel and the odor control branch pipe inlet through connecting edges, then the spatial chain graph attention layer first establishes an adjacency weight for the above connecting edges; then the flow constraint layer corrects the adjacency weight according to the airflow state sequence and the access control state sequence; finally, the path output layer generates a migration path code pointing to the inspection channel or pointing to the odor control branch pipe inlet; this migration path code is then input into the counterfactual action evaluation model to determine whether subsequent candidate actions will cause polluted gas to enter the personnel activity area.

[0069] In this embodiment, the counterfactual action evaluation model is an action intervention graph model based on a spatiotemporal state graph. The action intervention graph model takes the current spatiotemporal state graph as the baseline state, the migration path code as the object to be intervened, and the device state sequence as the basis for action execution. By simulating changes in the connected edge state, device call state, and migration direction before execution, counterfactual migration results under different action conditions are obtained. The counterfactual action evaluation model includes an action generation layer, an action injection layer, a counterfactual reconstruction layer, and an action output layer. Here, "counterfactual" means that before the actual execution of the control action, the candidate action is written into the spatiotemporal state graph as a hypothetical action, and the migration path of the pollutant gas is recalculated.

[0070] The action generation layer reads the migration path code and the device status sequence. The migration path code indicates the path of polluted gas from the source area node to the activity area node or the deodorization area node via the connected edge. The device status sequence indicates whether the ventilation equipment, deodorization equipment and alarm equipment associated with the connected edge are in a callable state. The action generation layer first determines the connected edge involved in the migration path code, then reads the device status sequence corresponding to the connected edge, and generates the first action code, the second action code and the third action code according to the device callability relationship. The first action code, the second action code and the third action code represent three candidate intervention methods, respectively. Each action code includes the connected edge that is acted upon, the execution device that is called, and the action status corresponding to the execution device.

[0071] The action injection layer is used to write candidate actions into the spatiotemporal state graph. Specifically, the action injection layer does not change the original spatiotemporal state graph, but instead copies the original spatiotemporal state graph to form an action graph. Then, the first action code is written into the copied spatiotemporal state graph to obtain the first action graph; the second action code is written into another copied spatiotemporal state graph to obtain the second action graph; and the third action code is written into yet another copied spatiotemporal state graph to obtain the third action graph. When writing the action code, the state of the affected connected edges, the state of the associated devices, and the corresponding path pointers are updated synchronously, so that each action graph represents a spatial connectivity state under a hypothetical control action.

[0072] The counterfactual reconstruction layer is used to reconstruct the migration paths of the first action graph, the second action graph, and the third action graph respectively. During reconstruction, the counterfactual reconstruction layer takes the source region node as the starting point, reads the updated connected edge state, device state, and airflow direction in the action graph, and redetermines the nodes that the pollutant gas can reach and the connected edges it passes through. The first counterfactual migration code is obtained from the first action graph, the second counterfactual migration code is obtained from the second action graph, and the third counterfactual migration code is obtained from the third action graph. Each counterfactual migration code records the starting node, connected edges passed through, the destination node, and the path direction under the corresponding candidate action, which is used to represent the pollution migration result that may be formed after the candidate action is executed.

[0073] The action output layer is used to organize candidate actions and their corresponding reconstruction results. The action output layer writes the first action code, the second action code, and the third action code into the candidate action code, and writes the first counterfactual transition code, the second counterfactual transition code, and the third counterfactual transition code into the counterfactual transition code. Thus, the subsequent counterfactual judgment step can simultaneously read the candidate actions and their corresponding counterfactual transition results to determine whether a certain candidate action will cause pollutant gas to enter the activity area node corresponding to the personnel location sequence, or whether it will spread to the non-target area through the open connected edge corresponding to the access control state sequence.

[0074] For example: The migration path encoding shows that a source node points to an active node via a connected edge, and the device state sequence indicates that the ventilation device and deodorization device corresponding to the connected edge are both in a callable state. Based on this, the action generation layer generates a first action code, a second action code, and a third action code. The action injection layer writes the three action codes into three action graphs respectively. The counterfactual reconstruction layer reconstructs the migration paths in the three action graphs respectively. If the first counterfactual migration code still points to the active node, while the second counterfactual migration code points to the deodorization node, the two will be distinguished and processed in the subsequent counterfactual judgment, so that the target action code is not generated simply based on whether the device can be started, but based on the migration reconstruction result before the candidate action is executed.

[0075] In this embodiment, the counterfactual determination is a data processing procedure that performs an exclusionary judgment on the counterfactual migration code. The counterfactual determination does not directly generate a new control action, but first determines whether the assumed execution result corresponding to the candidate action code will form an unacceptable contaminated migration path, and then determines the target action code from the candidate action codes that have not been excluded.

[0076] Specifically, the counterfactual transition code is generated by the counterfactual action evaluation model. It records the transition results obtained after the candidate action is written into the spatiotemporal state diagram. The predicted arrival node represents the node that the pollutant gas may reach under the candidate action condition, and the predicted traversed edge represents the connected edge that the pollutant gas may traverse under the candidate action condition. When performing counterfactual judgment, the control system first reads the predicted arrival node and predicted traversed edge from the counterfactual transition code, and compares them with the personnel position sequence and access control state sequence, respectively.

[0077] When the predicted arrival node matches the activity area node corresponding to the personnel location sequence, it indicates that the candidate action may cause pollutant gas to enter the personnel's area after it is assumed to be executed. The control system generates a first counter-evidence code and writes the first counter-evidence code into the corresponding candidate action code. The first counter-evidence code is used to mark the personnel exposure risk. It is not directly output as an alarm result, but serves as the basis for subsequently eliminating the candidate action code.

[0078] When the predicted traversed edge matches the connected edge corresponding to the access control state sequence, it indicates that the candidate action, if assumed to be executed, may cause polluting gas to diffuse outward through the currently open door, passage, maintenance port, or other connected path. The control system generates a second counter-proof code and writes the second counter-proof code into the corresponding candidate action code. The second counter-proof code is used to mark the diffusion risk on the open connected path.

[0079] After completing the above comparison, the control system reads whether the candidate action codes contain the first counter-proof code and the second counter-proof code. For candidate action codes containing the first counter-proof code, it is determined that there is a personnel exposure path; for candidate action codes containing the second counter-proof code, it is determined that there is an open connected edge diffusion path; for candidate action codes that do not contain the first counter-proof code and the second counter-proof code, they are determined as the target action codes.

[0080] To avoid determining the target action code solely based on the generation order of candidate action codes, the control system can also calculate the counter-evidence screening value corresponding to each candidate action code using the following formula. :

[0081] ;

[0082] The target action code is determined according to the following formula. :

[0083] ;

[0084] in, Indicates the first One candidate action code; This represents the set of candidate action codes formed by the first action code, the second action code, and the third action code; The personnel exposure item is set to a risk state when the predicted arrival node in the counterfactual transfer code corresponding to the kkk-th candidate action code is consistent with the activity area node corresponding to the personnel location sequence. This indicates that the open connected edge diffusion term is set to a risk state when the predicted traversed edge matches the open connected edge corresponding to the access control state sequence. This indicates a non-target migration term, used to characterize whether the counterfactual migration code corresponding to the candidate action code points to a non-deodorization zone node; This represents the deodorization guidance term, used to characterize whether the counterfactual transition code corresponding to the candidate action code points to the deodorization zone node; This indicates a non-executable item for a device, used to characterize whether the corresponding execution device in the device state sequence is in an uncallable state; Indicates an action is executable, when the first... When the execution device corresponding to each candidate action code is in a callable state. The status is now eligible for screening; , , , , The weight coefficients for the corresponding items can be determined by the sample cache or manually configured security priorities. Through the above formula, when the control system determines the target action code, it does not simply select the execution device that can be started, but incorporates the personnel location sequence, access control status sequence, counterfactual migration code, deodorization zone node and equipment status sequence into the screening, so that the personnel exposure risk, the risk of opening the connected edge diffusion, the non-target migration risk, the deodorization guidance result and the equipment execution conditions are processed uniformly in the same calculation process.

[0085] For example: If a certain counterfactual migration code shows that after a candidate action is executed, the predicted arrival node of the pollutant gas is the activity area node corresponding to the inspection channel, and the personnel location sequence shows that the inspection personnel are located in the activity area node, then the candidate action code is written into the first counterfactual code. If the predicted traversed edge corresponding to another candidate action is the connected edge corresponding to the maintenance door that is in the open state, then the candidate action code is written into the second counterfactual code. The remaining candidate action codes that are not written into the first and second counterfactual codes are determined as target action codes and used to convert them into linkage execution codes later.

[0086] In this embodiment, the linkage execution code is a device execution data structure converted from the target action code. It is used to convert the action result output by the model into control content that the execution device can recognize. The linkage execution code consists of action object code, action type code, execution edge code, and back sampling trigger code. The action object code is used to identify the execution device that needs to be called, the action type code is used to identify the type of action that the execution device should perform, the execution edge code is used to identify the connected edge that the action acts on, and the back sampling trigger code is used to identify the data content that needs to be re-collected after the action is executed.

[0087] During the transition, the control system first reads the action object information in the target action code and writes it into the action object code. The action object code corresponds to the device identifier in the device state sequence, thereby determining the specific execution device. The control system then reads the path action object in the target action code and writes it into the execution edge code, so that the action of the execution device is limited to the corresponding connected edge, rather than uniformly linking all devices in the closed space of the sewage treatment plant.

[0088] Subsequently, the control system generates corresponding execution instructions based on the action type code, causing the execution device to change the edge state of the connected edge. The edge state includes the on / off state, airflow guidance state, or deodorization access state of the connected edge under the current control action. Thus, the linkage execution code is not a simple on / off command, but binds the "execution device, action type, and connected edge" to the same control unit, ensuring that the control action corresponds to the aforementioned migration path code and counterfactual migration code.

[0089] The sampling trigger code is triggered after the execution equipment completes the action. Based on the sampling trigger code, the control system collects the gas monitoring sequence and the airflow state sequence after execution. The gas monitoring sequence after execution is used to indicate the gas changes after the target action is executed, and the airflow state sequence after execution is used to indicate the airflow changes on the connected edge after the target action is executed. The control system binds the gas monitoring sequence after execution, the airflow state sequence after execution, and the counterfactual migration code generated before execution to form a correction sample.

[0090] The calibration sample is not a separate historical record, but is used to record the correspondence between the "predicted migration result" and the "actual recovery result after execution". When writing, the control system writes the calibration sample into the sample cache of the process gas time-delay causal model, the pollution migration inference model and the counterfactual action evaluation model respectively. The sample cache is used to save the reference data that needs to be called when the model is updated in the future, so that the subsequent processing can correct the source term precursors, migration paths and candidate action generation process according to the gas changes and airflow changes after execution.

[0091] For example: The target action code corresponds to the connected edge between a source area node and a deodorization area node. The action object code determines the execution device associated with the connected edge. The execution edge code determines the connected edge. The action type code causes the execution device to change the edge state of the connected edge. After the action is executed, the back sampling trigger code calls the gas monitoring sequence and airflow state sequence near the connected edge. If there is a difference between the gas monitoring sequence after execution and the prediction result corresponding to the counterfactual transfer code, the difference is written into the correction sample along with the counterfactual transfer code for subsequent sample cache updates. Example 2

[0092] like Figure 2 As shown, this embodiment further improves the design based on Embodiment 1. The difference is that in the actual operation of Embodiment 1, it was found that the enclosed space of the sewage treatment plant is in a high-humidity, corrosive gas and airflow disturbance environment for a long time. Some gas monitoring sequences or airflow state sequences may experience short-term distortion, lag fall or single-point drift, resulting in inconsistency between the executed gas monitoring sequence and the executed airflow state sequence and the actual pollution migration results. It was not guaranteed that the correction samples written to the sample cache were all reliable samples. Based on this, the enclosed space air quality control method based on multi-sensor fusion and AI algorithm also includes: before forming the correction sample, performing a re-collection reliability judgment on the executed gas monitoring sequence and the executed airflow state sequence, and determining the writing method of the correction sample based on the re-collection reliability judgment result.

[0093] Specifically, after receiving the data acquisition trigger code, the control system does not directly write the gas monitoring sequence and the airflow state sequence after execution into the sample cache. Instead, it first generates a data acquisition discrimination group, which includes the counterfactual transition code before execution, the gas monitoring sequence after execution, the airflow state sequence after execution, the execution receipt status in the equipment status sequence, and the access control status sequence. The execution receipt status is used to indicate whether the executing equipment has completed the action according to the linkage execution code, and the access control status sequence is used to indicate whether the corresponding connected edge is still in the originally assumed on / off state after execution.

[0094] When performing reliable data acquisition determination, the control system first reads the predicted traversed edge and predicted arrived node in the counterfactual migration code, then reads the airflow change corresponding to the predicted traversed edge in the airflow state sequence after execution, and reads the gas change corresponding to the predicted arrived node in the gas monitoring sequence after execution. If the executing device has completed the action, and the access control state sequence does not show that the corresponding connected edge has been opened additionally, and the direction of the airflow change after execution corresponds to the path direction in the counterfactual migration code, then the data acquisition determination group is marked as the first reliable data acquisition group.

[0095] If the gas monitoring sequence changes after execution, but the airflow state sequence does not show a corresponding airflow change at the corresponding connected edge, or the access control state sequence shows an open connected edge that has not been recorded by the counterfactual migration code, the control system marks the data collection group as the first isolated data collection group. The first isolated data collection group is not directly written into the sample cache of the process gas time-delay causal model, the pollution migration inference model, and the counterfactual action assessment model, but is written into the isolated sample area and an isolation cause code is attached. The isolation cause code is used to indicate that the data collection data may be affected by sensor distortion, access control state changes, or non-target connected edge disturbances.

[0096] For data marked as the first reliable data acquisition group, the control system binds the executed gas monitoring sequence, the executed airflow state sequence, and the counterfactual migration code to form a correction sample and writes it into the sample cache. For data marked as the first isolated data acquisition group, the control system only saves it as data to be verified and does not participate in the update of the source term precursor, migration path, and candidate action generation process. Thus, the sample cache stores data that has been jointly verified by execution receipts, access control status, and airflow changes, avoiding the formation of incorrect associations in subsequent models due to short-term drift of a single gas sensor or temporary changes in connectivity edges.

[0097] For example, a target action code points to a deodorization zone node, and a counterfactual transition code indicates that pollutant gas should enter the deodorization zone node via a specified connected edge. After the action is executed, if the execution receipt status shows that the deodorization device has been invoked, the access control status sequence shows that the connected edge remains passable, and the airflow status sequence after execution shows that the airflow direction is consistent with the counterfactual transition code, then the collected data forms the first reliable collected data group and is written to the sample cache. If the gas monitoring sequence after execution shows that the gas in the active zone node changes significantly, but the airflow status sequence does not show that pollutant gas enters the active zone node via the predicted passing edge, and the access control status sequence shows that the maintenance door is opened, then the collected data forms the first isolated collected data group, which is not written to the sample cache but is retained as an isolated sample. Example 3

[0098] like Figure 3 As shown, based on the same inventive concept as the enclosed space air quality control method based on multi-sensor fusion and AI algorithm in the foregoing embodiments, this application provides an enclosed space air quality control system based on multi-sensor fusion and AI algorithm. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0099] Data acquisition module: The data acquisition module collects spatial partitioning relationships, connectivity relationships, and fused time-series data. The fused time-series data includes gas monitoring sequences, airflow status sequences, personnel location sequences, access control status sequences, equipment status sequences, and wastewater process sequences.

[0100] Graph construction module: Based on the spatial partitioning relationship and the connectivity relationship, the graph construction module converts the fused time series data into a spatiotemporal state graph, which includes source region nodes, active region nodes, deodorization region nodes and connecting edges;

[0101] Precursor identification module: The precursor identification module inputs the spatiotemporal state diagram into the process gas time-delay causal model, performs time-delay causal calculations on the wastewater process sequence, the equipment state sequence, and the gas monitoring sequence, and obtains the source term precursor code;

[0102] Migration Inference Module: The migration inference module inputs the spatiotemporal state diagram, the source term precursor code, the airflow state sequence, the personnel location sequence, and the access control state sequence into the pollution migration inference model to obtain the migration path code;

[0103] Action Pre-simulation Module: The action pre-simulation module inputs the migration path code and the device state sequence into the counterfactual action evaluation model to obtain candidate action codes and counterfactual migration codes;

[0104] The counterfact filtering module: Based on the personnel location sequence and the access control status sequence, the counterfact filtering module performs counterfact judgment on the counterfact transfer code to obtain the target action code;

[0105] Command output module: The command output module converts the target action code into linkage execution code and outputs the linkage execution code.

[0106] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0107] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application 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 application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.

Claims

1. A method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms, characterized in that the method... include: Collect spatial zoning relationships, connectivity relationships, and fused time-series data. The fused time-series data includes gas monitoring sequences, airflow status sequences, personnel location sequences, access control status sequences, equipment status sequences, and wastewater process sequences. Based on the spatial partitioning relationship and the connectivity relationship, the fused time series data is converted into a spatiotemporal state diagram, which includes source region nodes, active region nodes, deodorization region nodes and connecting edges; The spatiotemporal state diagram is input into the process gas time-delay causal model, and time-delay causal calculations are performed on the wastewater process sequence, the equipment state sequence, and the gas monitoring sequence to obtain the source term precursor code; The spatiotemporal state diagram, the source term precursor code, the airflow state sequence, the personnel location sequence, and the access control state sequence are input into the pollution migration inference model to obtain the migration path code; The migration path code and the device state sequence are input into the counterfactual action evaluation model to obtain candidate action codes and counterfactual migration codes. Based on the personnel location sequence and the access control status sequence, the counterfactual transition code is subjected to a counterfactual verification to obtain the target action code; The target action code is converted into a linkage execution code, and the linkage execution code is output.

2. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 1, characterized in that, The gas monitoring sequence consists of hydrogen sulfide, ammonia, methanethiol, odor concentration, and oxygen content sequences; the airflow status sequence consists of wind speed, wind direction, and pressure difference sequences; the wastewater process sequence consists of booster pump status sequence, liquid level status sequence, aeration status sequence, sludge return status sequence, sludge dewatering status sequence, chemical dosing status sequence, and tank cover status sequence; the equipment status sequence consists of ventilation equipment status sequence, deodorization equipment status sequence, and alarm equipment status sequence.

3. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 1, characterized in that, The transformation of the spatiotemporal state diagram includes: converting the spatial partitioning relationship into source area nodes, activity area nodes, and deodorization area nodes; converting the connectivity relationship into connectivity edges; encoding the gas monitoring sequence, personnel location sequence, and wastewater process sequence into node state codes; encoding the airflow state sequence, access control state sequence, and equipment state sequence into edge state codes; and writing the node state codes and edge state codes into the spatiotemporal state diagram along the acquisition time sequence.

4. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 1, characterized in that, The process gas time-delay causal model includes a perturbation fragment extraction layer, a response fragment alignment layer, a causal edge generation layer, and a precursor coding layer. The perturbation fragment extraction layer extracts state transition fragments from the wastewater process sequence and the equipment state sequence. The response fragment alignment layer extracts gas response fragments located after the state transition fragments from the gas monitoring sequence. The causal edge generation layer generates causal edge codes based on the sequential relationship between the state transition fragments and the gas response fragments. The precursor coding layer writes the causal edge codes, the source region nodes, and the connected edges into the source term precursor codes.

5. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 1, characterized in that, The pollution migration inference model includes a source region location layer, a spatial chain graph attention layer, a flow direction constraint layer, and a path output layer. The source region location layer reads the source region nodes from the source term precursor codes. The spatial chain graph attention layer assigns adjacency weights to the source region nodes, the connected edges, the active area nodes, and the deodorization area nodes based on the spatiotemporal state graph. The flow direction constraint layer writes the airflow state sequence and the access control state sequence into the adjacency weights. The path output layer generates the migration path code based on the written adjacency weights.

6. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 1, characterized in that, The counterfactual action evaluation model includes an action generation layer, an action injection layer, a counterfactual reconstruction layer, and an action output layer. The action generation layer generates a first action code, a second action code, and a third action code based on the migration path code and the device state sequence. The action injection layer sequentially writes the first action code, the second action code, and the third action code into the spatiotemporal state diagram to obtain a first action diagram, a second action diagram, and a third action diagram. The counterfactual reconstruction layer generates a first counterfactual migration code, a second counterfactual migration code, and a third counterfactual migration code based on the first action diagram, the second action diagram, and the third action diagram. The action output layer writes the first action code, the second action code, and the third action code into the candidate action code, and writes the first counterfactual transition code, the second counterfactual transition code, and the third counterfactual transition code into the counterfactual transition code.

7. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 1, characterized in that, The counterfactual determination includes: reading the predicted arrival node and predicted traversed edge from the counterfactual migration code; generating a first counterfactual code when the predicted arrival node is consistent with the activity area node corresponding to the personnel location sequence; generating a second counterfactual code when the predicted traversed edge is consistent with the connected edge corresponding to the access control state sequence; and determining the candidate action code that does not contain the first counterfactual code and does not contain the second counterfactual code as the target action code.

8. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 1, characterized in that, The linkage execution code consists of action object code, action type code, execution edge code, and data acquisition trigger code. The execution device is determined based on the action object code; the connected edge is determined based on the execution edge code; the execution device changes the edge state of the connected edge based on the action type code; the gas monitoring sequence and the airflow state sequence after execution are collected based on the data acquisition trigger code; the gas monitoring sequence, the airflow state sequence, and the counterfactual migration code after execution are used to form a correction sample; the correction sample is written into the sample cache of the process gas time-delay causal model, the pollution migration inference model, and the counterfactual action evaluation model.

9. The method for controlling air quality in enclosed spaces based on multi-sensor fusion and AI algorithms according to claim 8, characterized in that, Also includes: Based on the execution receipt, access control status sequence, and airflow status sequence after execution, the calibration samples are judged to be trustworthy. Trustworthy calibration samples are written into the sample cache, and untrustworthy calibration samples are written into the isolation sample area.

10. An air quality control system for enclosed spaces based on multi-sensor fusion and AI algorithms, characterized in that: The system for performing the method according to any one of claims 1 to 8, the system comprising: The data acquisition module collects spatial partitioning relationships, connectivity relationships, and fused time-series data. The fused time-series data includes gas monitoring sequences, airflow status sequences, personnel location sequences, access control status sequences, equipment status sequences, and wastewater process sequences. The graph construction module converts the fused time series data into a spatiotemporal state graph based on the spatial partitioning relationship and the connectivity relationship. The spatiotemporal state graph includes source region nodes, active region nodes, deodorization region nodes, and connecting edges. The precursor identification module inputs the spatiotemporal state diagram into the process gas time-delay causal model, performs time-delay causal calculations on the wastewater process sequence, the equipment state sequence, and the gas monitoring sequence, and obtains the source term precursor code. The migration inference module inputs the spatiotemporal state diagram, the source term precursor code, the airflow state sequence, the personnel location sequence, and the access control state sequence into the pollution migration inference model to obtain the migration path code; The action pre-simulation module inputs the migration path code and the device state sequence into the counterfactual action evaluation model to obtain candidate action codes and counterfactual migration codes; The counterfactual filtering module performs counterfactual discrimination on the counterfactual migration code based on the personnel location sequence and the access control status sequence to obtain the target action code. The instruction output module converts the target action code into a linkage execution code and outputs the linkage execution code.

Images