An all-process digital intelligent analysis system for environmental impact assessment information

Abstract

The application relates to the technical field of industrial environment monitoring and intelligent data analysis, and discloses a full-process digital intelligent analysis system for environment impact assessment information, which comprises a full-process analysis server, a multi-modal data acquisition subsystem and a visual interactive terminal. The server is internally provided with a heterogeneous environment causal dependence graph, and runs data mapping, dynamic topology evolution, forward calculation, double-layer shadow verification and counterfactual reasoning competition modules. The system uses a discrete event-driven graph to rewrite rules, updates topology and generates a benchmark graph; calculates the deviation between predicted values and measured values of pollutants, instantiates a shadow copy to perform parameter back propagation and physical conservation verification when the deviation exceeds the standard; if the quality balance verification fails, counterfactual reasoning is triggered, a variation set is generated based on a fault topology library, and the comprehensive fitness is calculated after spatial pruning to screen the optimal explanation graph. The application realizes accurate attribution of environmental hazards by combining physical conservation and causal reasoning.

Description

Technical Field

[0001] This invention relates to the field of industrial environmental monitoring and intelligent data analysis technology, specifically to a full-process digital intelligent analysis system for environmental impact assessment information. Background Technology

[0002] With the advancement of industrial digital transformation, environmental impact assessment and full-process supervision are gradually shifting from traditional static approval to dynamic digital monitoring. Modern heavily polluting industries such as chemical, pharmaceutical, and metallurgical industries are typically equipped with distributed control systems and online environmental monitoring instruments to collect real-time production data and pollutant emission data. Simulating and analyzing the entire process of pollution generation, treatment, and discharge using mathematical models is a key means to achieve pollution control and compliant supervision.

[0003] However, existing environmental impact assessment and monitoring systems still have technical limitations when dealing with complex industrial sites. Current analysis systems rely on pre-defined static process flow models to calculate pollutant transport, neglecting the real-time impact of frequent discrete control events (such as valve switching, bypass opening, and standby equipment activation) on physical connectivity in industrial sites. This mismatch between the static model and dynamic operating conditions causes the system to fail to accurately reflect the actual pollutant migration path at any given moment, resulting in significant calculation errors under non-steady-state conditions.

[0004] Furthermore, existing fault diagnosis methods based on pure data-driven approaches focus on numerical fitting accuracy, lacking physical interpretability and constraint mechanisms. When deviations occur between predicted and measured data, traditional techniques struggle to effectively distinguish whether such deviations stem from parametric errors such as sensor drift and equipment efficiency fluctuations, or from structural faults such as hidden leaks or illegal discharges. Lacking rigid verification of the law of conservation of mass and the ability to reason about counterfactual causal logic, existing systems cannot generate definitive fault attribution conclusions when faced with highly concealed unintended emissions, failing to meet the urgent environmental regulatory needs for accurate source tracing and evidence fixation. Therefore, this invention proposes a full-process digital intelligent analysis system for environmental impact assessment information to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a full-process digital intelligent analysis system for environmental impact assessment information, which solves the problem of difficulty in monitoring and attributing hidden faults in pollutant emissions under complex industrial environments.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a full-process digital intelligent analysis system for environmental impact assessment information, including a full-process analysis server, a multimodal data acquisition subsystem, and a visual interactive terminal.

[0008] The full-process analysis server is configured as the core computing device of the system, and the multimodal data acquisition subsystem is configured to collect analog monitoring data and discrete control data from the industrial production site and send them to the full-process analysis server through the industrial communication network.

[0009] The visualization and interactive terminal is configured to receive and display the analysis results generated by the end-to-end analysis server. The end-to-end analysis server includes a central computing processor and non-volatile storage media, which stores computer program instructions and heterogeneous environment causal dependency graph data structures.

[0010] The central computing processor is configured to execute computer program instructions to run the multimodal data mapping module, the dynamic topology evolution module, the forward computation and bias diagnosis module, the two-layer shadow verification module, and the counterfactual reasoning competition module.

[0011] The multimodal data mapping module maps analog monitoring data to node state vectors and discrete control data to discrete event vectors. The analog monitoring sensor group includes continuous monitoring instruments located at pollution-generating, treatment, and emission nodes, responsible for collecting analog monitoring data. The discrete controller interface unit connects to an industrial distributed control system or programmable logic controller (PLC), responsible for collecting the switching status signals of field actuators as discrete control data.

[0012] The heterogeneous environmental causal dependency graph data structure is defined as a directed acyclic graph, containing a set of nodes and a set of edges. The set of nodes is divided into a source node subset, a transformation node subset, and a sink node subset. The source node subset corresponds to the initial pollutant generation unit, the transformation node subset corresponds to pollutant treatment facilities or physical transmission channels, and the sink node subset corresponds to environmental monitoring points. Each directed edge in the edge set is associated with a transfer parameter matrix, which stores the transmission efficiency coefficients of material flow between physical nodes.

[0013] The dynamic topology evolution module updates the topology of the causal dependency graph of heterogeneous environments based on discrete event vectors and a pre-defined graph rewriting rule base, generating a baseline running graph. At each discrete time step, the module performs a topology update operation, applying connectivity enabling or connectivity blocking rules from the graph rewriting rule base based on changes in the element values ​​of the discrete event vectors. The connectivity enabling rule is configured to set the value of the corresponding coordinate in the adjacency matrix of the baseline running graph to 1 when an index element in the discrete event vector changes from 0 to 1; the connectivity blocking rule is configured to set the value of the corresponding coordinate in the adjacency matrix of the baseline running graph to 0 when an index element in the discrete event vector changes from 1 to 0.

[0014] The forward computation and deviation diagnosis module performs pollutant migration and transformation calculations within the baseline operating graph and calculates the system deviation metric between the predicted output vector and the measured vector. This module performs topological sorting on the baseline operating graph to generate a linear computation sequence and calculates the predicted output state vector of downstream nodes based on the state transfer equation. The state transfer equation includes a weighted sum of the output state of the upstream parent node and the transfer parameter matrix, as well as node bias terms. The module calculates the mean square error between the predicted output vector of the entire system corresponding to the subset of sink nodes and the sensor measured vector as the system deviation metric.

[0015] The two-layer shadow verification module is used to instantiate a copy of the shadow map in an independent memory space when the system deviation metric exceeds a preset threshold, and to perform parameter gradient backpropagation calculations and physical conservation constraint verification. The two-layer shadow verification module creates a copy of the baseline running map as the first-level shadow map, calculates the gradient of the system deviation metric with respect to the transfer parameter matrix, and corrects the transfer parameter matrix in the first-level shadow map based on the gradient descent method.

[0016] The two-layer shadow verification module performs physical conservation constraint verification on the corrected first-level shadow map and calculates the mass imbalance residual of the transformation nodes. If the mass imbalance residual exceeds the physical allowable error limit, the verification is deemed a failure, and a structural variation trigger signal is generated and sent to the counterfactual reasoning competition module. This process uses physical conservation laws as a criterion to distinguish between system parameter drift and topological structural faults.

[0017] The counterfactual reasoning competition module is used to generate a set of candidate variant maps based on the fault mode topology library when the physical conservation constraint verification fails, and to select the optimal interpretation map by calculating the comprehensive posterior fitness score to generate a hidden fault diagnosis conclusion.

[0018] The counterfactual reasoning competition module responds to the structural mutation trigger signal by invoking a fault mode topology library that stores implicit leakage mode templates and unexpected bypass mode templates. The counterfactual reasoning competition module applies mutation operators to perform nonlinear modification operations on the adjacency matrix of the baseline running graph, generating a set of candidate mutation graphs.

[0019] For the latent leakage mode template, the edge-increasing operation operator is executed to correct the element values ​​of the corresponding coordinates between non-sink nodes and environmental receptor nodes in the adjacency matrix from 0 to 1; for the unexpected bypass mode template, the reconnection operation operator is executed to block the path pointing to the governance facility node in the adjacency matrix and establish a short-circuit path across the governance facility node.

[0020] The counterfactual reasoning competition module is also configured to perform spatial pruning operations, read geographic information system data or factory 3D digital model data, calculate the Euclidean physical distance between nodes, and remove candidate variant maps whose Euclidean physical distance is greater than the preset physical connection limit distance from the candidate variant map set.

[0021] The overall posterior fitness score is a weighted sum of the fitting residuals, parameter drift penalty, and structural complexity. The fitting residuals are the norm distance between the predicted output vector of the candidate variogram and the measured vector; the parameter drift penalty is the difference between the fine-tuned transfer parameter matrix and the initial design parameter matrix; and the structural complexity is the Hamming distance between the variogram adjacency matrix of the candidate variogram and the adjacency matrix of the baseline variogram.

[0022] The counterfactual reasoning competition module selects the candidate variant map with the smallest comprehensive posterior fitness score as the optimal interpretation map, calculates the set of topological differences between the optimal interpretation map and the baseline running map, instantiates the set of topological differences as latent fault event records, and marks them at the corresponding coordinate positions on the visualized image through the visual interactive terminal.

[0023] This invention provides a full-process digital intelligent analysis system for environmental impact assessment information. It has the following beneficial effects:

[0024] 1. This invention utilizes a dynamic topology evolution module to drive graph rewriting of the rule base using discrete event vectors, achieving real-time synchronization between the causal dependency graph of heterogeneous environments and the physical state of industrial sites. This mechanism, based on updating the topology structure using discrete control signals, can map changes in the process flow caused by operations such as valve switching and equipment start-up and shutdown. It solves the calculation deviation problem caused by structural solidification under unsteady conditions in traditional environmental assessment models, ensuring that the baseline operating graph always accurately reflects the physical transport path of pollutants at the current moment.

[0025] 2. This invention utilizes a dual-layer shadow verification module to combine parameter gradient backpropagation with physical conservation constraint verification, establishing a rigid criterion for distinguishing between system parameter drift and structural hazards. By calculating the quality imbalance residuals of the shadow map replica in an independent memory space, the system only triggers the structural variation process when physical conservation fails, avoiding false alarms caused by sensor noise or parameter fluctuations, and improving the diagnostic accuracy and interpretability for structural faults such as hidden leakage or surreptitious discharge.

[0026] 3. This invention, based on a counterfactual reasoning competition module combined with a fault mode topology library and spatial pruning operations, achieves automated inversion and localization of unknown fault modes. By calculating a comprehensive posterior fitness score that includes fitting residuals, parameter drift, and structural complexity, the system mathematically filters fault topologies that conform to physically measured data. Under the constraint of geographic information data, it eliminates physically inaccessible false paths, thereby identifying and marking the locations of unexpected bypasses or hidden leaks, providing a traceability basis for environmental monitoring. Attached Figure Description

[0027] Figure 1 This is a system structure block diagram of the present invention;

[0028] Figure 2 This is a flowchart of the two-layer shadow verification and counterfactual reasoning competition logic of the present invention. Detailed Implementation

[0029] The technical solutions in 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.

[0030] See attached document Figure 1 -Appendix Figure 2 This invention provides a full-process digital intelligent analysis system for environmental impact assessment information. The full-process digital intelligent analysis system includes a full-process analysis server, a multimodal data acquisition subsystem, a visual interactive terminal, and an industrial communication network.

[0031] The end-to-end analysis server is the core computing device of the end-to-end digital intelligent analysis system. The end-to-end analysis server establishes a two-way data connection with the multimodal data acquisition subsystem through the industrial communication network, and establishes a one-way or two-way communication connection with the visualization interactive terminal through the encrypted transmission protocol.

[0032] The multimodal data acquisition subsystem is deployed at the industrial production site. It includes an analog quantity monitoring sensor group and a discrete quantity controller interface unit. The analog quantity monitoring sensor group comprises continuous monitoring instruments installed at pollution generation, treatment, and emission points. These instruments include, but are not limited to, continuous flue gas monitoring systems, flow meters, temperature sensors, and pressure transmitters.

[0033] Discrete quantity controller interface units are connected to industrial distributed control systems or programmable logic controllers via hardwiring or fieldbus. The discrete quantity controller interface unit is configured to acquire the on / off status signals of field actuators, including valves, fan motors, pump motors, and bypass switches.

[0034] The end-to-end analysis server is equipped with a system bus, a central processing unit (CPU), non-volatile storage media, and a network communication interface. The CPU, non-volatile storage media, and network communication interface are all physically connected and transmit electrical signals via the system bus.

[0035] Non-volatile storage media are used to store computer program instructions and heterogeneous environment causal dependency graph data structures. In non-volatile storage media, a heterogeneous environment causal dependency graph data structure is defined as a directed acyclic graph containing a set of nodes and a set of edges.

[0036] The node set is divided into source node subsets, transformation node subsets, and sink node subsets in the storage space. Each edge in the edge set is associated with a transfer parameter matrix. The transfer parameter matrix stores the transmission efficiency coefficients of material or energy flow between physical nodes.

[0037] The central processing unit (CPU) is configured to retrieve and execute computer program instructions from non-volatile storage media. While executing these instructions, the CPU receives analog status data from the analog monitoring sensor array via a network communication interface, and receives discrete event data from the discrete controller interface unit.

[0038] The central computing processor is configured to map analog state data to node state vectors and discrete event data to discrete event vectors. Based on the discrete event vectors, the central computing processor calls the graph rewriting rule base stored in non-volatile storage to perform topology update operations on the causal dependency graph data structure of the heterogeneous environment.

[0039] The internal logic circuitry of the central computing processor is further divided into a baseline graph computation unit and a shadow graph verification unit. The baseline graph computation unit is configured to perform forward evolution computation based on the current topology. The shadow graph verification unit is configured to instantiate at least one copy of the shadow graph in a separate memory address space.

[0040] The shadow map verification unit is configured to perform parametric gradient backpropagation and physical constraint verification operations in a shadow map replica. The central computing processor generates latent fault location coordinates by comparing the output values ​​of the baseline map calculation unit and the output values ​​of the shadow map verification unit.

[0041] The visualization interactive terminal is configured to receive and display a visualization image of the heterogeneous environment causal dependency graph generated by the full-process analysis server. The visualization interactive terminal marks the coordinates of latent faults at the corresponding coordinate positions in the visualization image and displays the confidence level value of the latent faults.

[0042] The computer program instructions stored in the non-volatile storage medium of the end-to-end analysis server are logically divided into a map initialization and construction module, a multimodal data mapping module, a dynamic topology evolution module, a forward computation and deviation diagnosis module, a two-layer shadow verification module, and a counterfactual reasoning competition module. The end-to-end analysis server calls and executes the instruction codes corresponding to the above modules through the central computing processor to realize the end-to-end analysis of environmental impact assessment information.

[0043] The graph initialization module is configured to parse digital engineering documents and establish an initial environmental causal dependency graph structure. This initial structure includes a set of nodes and a set of edges. The module assigns an initial transfer parameter matrix to each directed edge in the edge set. The values ​​of the initial transfer parameter matrix are derived from the theoretical pollution generation coefficient or design removal efficiency in the environmental impact assessment report.

[0044] The multimodal data mapping module is configured to vectorize the acquired time-series data. It converts analog monitoring data into node state vectors, with the dimension of the node state vectors matching the number of nodes in the node set. The module also converts discrete switching data into discrete event vectors.

[0045] The dynamic topology evolution module is configured to update the connectivity relationships of the graph based on discrete event vectors. The module stores a graph rewriting rule base and performs topology update operations at each time step. The graph instance generated by the dynamic topology evolution module, containing the current adjacency matrix, is defined as the baseline running graph.

[0046] The forward computation and bias diagnosis module is configured to perform pollutant migration and transformation calculations within the baseline operating graph. For any node in the graph, the forward computation and bias diagnosis module calculates the node's current output state vector based on the output states of its parent node set. The module further calculates the Euclidean distance between the predicted output vector and the measured vector of the sink node as the system bias.

[0047] The two-level shadow verification module is configured to activate when the system deviation exceeds a preset threshold. The two-level shadow verification module first creates a copy of the baseline running graph as the first-level shadow graph. The two-level shadow verification module performs backpropagation of the parameter gradients in the first-level shadow graph to calculate the corrected gradient of the transferred parameter matrix.

[0048] ;

[0049] in, To pass the correction increment of the corresponding element in the parameter matrix, The preset learning rate is used as the parameter. This is the system bias metric (loss function value) at the current moment. This is the gradient (partial derivative) of the bias metric with respect to the connection weights.

[0050] The dual-layer shadow verification module performs physical conservation constraint verification on the corrected parameters. The physical conservation constraint verification logic determines whether the absolute value of the difference between the total mass input flux and the total mass output flux within any closed subgraph region is less than the physical measurement error limit.

[0051] The counterfactual reasoning competition module is configured to start when the physical conservation constraint verification fails. The counterfactual reasoning competition module generates a set of candidate mutation graphs based on a failure mode library. Each candidate mutation graph contains at least one non-isomorphic topological change relative to the baseline running graph, including the addition or deletion of directed edges.

[0052] The counterfactual reasoning competition module is configured to compute the posterior fitness score for each candidate mutation map in parallel:

[0053] ;

[0054] in, For the first The combined posterior fitness score of the candidate variant maps , , These are the preset weighting coefficients. The fitting residual (data consistency measure) between the predicted output vector of the candidate map and the sensor measured vector. The information entropy of the parameter matrix is ​​used to transfer mutations (to constrain the physical rationality of parameter distribution). The complexity cost of topological variation (structural sparsity penalty).

[0055] In this embodiment of the invention, the heterogeneous environment causal dependency graph is constructed as a directed acyclic graph structure. The node set is strictly divided into three disjoint subsets at the data structure level: the source node subset, the transformation node subset, and the sink node subset.

[0056] Each source node in the subset of source nodes corresponds to an initial pollutant generation unit in physical space. These initial pollutant generation units include chemical reactors, combustion boiler furnaces, raw material storage tanks, and open painting stations. In the heterogeneous environment causal dependency graph, source nodes only serve as the starting point of directed edges and do not accept any material flow input from other nodes.

[0057] Each source node is associated with a source strength attribute vector. This vector contains the initial generation concentration, initial generation flow rate, and operating condition correlation factors. The values ​​of the source strength attribute vector are statically assigned by the graph initialization module based on the data from the engineering analysis section of the environmental impact assessment document.

[0058] Each conversion node in the subset corresponds to a pollutant treatment facility or a physical transport channel in the physical space. Pollutant treatment facilities include desulfurization towers, bag filters, activated carbon adsorption boxes, and regenerative thermal oxidizers (RTOs). Physical transport channels include exhaust pipes, chimneys, and atmospheric diffusion spaces. In the heterogeneous environment causal dependency graph, a conversion node simultaneously serves as the endpoint and starting point of directed edges. A conversion node receives material flow input from upstream nodes and outputs processed material flow to downstream nodes.

[0059] Each conversion node has a transfer function attribute. For a conversion node corresponding to a pollutant treatment facility, the transfer function attribute is defined as a removal efficiency matrix, which describes the mass reduction ratio of different types of pollutants passing through the node. For a conversion node corresponding to a physical transport channel, the transfer function attribute is defined as the physical attenuation coefficient. Or dilution factor, physical attenuation coefficient describes the natural degradation process of pollutants over time or distance.

[0060] Each sink node in the subset corresponds to an environmental monitoring point or environmentally sensitive target in physical space. Environmental monitoring points include installation points of fixed pollution source online monitoring systems, monitoring points for fugitive emissions at plant boundaries, and regional ambient air quality monitoring stations. In the heterogeneous environmental causal dependency graph, sink nodes only serve as the endpoints of directed edges; sink nodes do not output material flows to downstream nodes.

[0061] Each sink node is associated with an observation attribute vector. This vector stores real-time sensor measurements uploaded by the multimodal data acquisition subsystem. The dimension of the observation attribute vector is consistent with the pollutant type dimension output by the source node.

[0062] Each node in the heterogeneous environment causal dependency graph is assigned a state vector with a uniform format. The state vector includes a timestamp field, a pollutant concentration field, a gas velocity field, and a temperature field. The state vector of the source node is determined by the source strength attribute vector, the state vector of the transition node is obtained by calculating the weighted sum of the input connection edges through the transfer function attribute, and the state vector of the sink node is used to compare the Euclidean distance with the observed attribute vector.

[0063] In this embodiment of the invention, the edge set in the heterogeneous environment causal dependency graph consists of a series of directed edges. Each directed edge uniquely corresponds to a material flow transport channel in physical space, pointing from an upstream node to a downstream node. The physical forms of the material flow transport channel include process pipes, flues, wastewater ditches, and atmospheric diffusion paths in open spaces.

[0064] Each directed edge in the data structure is associated with a transitive weight matrix. The transitive weight matrix is ​​defined as a... A square matrix of dimensions, in which This represents the total number of controlled pollutant types in the system. The transfer weight matrix is ​​used to quantify the mass flux changes during the transport of material through directed edges.

[0065] Diagonal elements in the pass weight matrix Indicates the first The transport efficiency coefficient of a pollutant from an upstream node to a downstream node. The range of the transport efficiency coefficient is a closed interval. When a directed edge corresponds to a leak-free, closed pipe, the transmission efficiency coefficient is 1; when a directed edge corresponds to a branch with physical attenuation or branching, the transmission efficiency coefficient is less than 1.

[0066] Transmit weight matrix off-diagonal elements (in, ) indicates the first The first pollutant is transformed into the second during the transport process. The conversion coefficient of pollutants. When no chemical reaction occurs in the physical transport channel, the off-diagonal elements... It was assigned the value 0.

[0067] The graph initialization module is configured to perform static initialization of the transfer weight matrix. It reads the engineering analysis data table from the digitized environmental impact assessment document and extracts the material balance data under the design conditions. The module then calculates the initial transfer weight matrix using the following formula. The values ​​of each element:

[0068] ;

[0069] in, The diagonal elements (initial transmission efficiency coefficients) in the initial transmission weight matrix. This represents the theoretical transmission efficiency of the transmission channel in the engineering design documents. This represents the flow allocation ratio from the upstream node to the downstream node. It is the physical attenuation factor of pollutants over the transmission distance.

[0070] The graph initialization construction module is configured to perform matrix assembly operations, combining the calculated diagonal elements with pre-assigned 0 off-diagonal elements to construct a complete initial transitive weight matrix. The module then writes this initial transitive weight matrix to non-volatile storage, serving as the heterogeneous environment causal dependency graph at time steps. The baseline parameter state at that time. This baseline parameter state constitutes the numerical starting point for subsequent dynamic correction and backpropagation based on real-time data.

[0071] In this embodiment of the invention, the multimodal data mapping module is configured to perform time alignment operations on multi-source heterogeneous data. The analog monitoring data and discrete control data uploaded by the multimodal data acquisition subsystem have a non-uniform distribution on the original time axis. The multimodal data mapping module sets a unified discrete time step for the system. The multimodal data mapping module uses a linear interpolation algorithm to resample the continuously changing analog monitoring data to discrete time points. The multimodal data mapping module uses a zero-order hold algorithm to align the state-transitioning discrete control data to discrete time points.

[0072] The multimodal data mapping module is configured to construct a complete set vector of node states for the entire system. The complete set vector of node states is defined as a... A numerical matrix of dimension, where This represents the total number of physical nodes in the causal dependency graph of heterogeneous environments. This represents the dimension of the state characteristics of a single node.

[0073] The state characteristic dimensions include pollutant mass concentration, gas volumetric flow rate, thermodynamic temperature, and hydrostatic pressure components. For intermediate nodes without physical sensors, the multimodal data mapping module marks their corresponding state components as states to be calculated. These states are then derived through the parent node during the forward computation phase.

[0074] The multimodal data mapping module is configured to perform a normalization mapping operation on each element of the original analog quantity in the complete set vector of node states, generating normalized state values. The normalization mapping operation is performed according to the following formula:

[0075] ;

[0076] in, These are the normalized values. These are the original input data values. This is the minimum value (or statistical lower limit) of this data dimension. This is the maximum value (or statistical upper limit) for this data dimension.

[0077] The multimodal data mapping module is configured to construct discrete event vectors. A discrete event vector is defined as a... A binary sparse vector of dimension, where This indicates the total number of controlled actuators in the system. Controlled actuators include process shut-off valves, flue gas bypass dampers, induced draft fan frequency converters, and chemical dosing pumps. The multimodal data mapping module internally stores a physical address mapping table. This table defines a unique correspondence between hardware tag numbers in the industrial control system and their index positions in discrete event vectors.

[0078] The multimodal data mapping module assigns values ​​to discrete event vectors based on the real-time feedback signals from the controlled actuators. When the controlled actuators are in an active, running, or powered-on state, the multimodal data mapping module assigns the value 1 to the corresponding index element in the discrete event vector. When the controlled actuators are in a closed, stopped, or powered-off state, the multimodal data mapping module assigns the value 0 to the corresponding index element in the discrete event vector. The discrete event vectors are directly passed to the dynamic topology evolution module as input variables to drive the binary logic transformation of the graph topology.

[0079] In this embodiment of the invention, the dynamic topology evolution module is configured to receive discrete event vectors from the multimodal data mapping module. The dynamic topology evolution module stores a predefined graph rewriting rule base in a non-volatile storage medium. The graph rewriting rule base contains several logical mapping entries, each defining a deterministic causal correspondence between a specific index position in the discrete event vector and a specific coordinate position in the adjacency matrix.

[0080] The dynamic topology evolution module is configured to perform topology evolution at each discrete time step. Perform a topology update operation. The topology update operation follows the state transition equation:

[0081] ;

[0082] in, This is the adjacency matrix (topological state) of the heterogeneous environment causal dependency graph at the current moment. It is the topological dynamic evolution function (or state transition function). Let be the graph adjacency matrix from the previous time step. This refers to the vector of external discrete events (or changes in control signals) received at the current moment. This is a pre-defined set of topology reconstruction rule constraints.

[0083] The logical mapping entries in the graph rewriting rule base are divided into connectivity enabling rules and connectivity blocking rules. For connectivity enabling rules, when the dynamic topology evolution module detects an index of […] in the discrete event vector… When an element's value changes from 0 to 1, the dynamic topology evolution module sets the value of the corresponding coordinate in the adjacency matrix to 1. (Coordinate index) and Corresponding to the index as The physical actuator controls the upstream and downstream nodes. For connectivity blocking rules, when the index in the discrete event vector is... When the element value changes from 1 to 0, the dynamic topology evolution module will change the corresponding coordinates in the adjacency matrix. The value is set to 0.

[0084] The dynamic topology evolution module is configured to perform mutually exclusive logic operations to handle bypass switching scenarios. When a discrete event vector indicates a bypass valve opening event, the dynamic topology evolution module simultaneously performs two atomic operations based on the mutually exclusive entries in the graph rewrite rule base: the first atomic operation sets the adjacency matrix element representing the bypass pipe to 1, and the second atomic operation sets the adjacency matrix element representing the main treatment facility channel to 0. The mutually exclusive logic operations ensure that the flow structure in the causal dependency graph of the heterogeneous environment remains strictly consistent with the physical fluid flow direction in the industrial site.

[0085] After traversing and executing all applicable entries in the graph rewriting rule base, the dynamic topology evolution module generates the baseline running graph for the current time step. The baseline running graph contains the updated adjacency matrix and the transfer parameter matrix inherited from the previous time step. The baseline running graph is then transferred to the forward computation and bias diagnosis module, serving as the sole structural basis for subsequent pollutant migration and transformation calculations.

[0086] In this embodiment of the invention, the forward computation and deviation diagnosis module is configured to perform a step-by-step deduction operation of node states under the topological constraints of the baseline running graph. The forward computation and deviation diagnosis module first performs a topological sorting algorithm on all nodes in the baseline running graph to generate a linear computation sequence that satisfies causal dependencies. The linear computation sequence ensures that for any node in the sequence, all its parent nodes are arranged before that node.

[0087] The forward computation and deviation diagnosis module calculates the predicted output state vector of each downstream node at the current time step according to the linear computation sequence. For nodes in the source node subset, the predicted output state vector is directly assigned the source strong attribute vector. For nodes in the transformation node subset and sink node subset, the forward computation and deviation diagnosis module performs calculations according to the following state transfer equation:

[0088] ;

[0089] in, Output the state value (predicted value) of the target node at the time step. This is the activation response function for this node (used to describe the characteristics of nonlinear physicochemical reactions). The set of all upstream parent nodes pointing to a node. The transmission weight parameter (representing transmission efficiency or influence coefficient) is used to transfer data from the upstream node to the target node. This represents the output value of the upstream parent node at the current moment. This is the node bias term of the target node at the current moment (representing the intensity of the node's own independent emission sources or environmental background value).

[0090] For a node representing physical pipeline transmission, the physical transformation activation function is defined as a first-order dynamic decay function:

[0091] ;

[0092] in, Let be the activation response function of the node (specifically represented here as an exponential decay model following first-order dynamics). This is the aggregated input value for this node (i.e., the total amount of pollutants entering this node). This represents the reaction rate constant (or biochemical attenuation coefficient) corresponding to the node. This refers to the hydraulic residence time of pollutants within the treatment unit at this node.

[0093] The forward computation and deviation diagnosis module is configured to extract the state vectors corresponding to the sink node subset from the computation results of all nodes, and combine them to generate the predicted output vector for the entire system. Simultaneously, the forward computation and deviation diagnosis module obtains the sensor measurement vectors at the current moment from the multimodal data mapping module.

[0094] The forward computation and deviation diagnosis module is configured to calculate the deviation metric between the predicted output vector of the entire system and the measured vector from the sensors. The deviation metric is calculated using the mean square error formula:

[0095] ;

[0096] in, This is the overall deviation metric of the system at the current moment (usually represented by the mean squared error loss function). This refers to the total number of valid monitoring nodes (or the total number of dimensions of the output data) participating in error calculation. For the first The model prediction output value of each monitoring node at time t. For the first The actual measured values ​​of the sensors collected by each monitoring node at any given time.

[0097] The forward calculation and deviation diagnosis module will calculate the deviation metric value. The deviation metric is compared numerically with a preset system tolerance threshold. When the deviation metric is less than or equal to the system tolerance threshold, the forward calculation and deviation diagnosis module determines that the current system is in a steady state and does not trigger the subsequent correction process. When the deviation metric is greater than the system tolerance threshold, the forward calculation and deviation diagnosis module generates an anomaly trigger signal, which includes the current baseline running pattern and the deviation metric, and sends the anomaly trigger signal to the two-layer shadow verification module.

[0098] In this embodiment of the invention, the dual-layer shadow verification module is configured to initiate the first-level shadow verification process upon receiving an anomaly trigger signal from the forward computation and deviation diagnosis module. The dual-layer shadow verification module instantiates a first-level shadow graph object in an address space independent of the main runtime memory. The first-level shadow graph object has a node set, edge set, and initial transfer parameter matrix that are completely identical to the current time-based baseline runtime graph.

[0099] The two-layer shadow verification module is configured to perform parametric gradient backpropagation in the first-level shadow graph object. The module calculates the partial derivatives of the bias metric with respect to each element in the transfer parameter matrix using the chain rule. It then constructs the gradient matrix using these partial derivatives and calculates the corrected transfer parameter matrix according to the following parameter update formula.

[0100] ;

[0101] in, The passed weight parameters are updated based on gradient descent. The current passed weight parameter value before the update. The preset learning rate (a coefficient used to control the step size of parameter adjustment). This is the system bias metric (loss function) at the current moment. This is the gradient (partial derivative) of the deviation metric relative to the transferred weight parameter.

[0102] The dual-layer shadow verification module is configured to perform physical conservation constraint verification on the modified parameter set. The module iterates through each transformation node in the first-level shadow map object and establishes the mass balance equation for that node. The module calculates the mass imbalance residual of the node according to the following formula. :

[0103] ;

[0104] in, For the first The mass balance residual value at each physical node (used to measure whether the input and output of that node violate the law of conservation of mass). Pointing to a node The set of all upstream parent nodes, This is the updated propagation weight parameter from the upstream node to the current node (i.e., inflow efficiency). This is the output source strength value (or status output value) of the upstream node. For nodes The set of all downstream child nodes pointed to. For the current node The output status value.

[0105] The dual-layer shadow verification module compares the calculated mass imbalance residual with a preset physical tolerance limit. The physical tolerance limit is a constant value determined based on the sensor's measurement accuracy level.

[0106] When the mass imbalance residuals of all nodes in the first-level shadow map object are less than or equal to the physical allowable error limit, the two-layer shadow verification module determines that the current system deviation is caused by parameter drift. In this case, the two-layer shadow verification module will synchronously update the corrected parameter set to the baseline operating map of the main system and terminate the verification process.

[0107] When the residual mass imbalance of any node in the first-level shadow map object exceeds the physical allowable error limit, the two-layer shadow verification module determines that simple parameter adjustments cannot satisfy the physical conservation law, and the system has a structural anomaly not captured by the explicit topology. In this case, the two-layer shadow verification module generates a structural variation trigger signal and sends this signal to the counterfactual reasoning competition module to initiate the second-level verification process.

[0108] In this embodiment of the invention, the counterfactual reasoning competition module is configured to start the counterfactual reasoning engine after receiving a structural variation trigger signal sent by the two-layer shadow verification module. The counterfactual reasoning engine calls a pre-set fault mode topology library in a non-volatile storage medium. The fault mode topology library stores non-isomorphic topology transformation rule templates defined based on physical laws, including implicit leakage mode templates and unexpected bypass mode templates.

[0109] The counterfactual reasoning competition module is configured to generate a graph based on a benchmark running graph that includes... A set of candidate mutation maps for each element. Each candidate mutation map in the set is a counterfactual structural variant of the baseline running map. The counterfactual reasoning competition module generates candidate mutation maps by applying mutation operators:

[0110] ;

[0111] in, For the first generation generated based on the current base map Candidate variant maps. For the preset topological mutation operator, This is the baseline spectrogram structure before the mutation operation. For the first The specific mutation pattern used in this mutation. These are the control parameters corresponding to the mutation mode.

[0112] When the failure mode type is latent leakage mode, the counterfactual reasoning competition module executes the edge-increasing operation operator. The edge-increasing operation operator selects a non-sink node in the graph as the leakage source and selects either an environmental acceptor node or a physical environment node as the leakage sink. The edge-increasing operation operator corrects the element values ​​of the corresponding coordinates in the mutated adjacency matrix from 0 to 1.

[0113] ;

[0114] in, For the first Nodes in the adjacency matrix of candidate mutation maps With nodes The connection status values ​​between them To update the state value at this position to 1, The selection criteria for the mutation location are that the operation is performed only when the corresponding position in the baseline spectrum is 0 (originally no connection).

[0115] This operation creates a virtual directed edge at the mathematical graph theory level, which leads directly from the process node to the environmental emission node. This virtual directed edge at the physical level simulates a pipe rupture, seal failure, or illegal underground discharge channel.

[0116] When the failure mode type is an unexpected bypass mode, the counterfactual reasoning competition module executes the reconnection operator. The reconnection operator identifies a governance facility node in the graph and locates the upstream node pointing to that node and the downstream node that node points to. The reconnection operator modifies the mutated adjacency matrix, cutting off the normal path through the governance facility and establishing a short-circuit path directly across the governance facility.

[0117] ;

[0118] in, For the first The connection state values ​​from the input node to the processing node in the adjacency matrix of the mutation graph. To block this path, This represents the connection state value from the input node to the output node in the matrix. To establish this path connection.

[0119] This operation removes the interception and processing function of the treatment facility nodes on the material flow in the topology structure, and simulates the failure scenario of direct discharge of flue gas or complete shutdown of treatment facilities caused by internal leakage of valves at the physical level.

[0120] The counterfactual reasoning competition module is configured to perform spatial pruning on the candidate variant map set. This module reads Geographic Information System (GIS) data or 3D digital model data of the factory and calculates the Euclidean physical distance between upstream and downstream nodes. The module removes candidate variant maps whose physical distance exceeds a preset physical connectivity limit from the variant map set to ensure the physical plausibility of the generated counterfactual assumptions.

[0121] The selected candidate mutation maps are stored in memory as the input space for subsequent multi-hypothesis entropy reduction competition calculations. Each candidate mutation map independently maintains a private set of transfer parameter matrices, whose initial values ​​are inherited from the baseline running map and can be independently fine-tuned in subsequent competition steps.

[0122] In this embodiment of the invention, the counterfactual reasoning competition module is configured to perform parallel fitness evaluation calculations on each candidate map in the filtered candidate mutation map set. The counterfactual reasoning competition module allocates an independent computation thread to each candidate map and performs fast parameter fine-tuning based on gradient descent to make it fit the measured data as closely as possible under the current mutation topology.

[0123] The counterfactual reasoning competition module is configured to calculate the comprehensive posterior fitness score for each fine-tuned candidate graph. The comprehensive posterior fitness score is a weighted sum of the fitting residual term, the parameter entropy reduction term, and the structural complexity term, calculated as follows:

[0124] ;

[0125] in, For the first The combined posterior fitness score of the candidate variant maps To fit the residual term, This is a term to reduce the parameter entropy. For structural complexity, , , These are the weighting coefficients.

[0126] The fitted residual term quantifies the first The hypothetical graph's interpretability of the current observation data is used. The counterfactual reasoning competition module calculates the L2 norm distance between the predicted output vector and the observed vector of the fine-tuned graph:

[0127] ;

[0128] in, To fit the residual term, For the first The full system prediction output vector generated from the candidate mutation maps after parameter fine-tuning. This represents the sensor measurement vector uploaded by the multimodal data acquisition subsystem at the current moment. The L2 norm distance (Euclidean distance) operator between vectors is used to quantify the overall deviation between predicted and measured values ​​in multidimensional space.

[0129] The parameter entropy reduction term quantifies the naturalness of model parameter adjustments, based on the inverse application of the maximum entropy principle (entropy reduction principle), i.e., prioritizing the model with the smallest parameter correction magnitude. The counterfactual reasoning competition module calculates the relative entropy or Frobenius norm between the fine-tuned parameter matrix and the initial design parameter matrix:

[0130] ;

[0131] in, This is a term to reduce the parameter entropy. For the first The transfer parameter matrix of each candidate mutation map after gradient descent fine-tuning. The parameter matrix is ​​used to transfer the design parameters to the initial state of the system, which is a set of reference parameters that conforms to engineering design or physical common sense. The Frobenius norm operator for matrices is used to calculate the overall difference between the fine-tuned parameters and the initial design parameters.

[0132] The structural complexity term quantifies the cost of structural variation of the counterfactual assumption, based on Occam's razor, which prioritizes the interpretation with the least structural modification. The counterfactual reasoning competition module calculates the Hamming distance between the mutated adjacency matrix and the baseline adjacency matrix, i.e., the total number of binary elements that undergo flipping.

[0133] ;

[0134] in, For structural complexity, For the first The mutation adjacency matrix corresponding to each candidate mutation map is in the coordinate system. The element value at a given location represents the node connectivity state under the counterfactual assumption. The Hamming distance calculation operation between two adjacency matrices is to count the total number of directed edges in the entire graph that have undergone connection state flips (from disconnected to connected or from connected to disconnected).

[0135] The counterfactual reasoning competition module is configured to perform a global optimization competition operation. This module iterates through the comprehensive posterior fitness scores of all candidate graphs and identifies the index with the lowest score.

[0136] ;

[0137] in, The index is the unique identifier that is determined to be the optimal interpretation map after global selection. The variant map corresponding to this index is the final fault diagnosis conclusion output by the system. The parameter optimization operator that minimizes the objective function represents traversing all possible candidate indices. From this, we select the specific index that minimizes the function value. For the first The overall posterior fitness score of each candidate variant map indicates that the lower the score, the better the corresponding counterfactual hypothesis performs in terms of fitting accuracy, parameter rationality, and structural simplicity (i.e., the lowest cost).

[0138] The counterfactual reasoning competition module determines the candidate variant graph corresponding to the index as the optimal explanatory graph. The counterfactual reasoning competition module calculates the set of topological differences between the optimal explanatory graph and the baseline running graph:

[0139] ;

[0140] in, This is a set of topological differences used to instantiate latent fault event records. It contains all directed edges that have changed relative to the baseline state under the optimal counterfactual hypothesis. For the node Pointing to node The directed edge coordinate tuples represent specific physical connection paths in the causal dependency graph of heterogeneous environments. The variant graph determined as the optimal interpretation is located in the adjacency matrix coordinates. The element value at that location represents the connectivity under fault conditions. The baseline running graph in the adjacency matrix coordinates The element value at that location represents the connectivity under normal operating conditions.

[0141] The counterfactual reasoning contest module instantiates the difference set as a hidden fault event record. If the difference set contains elements... and Counterfactual reasoning competition module generates nodes To node The diagnosis indicates the presence of a latent leak. If the difference set contains elements... and Counterfactual reasoning competition module generates nodes To node The diagnostic conclusion is that the transmission path is blocked.

[0142] The counterfactual reasoning competition module sends a data packet containing the diagnostic conclusion, the optimal explanatory graph, and the corresponding confidence score to the visual interactive terminal for display, completing a full counterfactual reasoning closed loop.

Claims

1. A full-process digital intelligent analysis system for environmental impact assessment information, characterized in that, include: The end-to-end analysis server is configured as the core computing device of the system. The multimodal data acquisition subsystem is configured to acquire analog monitoring data and discrete control data from the industrial production site and send them to the full-process analysis server via an industrial communication network. A visual interactive terminal is configured to receive and display the analysis results generated by the full-process analysis server; The full-process analysis server includes a central computing processor and a non-volatile storage medium, which stores computer program instructions and heterogeneous environment causal dependency graph data structures. The central computing processor is configured to execute the computer program instructions to run the following modules: A multimodal data mapping module is used to map the analog quantity monitoring data into node state vectors and the discrete quantity control data into discrete event vectors. The dynamic topology evolution module is used to update the topology of the heterogeneous environment causal dependency graph based on the discrete event vectors and a preset graph rewriting rule base, and generate a benchmark running graph. The forward calculation and deviation diagnosis module is used to perform pollutant migration and transformation calculations in the baseline operation spectrum and calculate the systematic deviation metric between the predicted output vector and the measured vector. The dual-layer shadow verification module is used to instantiate a shadow map copy in an independent memory space and perform parameter gradient backpropagation calculation and physical conservation constraint verification when the system deviation metric exceeds a preset threshold. The counterfactual reasoning competition module is used to generate a set of candidate mutation maps based on the fault mode topology library when the physical conservation constraint verification fails, and to select the optimal interpretation map by calculating the comprehensive posterior fitness score to generate a hidden fault diagnosis conclusion. The dual-layer shadow verification module is configured as follows: Create a copy of the baseline running graph as a first-level shadow graph; Calculate the gradient of the system deviation metric with respect to the transfer parameter matrix, and correct the transfer parameter matrix in the first-level shadow map based on the gradient descent method; Perform the physical conservation constraint check on the corrected first-level shadow map and calculate the mass imbalance residual of the conversion node; If the mass imbalance residual is greater than the physical allowable error limit, the verification is determined to be unsuccessful and a structural variation trigger signal is generated and sent to the counterfactual reasoning competition module. The counterfactual reasoning competition module is configured as follows: In response to the structural variation trigger signal, the fault mode topology library is invoked, which stores hidden leakage mode templates and unexpected bypass mode templates. The adjacency matrix of the baseline running graph is modified nonlinearly by applying the mutation operator to generate the candidate mutation graph set; Specifically, for the latent leakage mode template, the edge-increasing operation operator is executed to change the element values ​​of the corresponding coordinates between non-sink nodes and environmental receptor nodes in the adjacency matrix from 0 to 1. For the unexpected bypass mode template, the reconnection operation operator is executed to block the path pointing to the governance facility node in the adjacency matrix and establish a short-circuit path across the governance facility node.

2. The full-process digital intelligent analysis system for environmental impact assessment information according to claim 1, characterized in that, The multimodal data acquisition subsystem includes: The analog quantity monitoring sensor group includes continuous monitoring instruments installed at pollution generation nodes, treatment nodes, and emission nodes for collecting the analog quantity monitoring data; The discrete quantity controller interface unit is connected to an industrial distributed control system or a programmable logic controller to collect the switching status signals of field actuators as the discrete quantity control data.

3. The full-process digital intelligent analysis system for environmental impact assessment information according to claim 1, characterized in that, The heterogeneous environment causal dependency graph data structure is defined as a directed acyclic graph, which includes: a set of nodes, divided into a subset of source nodes, a subset of transformation nodes, and a subset of sink nodes; Wherein, the source node subset corresponds to the initial pollutant generation unit, the conversion node subset corresponds to the pollutant treatment facility or physical transmission channel, and the sink node subset corresponds to the environmental monitoring point; The set of edges, each of which is associated with a transfer parameter matrix, stores the transmission efficiency coefficients of material flow between physical nodes.

4. The full-process digital intelligent analysis system for environmental impact assessment information according to claim 1, characterized in that, The dynamic topology evolution module is configured as follows: At each discrete time step, a topology update operation is performed, and based on the changes in the element values ​​in the discrete event vector, the connectivity enabling rules or connectivity blocking rules in the graph rewrite rule base are applied. The connectivity enabling rule is configured such that when the index element in the discrete event vector changes from 0 to 1, the value of the corresponding coordinate in the adjacency matrix of the baseline running graph is set to 1. The connectivity blocking rule is configured such that when the index element in the discrete event vector changes from 1 to 0, the value of the corresponding coordinate in the adjacency matrix of the baseline running graph is set to 0.

5. The full-process digital intelligent analysis system for environmental impact assessment information according to claim 1, characterized in that, The forward calculation and deviation diagnosis module is configured as follows: Perform topological sorting on the baseline running graph to generate a linear computation sequence; The predicted output state vector of the downstream node is calculated based on the state transfer equation, which includes the weighted sum of the output state of the upstream parent node and the transfer parameter matrix, as well as the node bias term. The mean square error between the predicted output vector of the entire system corresponding to the subset of sink nodes and the measured vector of the sensor is calculated as the system deviation metric.

6. The full-process digital intelligent analysis system for environmental impact assessment information according to claim 1, characterized in that, The counterfactual reasoning competition module is also configured to perform spatial pruning operations: Read data from geographic information systems or 3D digital models of factories; Calculate the Euclidean physical distance between nodes; Candidate variant maps whose Euclidean physical distance is greater than the preset physical connection limit distance are removed from the candidate variant map set.

7. The full-process digital intelligent analysis system for environmental impact assessment information according to claim 1, characterized in that, The comprehensive posterior fitness score calculated by the counterfactual reasoning competition module is composed of a weighted sum of the fitting residual term, the parameter entropy reduction term, and the structural complexity term. The fitting residual term is the norm distance between the predicted output vector and the measured vector of the candidate mutation map; The parameter entropy reduction term is the magnitude of the difference between the fine-tuned transfer parameter matrix and the initial design parameter matrix. The structural complexity term is the Hamming distance between the adjacency matrix of the candidate mutation graph and the adjacency matrix of the baseline running graph.

8. The full-process digital intelligent analysis system for environmental impact assessment information according to claim 7, characterized in that, The counterfactual reasoning competition module is further configured as follows: The candidate variant map with the smallest comprehensive posterior fitness score is selected as the optimal interpretation map. Calculate the set of topological differences between the optimal interpretation map and the baseline running map; The set of topological differences is instantiated as a record of latent fault events, and then marked at the corresponding coordinates in the visualized image via the visual interactive terminal.

Images