Chemical P&ID process drawing intelligent analysis and change management method based on image processing

Abstract

This invention relates to the field of intelligent processing of chemical process diagrams and industrial image recognition technology, specifically to an intelligent parsing and change management method for chemical process diagrams (P&ID) based on image processing. The method includes: acquiring a two-dimensional degraded document image containing unstructured modification traces; extracting underlying image connectivity features and character features; inputting a graph neural network to reconstruct an initial logical topology network; dividing local regions and detecting logical conflicts using deterministic topology rules, quantifying the probability value of topological logical conflicts; comparing with a preset danger threshold, actively severing logical connections in high-risk areas and marking them as isolated sub-graph regions awaiting manual review, while retaining connectivity paths in low-risk areas to generate a global topology network; updating the global topology network based on manual review feedback, and outputting the final document drawing change parsing version. This invention significantly improves the recognition stability of complex on-site drawings.

Description

Technical Field

[0001] The invention relates to the field of intelligent processing of chemical process diagrams and industrial image recognition technology, specifically a method for intelligent parsing and change management of chemical P&ID process drawings based on image processing. Background Technology

[0002] With the increasing demand for digital operation and maintenance and overhaul and renovation management of petrochemical plants, intelligent analysis and change management of chemical pipelines and instrumentation diagrams has become an important research direction in the field of drawing management. In order to achieve rapid identification, structured extraction and version management of on-site modification information, automatic understanding and topology reconstruction analysis of drawing images have become particularly important. In the process of analyzing chemical pipeline and instrumentation diagrams, image processing, character recognition, and topology reasoning technologies can extract equipment symbols, pipeline relationships, and text annotations to a certain extent. However, the drawings used on-site often have issues such as manual red line modifications, oil stains, creases, breaks, cross-page line additions, and photographic distortion. This leads to unstable element recognition, misjudgment of connectivity relationships, and frequent local logical conflicts. Especially in scenarios such as unit shutdown and maintenance, commissioning and switching, and temporary venting, if the connection relationships of the entire diagram are directly generated based solely on image recognition results, local errors can easily spread to errors throughout the diagram, thereby affecting the use of subsequent control systems, version control system updates, and process safety verification. Therefore, effectively processing two-dimensional degraded document images containing unstructured modification traces, reconstructing a verifiable initial logical topology network of drawings, and combining process material flow restrictions and valve normally open and normally closed logic to identify, isolate, and manually verify high-risk areas, thereby outputting reliable structured drawing change results, is crucial to ensuring the accuracy, safety, and traceability of drawing change management in chemical plants. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent parsing and change management method for chemical P&ID process drawings based on image processing, and to solve the following technical problems: It avoids local misjudgments from spreading to whole-drawing errors in noisy or heavily hand-drawn interference scenarios, and makes it easier to ensure the security boundary of the global topology network and form a closed loop for drawing change management.

[0004] The objective of this invention can be achieved through the following technical solutions: The method for intelligent parsing and change management of chemical P&ID process drawings based on image processing is characterized by including: acquiring a two-dimensional degraded document image containing unstructured modification traces, and extracting the underlying image connectivity features and character features of the two-dimensional degraded document image using a multimodal visual model and optical character recognition technology. The underlying image connectivity features and character features are input into a graph neural network for feature aggregation to reconstruct the initial logical topology network of the two-dimensional degraded document image; The initial logical topology network is divided into subgraphs based on a preset graph segmentation algorithm to obtain multiple local regions; The connected paths in the initial logical topology network are input into the graph theory rule engine, and logical conflict nodes that violate the physical conservation law are detected by combining the preset deterministic topology rules. Based on the detection results, the topological logical conflict probability value of each local region is quantified. The topological logical conflict probability value represents the probability of a logical paradox between image semantic understanding and physical rules. The probability value of the topological logical conflict is compared with a preset danger threshold that characterizes the tolerance of logical errors. If the probability value of the topological logical conflict is not less than the preset danger threshold, the prediction of the connectivity between the primitive nodes in the local area is abandoned, the logical connectivity of the local area in the initial logical topology network is actively cut off, and the local area is marked as an isolated subgraph area to be manually reviewed. If the probability value of the topological logic conflict is less than the preset danger threshold, the connectivity path of the local area is retained, and a global topology network is generated. Based on the isolated subgraph area to be manually reviewed, a manual review defense line is triggered. The corrected connectivity relationship of the isolated subgraph area to be manually reviewed is received as the manual review feedback result. The global topology network is updated in combination with the manual review feedback result, and the final document drawing change parsing version is output, forming a closed loop for drawing change management.

[0005] Preferably, the acquisition of a two-dimensional degraded document image containing unstructured modification traces, and the extraction of the underlying image connectivity features and character features of the two-dimensional degraded document image using a multimodal visual model and optical character recognition technology, includes: Obtain a scanned copy of the process drawing with hand-drawn change annotations, degraded noise, and cross-page lines as the two-dimensional degraded document image; The multimodal visual model is used to denoise the two-dimensional degraded document image and identify primitive node features and spatial position relationships. Text attribute features are extracted using the optical character recognition technology, and the primitive node features, spatial position relationships and text attribute features are fused into the underlying image connectivity features and character features.

[0006] Preferably, the process of inputting the underlying image connectivity features and character features into a graph neural network for feature aggregation to reconstruct the initial logical topology network of the document includes: Construct a graph structure; wherein the nodes of the graph structure include the graph primitive node features and the text attribute features, and the edges of the graph structure represent the connectivity assumptions derived from the spatial location relationships; perform deep learning on the graph structure using multi-layer graph convolution operations to aggregate neighbor node information; predict the connectivity probability between nodes, and generate the initial logical topology network based on the connectivity probability.

[0007] Preferably, quantifying the topological logic conflict probability value of the local region based on the detection results includes: The degradation noise intensity of the acquired two-dimensional degraded document image is evaluated, and the topological logical conflict probability value of the local region is calculated based on the density of the logical conflict nodes, the degradation noise intensity, and the connectivity probability output by the graph neural network.

[0008] Preferably, calculating the topological logical conflict probability value of the local region based on the density of the logically conflicting nodes, the intensity of the degraded noise in the two-dimensional degraded document image, and the connectivity probability output by the graph neural network includes: After normalizing the density of the logically conflicting nodes and the intensity of the degraded noise respectively, the nodes are weighted and summed according to preset weights to obtain the conflict base number. Divide the conflict base number by the connectivity probability to obtain the comprehensive risk index; normalize the comprehensive risk index to output the topological logic conflict probability value.

[0009] Preferably, actively severing the logical connectivity of the local area in the initial logical topology network and marking the local area as an isolated subgraph area to be manually reviewed includes: blocking all undetermined connectivity paths within the local area from extending outward; The node states within the local area are reset to an undefined state; a defensive isolation boundary is generated around the local area in the initial logical topology network to prevent logical errors from cascading outwards; the area within the defensive isolation boundary is encapsulated as the isolated subgraph region to be manually reviewed.

[0010] Preferably, if the probability value of the topological logic conflict is less than the preset danger threshold, then retaining the connected paths of the local area and generating a global topological network includes: verifying that the connected paths within the local area conform to the deterministic topological rules; The verified connectivity paths are solidified into the initial logical topology network; all verified regions are merged to dynamically generate the global topology network that reflects the actual physical connectivity relationships.

[0011] Preferably, triggering the manual review defense based on the isolated sub-graph area to be manually reviewed, updating the global topology network in combination with the feedback results of manual review, and outputting the final document drawing change parsing version includes: pushing alarm information containing the isolated sub-graph area to the preset review end, as well as image slices of the corresponding local area in the two-dimensional degraded document image; Receive manually inputted corrected connectivity relationships; replace the isolated subgraph regions to be manually reviewed with the corrected connectivity relationships, and recalculate the overall topology state; convert the updated global topology network into structured change data, and output the final document drawing change parsing version.

[0012] Preferably, the two-dimensional degraded document image is a scanned copy of a chemical pipeline and instrumentation diagram; the unstructured modification traces are manual changes and annotations made on-site. The deterministic topology rules include chemical material flow restriction rules and valve normally open and normally closed logic rules.

[0013] The beneficial effects of this invention are: 1. This invention utilizes multimodal visual models and optical character recognition technology to extract features, and inputs them into a graph neural network for feature aggregation. This method effectively overcomes visual interference caused by degraded images such as hand-drawn modifications and oil stains, and merges scattered primitives and text to construct a more accurate initial logical topology network, significantly improving the recognition stability of complex on-site drawings. 2. This invention performs subgraph partitioning on the initial network and inputs the connected paths into a graph theory rule engine. It then combines deterministic topology rules to detect conflicting nodes and quantifies the probability value of topological logical conflicts. This design deeply integrates visual recognition with the physical rules of chemical processes, achieving accurate quantitative evaluation of potential logical paradoxes between image semantic understanding and real processes. 3. This invention compares the conflict probability value with the danger threshold, actively abandons the prediction of high-risk local areas and cuts off their logical connectivity in the initial network; this active isolation mechanism generates a defensive boundary, effectively preventing uncertain errors caused by distortion or illegal modification in local areas from cascading outwards, and ensuring the security boundary of the global topology. 4. This invention retains paths in low-risk areas to generate a global network and triggers manual review defenses for isolated subgraphs. This mechanism enables the system to automatically allow access to trusted areas for network integration, while focusing human effort precisely on high-risk localities. Combined with review feedback, the network is dynamically updated, ultimately outputting a highly accurate structured change resolution version, forming a secure and efficient management closed loop. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating the intelligent parsing and change management method for chemical P&ID process drawings based on image processing provided in this application embodiment. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 A method for intelligent parsing and change management of chemical P&ID process drawings based on image processing, including: acquiring a two-dimensional degraded document image containing unstructured modification traces, and extracting the underlying image connectivity features and character features of the two-dimensional degraded document image using a multimodal visual model and optical character recognition technology; The underlying image connectivity features and character features are input into a graph neural network for feature aggregation to reconstruct the initial logical topology network of the two-dimensional degraded document image. The initial logical topology network is divided into subgraphs based on a preset graph segmentation algorithm to obtain multiple local regions. The connected paths in the initial logical topology network are input into a graph theory rule engine, and logical conflict nodes that violate physical conservation laws are detected by combining preset deterministic topology rules. The topological logical conflict probability value of each local region is quantified based on the detection results. The topological logic conflict probability value represents the probability of a logical paradox between image semantic understanding and physical rules. The topological logic conflict probability value is compared with a preset danger threshold that characterizes the tolerance for logical errors. If the topological logic conflict probability value is not less than the preset danger threshold, the prediction of the connectivity between primitive nodes in the local area is abandoned, the logical connectivity of the local area in the initial logical topology network is actively cut off, and the local area is marked as an isolated subgraph area to be manually reviewed. If the probability value of the topological logic conflict is less than the preset danger threshold, the connectivity path of the local area is retained, and a global topology network is generated. Based on the isolated subgraph area to be manually reviewed, a manual review defense line is triggered. The corrected connectivity relationship of the isolated subgraph area to be manually reviewed is received as the manual review feedback result. The global topology network is updated in combination with the manual review feedback result, and the final document drawing change parsing version is output, forming a closed loop for drawing change management.

[0018] This embodiment provides an intelligent drawing parsing and change management mechanism for large-scale shutdown and maintenance of petrochemical plants. Specifically, the main application scenario is set as follows: During a 72-hour shutdown and maintenance of an ethylene cracking unit, on-site engineers continuously upload P&ID scan images with hand-drawn red lines, oil stains, creases, and distortion from mobile phone photography. The system needs to first reconstruct a verifiable initial logical topology network without completely relying on the image inference results, then actively downgrade and isolate high-risk areas, and output a structured change version after manual confirmation for use by the subsequent control system and version library. In practice, a two-dimensional degraded document image can be divided into multiple image blocks, such as region A, region B, and region C. Two types of basic information can be extracted from each region: one is the underlying image connectivity features, which reflect whether line segments are continuous, whether endpoints are connected, whether arrows are pointing, and the positional relationship of primitives such as valves, pumps, and containers. Another type is character features, reflecting information such as pipeline numbers, valve tag numbers, medium flow direction descriptions, and handwritten modifications. For ease of explanation, we can assume that four nodes are identified in region A: node N1 is the pump, node N2 is the valve, node N3 is the pipeline number P-101, and node N4 is the heat exchanger. We can also initially obtain three candidate edges: E12 indicates that there is a connection between the pump and the valve, E24 indicates that there is a connection between the valve and the heat exchanger, and E34 indicates that there is a labeling relationship between the pipeline text and the heat exchanger. After constructing the above nodes and edges into a graph structure, the graph neural network is input into the graph neural network. The graph neural network aggregates the neighbor information and outputs the connectivity probability of each candidate edge. For example, the connectivity probability of E12 is 0.93, the connectivity probability of E24 is 0.88, and the home probability of E34 is 0.91. Based on this result, an initial logical topology network is generated. On this basis, graph segmentation is performed on the entire graph, which can be divided into local regions R1, R2, and R3 according to connectivity density, number of nodes, or page spatial location. Specifically, the graph segmentation algorithm can adopt spectral clustering algorithm, Louvain community detection algorithm, or nonnegative matrix factorization algorithm based on random walk, which can achieve high cohesion and low coupling of local regions by minimizing the sum of the weights of connected edges across subgraphs. Schematic diagram: R1 corresponds to the reaction feed branch, R2 corresponds to the reflux control branch, and R3 corresponds to the utility interface area. Next, the connected paths in each local region are fed into the graph theory rule engine; the rule engine does not rely directly on the appearance of the image, but checks whether the connection relationship of the region violates deterministic process rules; for example, a material flow path should not form a closed dead loop without a recycling device, normally closed valves should not be resolved as long-term open main channels, and materials should not flow in reverse without the presence of driving force elements. For ease of explanation, we can assume that in region R2, the system infers the path N5→N6→N7→N5 from the image, forming a closed loop. However, there is no circulation pump symbol in this region, and the OCR extracts a label next to the valve as NC, which means normally closed. Then the graph theory rule engine can determine that N6 or the corresponding edge E67 belongs to a logically conflicting node or an associated node of a conflicting edge. During the quantization phase, the system generates a topological logic conflict probability value for each local region. This probability value is not solely determined by the image recognition confidence level, but is derived by combining the logic conflict detection results, the degree of image degradation, and the topological prediction confidence level. Indicatively, the ratio of conflicting nodes to the total number of nodes in region R2 is 0.4, the image noise intensity assessment is 0.7, and the average confidence of the connected edges in this region by the graph neural network is 0.55. After normalization and weighting, the topological logic conflict probability value of R2 is 0.82. If the preset danger threshold is 0.75, then R2 is determined to be a high-risk region. For high-risk areas, the system does not continue to pursue automatic graph completion, but actively abandons further prediction of the connectivity between graph element nodes in the area and cuts off the logical connection between the area and the external network. For example, if R2 originally has a candidate edge between it and R1 and R3, the system freezes these two cross-regional candidate edges, no longer allowing the unconfirmed path of R2 to affect the flow direction judgment of other areas, and at the same time encapsulates R2 as an isolated subgraph area to be manually reviewed. Conversely, for areas below the danger threshold, such as R1 with a collision probability of only 0.21, their connectivity paths are preserved and incorporated into the global topology network. The system sends alarm information for the isolated area and corresponding image slices to the review end; after the review engineer corrects the valve status, disconnection attribution, or cross-page connection relationship in R2 on the terminal, the system uses the correction results to replace the isolated subgraph and recalculates the topology status of the entire graph. For example, if the engineer confirms that N6 is not a non-closed valve, but a new bypass valve is added, then the path can form a legal branch from N5→N8→N7, and the original closed-loop risk is eliminated; the system updates the global topology network accordingly and outputs the final document drawing change parsing version, which can be represented as a structured equipment connection table, change difference record and traceable version number; In an optional anomaly handling mechanism, when a local area in an image simultaneously exhibits three anomalies—character loss, numerous broken lines, and unclear cross-page indexes—if a stable set of nodes cannot be formed, the system can directly mark the area as an unmapped area, without entering the automatic topology reconstruction process, and instead submit it for review in the form of original image slices and a preliminary list of suspicious points. If the conflict probability value of a certain area is exactly equal to the danger threshold, it will be handled as a high-risk strategy to avoid accidental release at the boundary state; if the manual review does not return within the specified time, the system can maintain the isolation state and only output the change results of some trusted areas, and not allow high-risk areas to enter the official version repository. For example, during the overhaul of the aforementioned ethylene cracking unit, a photo taken by a mobile phone involving the outlet branch of the cracked gas compressor was uploaded on-site. The photo showed oblique reflections, and the original fine lines were covered by oil stains. The engineer then manually drew cross-page connecting lines on it. The system first identifies the compressor, shut-off valve, check valve, and branch to the quench tower from the image, and then constructs an initial topology. Because this branch is inferred to be an abnormal path for material to return to the upstream stage after the outlet check valve, and the image at this location has strong noise, it is ultimately designated as an isolated subgraph. The review engineer confirms on the image slice that the hand-drawn arrows actually represent the removal of the old line and the connection of the new line. Based on this, the system rewrites the connection relationship and then outputs the official change version. The purpose of this step is to transform the automatic identification and process safety from a sequential dependency relationship to a risk-layered relationship, so that the system can prioritize the safety boundary of the global topology network in high-noise and hand-drawn interference scenarios, and avoid the spread of local misjudgments into whole-map errors.

[0019] In a preferred embodiment of the present invention, obtaining a two-dimensional degraded document image containing unstructured modification traces and extracting the underlying image connectivity features and character features of the two-dimensional degraded document image using a multimodal visual model and optical character recognition technology includes: obtaining a scanned copy of a process drawing with hand-drawn change annotations, degraded noise, and cross-page lines as the two-dimensional degraded document image; The multimodal visual model is used to denoise the two-dimensional degraded document image and identify primitive node features and spatial position relationships. Text attribute features are extracted using the optical character recognition technology, and the primitive node features, spatial position relationships and text attribute features are fused into the underlying image connectivity features and character features.

[0020] This embodiment provides a processing step for front-end image acquisition and feature fusion. Specifically, in the above-mentioned device overhaul scenario, relying solely on simple binarization or traditional OCR often cannot stably handle problems such as perspective distortion caused by mobile phone photography, hand-drawn red lines covering the original black lines, and local occlusion caused by oil stains. Therefore, a multimodal visual model and character recognition collaborative mechanism are introduced at the input layer to form a unified feature expression suitable for subsequent image construction. Specifically, the two-dimensional degraded document images can come from scanners, industrial tablet camera terminals, or mobile phone upload terminals; after the images are accessed, they are preprocessed, including page correction, brightness normalization, edge cropping, and background cleaning; For ease of explanation, we can assume that the original image is a 3000×2000 pixel P&ID photo with a shadow in the upper left corner, a crease in the lower right corner, and a red hand-drawn arrow in the center crossing two original pipelines. The system can first perform corner detection on the page, correct the tilt angle from 7 degrees to within 1 degree, and then cut the image into several local windows, for example, each window is 512×512 pixels. The multimodal vision model simultaneously extracts shape, line type, color, and relative position features in each window; for example, it identifies an approximately triangular valve symbol, two collinear pipes broken in the middle, a red hand-drawn arrow, and blurry text next to it in window W1. The model does not process these contents separately, but outputs a combination of features including primitive category, location box, orientation vector and local connectivity score; schematically, the valve primitive can be assigned a category code V2, center coordinates of 120 and 86, and main direction of 0 degrees; the endpoint spacing of pipeline segments L1 and L2 is 8 pixels, but they are highly continuous in color and direction, so the connectivity candidate score can be 0.78. In the character recognition section, OCR can be performed on the original image and the enhanced image separately to improve the recall rate of weak contrast text. For example, when the original grayscale image and the enhanced high-contrast image are input to the same area at the same time, the former recognizes FV-102, and the latter recognizes the handwritten characters changed from normally closed to normally open. The system then attaches the text to the corresponding valve primitive based on the proximity of the position and the text line belonging. The text attribute features formed in this way not only include the character content, but also the text orientation, character box position, distance from the primitive, handwriting or printing type and other additional attributes. During the fusion stage, primitive node features, spatial position relationships, and text attribute features can be written into a unified record. The unified record is specifically represented as a high-dimensional feature tensor or a JSON / XML format node attribute table with structured fields. Spatial position relationships are represented by normalized center coordinates and direction vectors, and text attribute features are represented by word embedding vectors. The three are then concatenated into the underlying image connectivity features and character features after dimensional alignment through a multilayer perceptron or a fully connected layer. For example, in addition to retaining the valve category, node N2 is also associated with the string FV-102NO, as well as its spatial proximity to upstream and downstream pipelines; This can be simply illustrated as follows: if the primitive feature vector has a length of 8, the text feature vector has a length of 6, and the positional relationship vector has a length of 4, then the fusion will result in a node representation of length 18, which can be directly used for the subsequent graph construction network. In an optional exception handling mechanism, if the text recognition in a certain window is empty, but the boundary of the graphic element is clear, the system still retains the graphic element and spatial relationship and assigns a default label to the text; if the confidence of graphic element recognition is low but the text is clear, for example, only P-201A is recognized but the pump symbol is not stably recognized, the system can first construct a text node and attach it to the nearby line segment, and wait for the subsequent mapping stage to supplement the category through neighborhood information. If the handwritten red line completely overlaps with the original black line, the model cannot determine which line is the effective modification line. In this case, the conflict marker is written into the node attribute for the subsequent rule engine to check. For example, in the compressor outlet drawing of the pyrolysis unit, a mobile phone photo simultaneously contains hand-drawn text "→ changed to flare" and original printed text "→ quench tower". The multimodal visual model identifies two sets of arrows in opposite directions and two sets of text at the same location. OCR extracts the flare and quench tower respectively. Then, based on color layering, it is determined that the red hand-drawn arrow is the newly added modification, and the black printed arrow is the original design flow direction. After fusion, the node attributes retain both the original flow direction and the changed flow direction semantics for selection during subsequent topology reconstruction. The purpose of this step is to encode geometric relationships, textual semantics, and modification traces in the image into computable features, thereby reducing the reliance of subsequent mapping on a single visual channel.

[0021] In a preferred embodiment of the present invention, inputting the underlying image connectivity features and character features into a graph neural network for feature aggregation to reconstruct the initial logical topology network of the document includes: constructing a graph structure; wherein, the nodes of the graph structure contain the graph primitive node features and the text attribute features, and the edges of the graph structure represent the connectivity assumptions derived from the spatial positional relationships; The graph structure is subjected to deep learning using multi-layer graph convolution operations to aggregate neighbor node information; the connectivity probability between nodes is predicted, and the initial logical topology network is generated based on the connectivity probability.

[0022] This embodiment provides a reconstruction step for an initial logical topology network. Specifically, after the aforementioned feature extraction is completed, if lines are directly connected based solely on pixel contact relationships, misconnections are likely to occur in scenarios involving broken lines, page spans, overlapping drawings, and handwritten overlays. Therefore, in this embodiment, the drawing is first understood as a candidate relationship graph, and then the graph neural network filters each candidate relationship based on the neighborhood context. Specifically, when constructing a graph structure, nodes include at least equipment nodes, instrument nodes, valve nodes, pipeline nodes, and text nodes; edges do not initially represent actual connectivity, but rather represent an assumption of possible connectivity. For example, if a valve symbol is 3 pixels away from the left line segment and in the same direction, a candidate edge is created; if a text is within 10 pixels of the end of a road arrow, a belonging edge is created. For ease of explanation, we can assume a small graph with 6 nodes: N1 is the compressor, N2 is the outlet pipeline, N3 is the shut-off valve, N4 is the branch arrow, N5 is the text to the torch, and N6 is the text to the quench tower; the candidate edges include E12, E23, E34, E45, and E46, where E45 and E46 have a topological conflict because the arrow can only correspond to one primary destination; Graph neural networks can employ multi-layer graph convolution or equivalent neighborhood aggregation structures. Specifically, to accurately capture the relative constraints between local primitives, each layer of the network follows a clear message passing mechanism: based on candidate edges, the features of neighboring nodes and edge attributes are weighted and concatenated to form a neighborhood feature vector. After nonlinear mapping, this vector is fused with the central feature of the current node, thus completing one expansion of the receptive field. The iterative calculation method for node features is expressed as follows: in, Let be the hidden feature vector output at layer l+1 after node i has undergone one neighbor information aggregation and feature update. Let i be the hidden feature vector of node i in layer l. This represents the hidden feature vector of the corresponding neighbor node j at layer l. Let i represent the set of neighboring nodes. This represents the pooling aggregation function. Let l be the learnable weight parameter matrix of the l-th layer. This is the feature transformation weight matrix of the l-th layer for the center node itself. For activation functions; The first layer mainly integrates local adjacency information. For example, N4 obtains the context from the outlet valve after N3 and two text candidates from N5 and N6. The second layer introduces a wider range of information. For example, N4 further perceives that N1 is the compressor outlet. Combined with the process topology rules, going directly to the flare is usually an abnormal venting branch, while going to the quench tower is more in line with the normal path. After the network output, the connectivity probability of each candidate edge is calculated through the fully connected layer. For example, the probability of E45 may be 0.32 and the probability of E46 may be 0.81. In the initial logical topology, E46 is retained first. For example, for a broken line that is obscured by a stain, we can assume that there is a 12-pixel gap between line segments L1 and L2. If we only look at the image, the connectivity is insufficient. However, L1 is connected to the pump outlet upstream, and L2 is connected to the heat exchanger E-201 downstream. The two lines are basically aligned. After the graph neural network integrates the neighborhood, it can increase the probability of this broken connection to 0.87, allowing it to enter the initial logical topology network. This shows that the graph construction process is not a simple tracking of pixel lines, but a holistic judgment of nodes, relationships, and context. As an anomaly tolerance mechanism, if there are no candidate edges around a node, the system will still include the node in the isolated node set to retain information for subsequent manual review; if the probabilities of a group of candidate edges are close, for example, E45 is 0.49 and E46 is 0.51, and the minimum confidence difference is not reached, neither edge will be fixed as a determined edge, but will be written into the initial logical topology network as an edge pending decision; if there is a clear conflict in node categories, for example, the same location is identified as both a valve and a pump, the category with higher confidence is the main category, and the other category is retained as a conflict label. For example, during the analysis of drawings with hand-drawn red lines showing signs of modification during the overhaul of the pyrolysis unit, a hand-drawn line changed the branch that originally went to the quench tower to the flare, but the hand-drawn line intersected and overlapped with the original printed line. By constructing a candidate graph and performing neighborhood aggregation, the network identifies that the red line is more strongly associated with the text "temporary voiding", thereby increasing the probability of removing the torch edge in the changed version, while the original removal of the rapid cooling tower edge is reduced to a pending state in this local area; subsequently, the rule engine determines whether the temporary voiding complies with the working condition permit. The purpose of this step is to incorporate break lines, cross-page markers, and local fuzzy information into a unified reasoning framework through candidate relationship graphs and neighborhood aggregation mechanisms, thereby breaking the isolated state of nodes during the fusion of underlying features and providing a computable initial logical topology network for subsequent process rule verification.

[0023] In a preferred embodiment of the present invention, quantifying the topological logical conflict probability value of the local region based on the detection result includes: evaluating the degradation noise intensity of the acquired two-dimensional degraded document image, and calculating the topological logical conflict probability value of the local region based on the density of the logical conflict nodes, the degradation noise intensity, and the connectivity probability output by the graph neural network. This embodiment provides a local risk quantification step; specifically, after the aforementioned initial logical topology network is established, if a binary judgment is made solely based on whether a conflict is found, two types of problems are likely to occur: one is that slightly noisy areas are over-intercepted, causing unnecessary manual burden; Another type is where areas with severely degraded images but where conflicts have not yet fully manifested are mistakenly allowed to pass through; therefore, this embodiment introduces a comprehensive quantitative index composed of the density of logical conflict nodes, the intensity of degraded noise, and the connectivity probability. Specifically, the density of logical conflict nodes can be counted for each local area; the density of logical conflict nodes can be understood as the proportion of the number of nodes that the rule engine determines are inconsistent with physical conservation, equipment logic or valve status to the total number of nodes in that area. For example, if there are 10 nodes in a local region R4, and 3 of them are involved in the conduction of a normally closed valve without power backflow or an invalid loop anomaly, then the density of logically conflicting nodes can be taken as 0.3; the intensity of the degradation noise comes from the input quality assessment of the two-dimensional degraded document image, which can comprehensively consider blurriness, the proportion of smudges, uneven brightness, crease occlusion, and hand-drawn coverage; for example, the intensity of the degradation noise in the same region is assessed as 0.65; the connectivity probability can be taken as the average confidence of the retained candidate edges in the region, for example, 0.58; Combining the above three factors yields a region-level topological logic conflict probability value; intuitively, the denser the conflicting nodes, the worse the graph, and the more unstable the connected edges, the higher this value will be. Unlike methods that rely solely on the number of conflicts, this quantification method can distinguish between images that are clear but have a real design flaw and images with degraded noise levels exceeding a set baseline, causing the system to struggle to patch the lines. This provides a more robust basis for subsequent isolation decisions. In an anomaly tolerance mechanism, if a region has no effective connected edges, making it impossible to directly calculate the average connectivity probability, a default low confidence value can be used, or the region can be directly marked as a high-risk candidate region for more stringent review. If the number of nodes in a certain area is too small, such as only one text node and one incomplete line segment, the conflict density statistics are unstable and can be merged with neighboring areas before recalculation; if the intensity of degraded noise is extremely low but the density of logical conflict nodes is extremely high, it indicates that there is more likely a real major change or error in the drawing itself. In this case, the high-risk judgment should still be retained and the alarm level should not be reduced because the image quality is good. For example, in the branch diagram of the cracked gas compressor outlet, region R4 includes a branch to the quench tower, a flare bypass, and a check valve. The rule engine detected an anomaly in the path inferred from this region, where the flow back to the previous stage after the check valve. At the same time, the image is discontinuous due to reflections and creases, and the confidence of multiple edges output by the graph neural network is also low. After comprehensive analysis, R4 has a high probability value of topological logic conflict and is therefore included in the priority isolation and priority review queue. The purpose of this step is to integrate image quality, topological anomalies, and model confidence into a single risk scale, so that there is a repeatable and traceable basis for whether a local area needs to be actively downgraded.

[0024] In a preferred embodiment of the present invention, the calculation of the topological logical conflict probability value of the local region based on the density of the logical conflict nodes, the intensity of the degrading noise, and the connectivity probability output by the graph neural network includes: normalizing the density of the logical conflict nodes and the intensity of the degrading noise respectively, and then performing a weighted summation according to a preset weight to obtain the conflict base number. Divide the conflict base number by the connectivity probability to obtain the comprehensive risk index; normalize the comprehensive risk index to output the topological logic conflict probability value.

[0025] This embodiment provides a probability value calculation method that can be directly implemented. Specifically, under the above quantitative framework, if the specific calculation path is not disclosed, the risk ranking may be distorted due to the inconsistency of numerical scales in different regions during actual deployment. Therefore, this embodiment further provides a complete processing flow of normalization, weighting, division by confidence level, and final output. Specifically, the density of logically conflicting nodes and the intensity of degrading noise can be normalized first to obtain the normalized density of logically conflicting nodes. and the intensity of the normalized degraded noise The collision cardinality is obtained by weighting and summing the normalized data according to preset weights, and then dividing the collision cardinality by the connectivity probability. Derivation of comprehensive risk index The specific derivation formula used is as follows: in, and To satisfy the preset weights that sum to 1, a small positive number is preset. Used as a lower bound to prevent the denominator from being zero; preset weights and The aforementioned risk cutoff extreme value The preset danger threshold is an empirical hyperparameter obtained by grid search of historical petrochemical unit drawing change datasets or by joint calibration by expert experience, and can be dynamically and adaptively adjusted according to the average degradation noise intensity of the current input drawing. For example, the density of logically conflicting nodes is naturally between 0 and 1 and can be used directly; if the original score range of the degraded noise intensity is 0 to 100, then divide by 100 to obtain the normalized intensity of the degraded noise. Suppose the density of logically conflicting nodes in a certain region is 0.40, and the original score of the degraded noise intensity is 70. Then the intensity of the normalized degraded noise is... The value is 0.70; let the corresponding density weight be... The intensity weight is 0.6. If the value is 0.4, then the conflict base number of the numerator of the formula can be calculated as 0.6×0.40+0.4×0.70, which gives 0.52; If the connectivity probability output by the neural network of this region graph It is 0.65 and this value is greater than The comprehensive risk index is 0.52 ÷ 0.65 = 0.80. The reason for using division here is that when the density of conflicting nodes and the intensity of degrading noise are comparable, the lower the connectivity probability, the more unstable the system's connection judgment of the region, and the implicit risk should be amplified accordingly. Then, the comprehensive risk index Perform normalization processing and output the topological logic conflict probability value between 0 and 1. To ensure safety boundaries, the system adopts a truncated linear mapping structure, expressed as follows: in, The set risk cutoff extreme value; for example, the comprehensive risk index is set through parameters. Linear compression is applied to the range of 0 to 1. Any value exceeding the upper limit of 1.2 will be truncated by the min function and retained as 1. In the example above, 0.80 can be output as approximately 0.67 after mapping. If the preset danger threshold is 0.60, the area will enter the subsequent isolation process. To facilitate a microscopic explanation, another set of parallel data can be provided; the density of logical conflict nodes in region R5 is 0.20, the intensity score of degraded noise results in a normalized data of 0.30, the connectivity probability is 0.90, then the conflict base number is 0.24, the comprehensive risk index is approximately 0.27, and the normalized probability value is approximately 0.23. The density of logically conflicting nodes in region R6 is the same as that in R5, but the normalized intensity of the degraded noise increases to 0.80 and the connectivity probability decreases to 0.45. Therefore, the conflict base number is 0.44, the comprehensive risk index is about 0.98, and the normalized probability value is close to 0.82. The comparison shows that the same number of logically conflicting nodes does not mean the same comprehensive risk. As a boundary case handling mechanism, if the connectivity probability is close to 0, a preset lower limit is applied using the formula above, such as a value of 0.1. This can effectively avoid numerical overflow caused by an excessively small denominator; if a certain area lacks intensity assessment results for degraded noise, the average intensity of degraded noise in neighboring areas or the global average on the page can be used instead, but an estimated substitution mark is added at the same time so that the review end can track it. For example, in the above-mentioned pyrolysis device drawings, the intensity of the degraded noise in the torch bypass area is higher than the benchmark evaluation value due to the reflection of mobile phones, and the newly added hand-drawn connecting lines overlap with the original lines, which reduces the connectivity probability. Based on this, the system calculates that the topological logic conflict probability value of this area exceeds the threshold, while another area with only ordinary text corrections has a probability value lower than the preset danger threshold because the image quality is better and the connecting edges are stable, thereby realizing the system's automatic differentiated handling diversion defense line. The purpose of this step is to provide a structured, reproducible, and parameterizable quantitative extrapolation and calculation system, so that regional risk quantification and isolation decisions do not rely on opaque extrapolation processes based on human experience, but are instead built on unified and transparent numerical calculation rules.

[0026] In a preferred embodiment of the present invention, actively severing the logical connectivity of the local region in the initial logical topology network and marking the local region as an isolated subgraph region to be manually reviewed includes: blocking all undetermined connectivity paths extending outward within the local region; resetting the node state within the local region to an undefined state; generating a defensive isolation boundary around the local region in the initial logical topology network to prevent logical errors from cascading outward; and encapsulating the region within the defensive isolation boundary as the isolated subgraph region to be manually reviewed.

[0027] This embodiment provides an active isolation mechanism for high-risk areas. Specifically, after the aforementioned risk quantification is completed, if only a warning label is attached to the high-risk area without changing the network structure, the unstable connectivity in the area may still continue to affect other areas in full-graph reasoning, path tracing, or version comparison. Therefore, it is necessary to perform a substantial disconnection on the area at the logical level. Specifically, the system blocks all undetermined connecting paths within a local area from extending outward. For example, suppose a high-risk area R6 contains nodes N21, N22, and N23, and has candidate cross-area edges E21-30 and E23-41 with the external area. If these two cross-area edges have not yet been manually confirmed, the system changes their status from candidate propagation to blocked non-propagation, so that any path deduction within R6 cannot automatically affect the areas where nodes N30 and N41 are located. The node status within this area is reset to an undefined state. This reset does not mean deleting the node, but rather setting the node's connectivity, flow direction, or valve position to undefined, while preserving the image evidence and original features. For example, valve node N22 might have been inferred to be normally open, but due to the high risk of the area, the system retracts this inference, retaining only the basic information that the valve element's text may conflict with NO / NC. This prevents subsequent modules from mistakenly treating provisional inferences as confirmed facts. Furthermore, a defensive isolation boundary is generated in the initial logical topology network; this boundary can be understood as a special boundary marker that records the set of nodes within the boundary, the contact points between the boundary and the outside, and the reason for isolation. In the underlying graph structure, the step of generating a defensive isolation boundary is specifically manifested as follows: in the adjacency matrix of the initial logical topology network, the weight of the connection edge that crosses the local region and the external global network is forcibly set to zero, and the boundary contact point is marked as an absorbing node that cannot be crossed in the node state table, thereby truncating the message transmission and path deduction of the graph neural network at the logical level. For example, the boundary contact points of R6 include the inlet end B1 and the outlet end B2. The boundary attributes record the reasons as: high noise + backflow conflict + low confidence. If any subsequent path search encounters B1 or B2 and no manual correction result is received, it can only stop at the boundary and cannot be penetrated. The region within the boundary is encapsulated as an isolated sub-graph region to be manually reviewed; the encapsulated object can contain the original image slice, node list, candidate edge list, conflict rule list, and suggested review points; For example, checking the position of valve N22, verifying whether the red hand-drawn line extends across pages, and verifying whether the text 'removal' applies to the main line; in this way, manual review does not need to reinterpret the entire diagram, but focuses on high-risk areas; In an optional anomaly handling mechanism, if there are strongly defined edges within a high-risk area that have been confirmed by other sources, such as edges from a previously approved version that have not been modified in this instance, these strongly defined edges can be retained as read-only edges, but they are still prohibited from being used as new evidence for the outward expansion of the high-risk area. If two high-risk areas are adjacent and their boundaries overlap, they can be merged into a larger isolation sub-map to avoid boundary overlap causing difficulties in the review process. If the area of ​​an isolation sub-map is too large and exceeds the single-screen capacity of the review end, it can be further divided into multiple sub-packages according to contact points or process units, but each sub-package still maintains an overall isolation state. For example, in the cracked gas compressor outlet area, the system finds that the hand-drawn vent line to the flare and the original main line to the quench tower are intertwined in the same window, and the check valve symbol is obscured by creases. At this time, the system encapsulates this local area as an isolated sub-graph, blocking its automatic connection with the upstream and downstream main processes, and marks the inlet at the boundary as: compressor outlet; outlet as: quench tower inlet, flare main pipe. The review engineer only needs to confirm which line is the effective modification line in this encapsulated area to restore the corresponding main process. The purpose of this step is to proactively prevent uncertain connectivity relationships from having a spillover effect on the overall graph topology before high-risk local conditions are clarified, thereby achieving a global security-first processing strategy.

[0028] In a preferred embodiment of the present invention, if the probability value of the topological logical conflict is less than the preset danger threshold, the connection path of the local area is retained and a global topology network is generated, which includes: verifying that the connection path in the local area conforms to the deterministic topology rule; solidifying the verified connection path into the initial logical topology network; merging all verified areas and dynamically generating the global topology network that reflects the actual physical connection relationship.

[0029] This embodiment provides a mechanism for retaining and merging low-risk areas. Specifically, in the previous embodiment, high-risk areas were isolated to avoid the risk of the spread of uncertain areas. However, if all areas are conservatively isolated, the automation coverage will be too low. Therefore, for local areas below the danger threshold, a robust release and solidification process is needed to form a usable global topology network. Specifically, verify whether the connected paths within a local area satisfy deterministic topology rules; for example, there exists a path N31→N32→N33→N34 within a certain area, where N32 is a valve and N33 is a flow meter; In addition to confirming that the four nodes are connected, the system also needs to verify: whether the arrow direction is consistent with the text description, whether the valve status allows the current path to be valid, whether there is a meaningless closed loop, and whether the instruments are in a reasonable position; only when the image evidence and process rules of the path are both approved will the path be considered a solidifiable path. During the solidification process, the verified edges in the region can be upgraded from the candidate edge state to the verified edge state, and the source and version information are recorded at the same time. For example, the original probabilities of E31-32 are 0.91, E32-33 are 0.88, and E33-34 are 0.93. After all three edges pass the rule verification, they are written into the formal edge set of the global topology. If there is another branch road E32-35 in the same area, although the image confidence is high, it conflicts with the rule that normally closed bypass valves should not be normally open, so it will not be fixed. After each low-risk area is solidified, they are merged according to cross-regional contact points to dynamically generate a global topology network that reflects the actual physical connection relationship. The so-called dynamic generation means that after each new area is processed or a manual correction is received, the global network can recalculate its connected components, material flow direction and key equipment association, rather than having to recalculate and deduce the topology network of the entire drawing. For ease of explanation, it can be assumed that R1, R3 and R5 are all verified, and R2 and R4 are in an isolated state. Then, three trusted topologies are first formed in the global network, and R1 and R3 are connected through the confirmed cross-region interface, while R5 is not included in the main process because it has reached the isolation boundary. In an optional anomaly handling mechanism, if the risk value of a certain region is lower than the threshold, but there are individual edges in the region that are in an unresolved state, then partial solidification can be implemented for that region, that is, only the verified connected paths are solidified, and the unresolved edges remain in the candidate state. If two low-risk areas disagree on the classification of the same node at their boundary, for example, one considers it a stop valve while the other considers it a check valve, then node classification arbitration will be triggered before merging. If necessary, the conflict point will be transferred to local pending review without affecting the network connection of the rest. If new cross-regional conflicts occur in the global network after dynamic merging, for example, two originally independent and legal areas are connected to form an unreasonable loop, then the risk level of the relevant interface area will be raised again and transferred to isolation processing. For example, at the overhaul site, the steam tracing branch, cooling water branch, and main material outlet branch were uploaded by different work teams. After the system verified their local paths, it confirmed that the steam tracing branch and the main material outlet were only coupled through a heat exchanger rather than directly mixed, and the valve position of the cooling water branch was consistent with the label. Therefore, these low-risk areas were solidified into the formal topology. Although the flare bypass area was still pending review, the global network could still output the main trusted structure first, so that the maintenance command center could view the confirmed scope of changes. The purpose of this step is to ensure that the system can guarantee the safety boundaries while preventing the output of usable results for the entire drawing from being blocked by individual high-risk local areas, thus achieving priority networking of trusted areas.

[0030] In a preferred embodiment of the present invention, triggering a manual review defense based on the isolated sub-graph region to be manually reviewed, updating the global topology network in conjunction with the feedback results of the manual review, and outputting the final document drawing change parsing version includes: pushing alarm information containing the isolated sub-graph region to a preset review terminal, as well as image slices of the corresponding local region in the two-dimensional degraded document image; and receiving manually input corrected connectivity relationships. The isolated subgraph regions awaiting manual review are replaced with the corrected connectivity relationships, and the overall topology is recalculated. The updated global topology network is then converted into structured change data, and the final document drawing change parsing version is output.

[0031] This embodiment provides a closed-loop manual review and result write-back mechanism; specifically, although the aforementioned isolation strategy can prevent high-risk areas from polluting the global network, if there is no clear review entry and write-back mechanism, the isolated area will remain for a long time and will not be able to form a final deliverable version. Therefore, this embodiment transforms the isolated subgraph into an operable review task, and automatically updates the global topology and structured change results after the review is completed; Specifically, the system pushes alarm information and corresponding image slices to the review end; the alarm information may include at least the area number, risk value, main conflict type, involved equipment or valve tag number, and boundary contact point with upstream and downstream; the image slices may display the original image, enhanced image, and current candidate topology overlay image; For example, for region R6, the reviewer can simultaneously see the hand-drawn red line in the original photo, the valve node frame identified by the system, and the candidate connection direction; in this way, the review engineer can complete the judgment locally without switching between multiple systems. When receiving manual input, the review end can provide several standardized operations, such as confirming connectivity, deleting connectivity, modifying valve positions, specifying cross-page corresponding points as removal lines, and adding device nodes. For ease of explanation, it can be assumed that the engineer deleted the original candidate edge E21-30, confirmed the new edge E22-41, and changed the state of node N22 from undefined to normally open. After receiving the data, the system not only records the result itself, but also the operator, timestamp, and basis for correction, so as to facilitate version traceability. During the write-back phase, the system replaces the original candidate relationships in the isolated subgraph region with corrected connectivity relationships, and then recalculates the topology of the entire graph. For example, the main path compressor outlet → quench tower inlet, which was originally disconnected due to isolation, may be restored or legally rewritten as compressor outlet → temporary vent → flare main. The system re-runs cross-regional path checks, loop checks, and flow consistency checks to confirm that the correction will not introduce new process conflicts. The updated global topology network is converted into structured change data; this structured data may include device connection tables, node attribute tables, tables showing differences before and after changes, lists of devices to be removed and added, version numbers, and audit logs. To illustrate, a certain record can be written as: the original connection from V-102 outlet to quench tower inlet has been changed to V-102 outlet to flare main temporary vent line, with the status being approved and effective, and the reviewer and time being written simultaneously; thus forming the final document drawing change parsing version; In an optional exception handling mechanism, if the reviewer only provides partial corrections, such as only confirming the deletion of a mistakenly connected edge but not confirming the addition of a new path, the system can perform a partial update on the subgraph and maintain the isolated state of the remaining unresolved parts. If different reviewers give conflicting conclusions for the same area, it can enter the second review queue, and the original isolation status will not be lifted; if new rule conflicts are triggered when the full map topology is recalculated after writing back, the system will not directly release the final version, but will repackage the new conflicting area for review and keep the release status suspended in the version repository. For example, during a major overhaul of a pyrolysis unit, after receiving the flare bypass isolation area, the review engineer discovered through slicing that the red hand-drawn arrow on the original diagram indicated that it was for commissioning, meaning that the branch was for temporary venting and should not cover the original long-term production path; the engineer then confirmed the addition of a temporary venting line while retaining the original main line. The system updates the global topology accordingly, generating a change resolution version with an expiration date, which clearly distinguishes between long-term valid connections and temporary connections during the trial run. The purpose of this step is to write back the manual judgments of high-risk localities into the global topology network in a structured manner, forming a traceable and publishable change management closed loop.

[0032] In a preferred embodiment of the present invention, the two-dimensional degraded document image is a scanned copy of a chemical pipeline and instrumentation diagram; the unstructured modification traces are manual changes and annotations made on-site; the deterministic topology rules include chemical material flow direction restriction rules and valve normally open and normally closed logic rules. This embodiment provides a method for defining scene constraints and rule sets; specifically, in order to make the aforementioned image processing, mapping, risk quantification, isolation and write-back processes repeatable in chemical engineering scenarios, this embodiment further clarifies the applicable boundaries of input objects, modification sources and deterministic topology rules; Specifically, the input images are preferably scanned copies of chemical pipeline and instrumentation diagrams or their photographic digitization results; these drawings have typical elements such as equipment symbols, valve symbols, instrument tag numbers, pipeline numbers, arrow flow directions, and cross-page indexes, making them suitable for extracting structured information through the aforementioned multimodal visual processing and mapping mechanisms; on-site manual changes can be expressed as red line corrections, circle annotations, handwritten new text, hand-drawn arrows, strikethroughs, or cross-page fill lines; these modifications are usually not standardized electronic objects, but are directly superimposed on the original drawings, thus requiring a risk-controlled analysis process; Deterministic topology rules can include at least two categories; the first category is rules restricting the flow of chemical materials, such as that the main process should not flow backwards when there is no driving equipment or differential pressure support; certain branches can only flow from upstream to downstream and should not form uncontrolled backflow; and closed loops should not be automatically formed in areas where no circulation equipment is installed. The second category is valve normally open and normally closed logic rules. For example, valves marked as normally closed should not be automatically considered as conducting in the normal process path. If a valve marked as normally open is manually changed to normally closed, its upstream and downstream material paths must be restricted synchronously. The system can write these rules into executable inspection templates and call them repeatedly during topology verification and regional risk quantification. For ease of microscopic explanation, we can assume that there are three valves in a certain drawing: V1 is marked NO, V2 is marked NC, and V3 is not explicitly marked; if the system infers that a path must pass through V2 to be valid, then the path should be judged as a conflict path by the rule engine; if there is a backflow between V1 and V3, but there is no pump or compressor in between, then it should also be marked as abnormal. For example, if a diagram shows a hand-drawn arrow connecting the product line to the main flare pipe, the system needs to combine this with the rule that temporary venting is allowed but subject to operating conditions, and classify it as a highly sensitive path that requires review, rather than simply connecting it automatically based on the image connection. In an optional exception handling mechanism, if the input document is not a P&ID but an equipment layout drawing, civil engineering drawing or electrical schematic diagram, it can be diverted to other parsing processes during the pre-classification stage and will not enter this method; if the on-site manual annotation does not contain any process connections and only has signatures, dates or approval marks, the system can archive it as non-topology modification information and will not participate in connectivity reconstruction. If some valves do not have normally open or normally closed markings, they can be treated as unknown valve positions. Instead of directly negating or confirming the path based on this, the uncertainty is increased in the risk calculation. For example, at the aforementioned ethylene cracking unit overhaul site, the uploaded drawings were all P&ID scans. The engineer used a red pen to write "test run vent to flare" on the compressor outlet branch and marked a normally closed valve as open. The system uses chemical material flow restriction rules to determine that the newly added branch should not flow back to the compressor inlet. It also uses normally open and normally closed valve logic rules to identify that this is a critical state change, so it is included in the high-risk review scope. After the review is completed, the final version of the change output by the system clearly reflects the linkage between the valve position adjustment and the temporary venting branch. The purpose of this step is to clarify the applicable objects and process rules of this method, so that the entire analysis and change management process is based on the professional constraints of the chemical P&ID scenario.

[0033] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for intelligent parsing and change management of chemical P&ID process drawings based on image processing, characterized in that: include: A two-dimensional degraded document image containing unstructured modification traces is obtained, and the underlying image connectivity features and character features of the two-dimensional degraded document image are extracted using a multimodal visual model and optical character recognition technology. The underlying image connectivity features and character features are input into a graph neural network for feature aggregation to reconstruct the initial logical topology network of the two-dimensional degraded document image; The initial logical topology network is divided into subgraphs based on a preset graph segmentation algorithm to obtain multiple local regions; The connected paths in the initial logical topology network are input into the graph theory rule engine, and logical conflict nodes that violate the physical conservation law are detected by combining the preset deterministic topology rules. Based on the detection results, the topological logical conflict probability value of each local region is quantified. The topological logical conflict probability value represents the probability of a logical paradox between image semantic understanding and physical rules. The probability value of the topological logical conflict is compared with a preset danger threshold that characterizes the tolerance of logical errors. If the probability value of the topological logical conflict is not less than the preset danger threshold, the prediction of the connectivity between the primitive nodes in the local area is abandoned, the logical connectivity of the local area in the initial logical topology network is actively cut off, and the local area is marked as an isolated subgraph area to be manually reviewed. If the probability value of the topological logic conflict is less than the preset danger threshold, the connectivity path of the local area is retained, and a global topology network is generated. Based on the isolated subgraph area to be manually reviewed, a manual review defense line is triggered. The corrected connectivity relationship of the isolated subgraph area to be manually reviewed is received as the manual review feedback result. The global topology network is updated in combination with the manual review feedback result, and the final document drawing change parsing version is output, forming a closed loop for drawing change management.

2. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to claim 1, characterized in that, A two-dimensional degraded document image containing unstructured modification traces is obtained. The underlying image connectivity features and character features of the degraded document image are extracted using a multimodal visual model and optical character recognition technology, including: Obtain a scanned copy of the process drawing with hand-drawn change annotations, degraded noise, and cross-page lines as the two-dimensional degraded document image; The multimodal visual model is used to denoise the two-dimensional degraded document image and identify primitive node features and spatial position relationships. Text attribute features are extracted using the optical character recognition technology, and the primitive node features, spatial position relationships and text attribute features are fused into the underlying image connectivity features and character features.

3. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to claim 2, characterized in that, The process of inputting the underlying image connectivity features and character features into a graph neural network for feature aggregation to reconstruct the initial logical topology network of the document includes: Construct a graph structure; wherein the nodes of the graph structure include the graph primitive node features and the text attribute features, and the edges of the graph structure represent the connectivity assumptions derived from the spatial location relationships; perform deep learning on the graph structure using multi-layer graph convolution operations to aggregate neighbor node information; predict the connectivity probability between nodes, and generate the initial logical topology network based on the connectivity probability.

4. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to claim 3, characterized in that, The probability value of topological logical conflict in the local region is quantified based on the detection results, including: The degradation noise intensity of the acquired two-dimensional degraded document image is evaluated, and the topological logical conflict probability value of the local region is calculated based on the density of the logical conflict nodes, the degradation noise intensity, and the connectivity probability output by the graph neural network.

5. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to claim 4, characterized in that, The calculation of the topological logical conflict probability value of the local region, based on the density of the logically conflicting nodes, the intensity of the degraded noise in the two-dimensional degraded document image, and the connectivity probability output by the graph neural network, includes: After normalizing the density of the logically conflicting nodes and the intensity of the degraded noise respectively, the nodes are weighted and summed according to preset weights to obtain the conflict base number. Divide the conflict base number by the connectivity probability to obtain the comprehensive risk index; normalize the comprehensive risk index to output the topological logic conflict probability value.

6. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to claim 1, characterized in that, Actively severing the logical connectivity of the local area in the initial logical topology network and marking the local area as an isolated subgraph area to be manually reviewed includes: blocking all undetermined connectivity paths within the local area from extending outward; The node states within the local area are reset to an undefined state; a defensive isolation boundary is generated around the local area in the initial logical topology network to prevent logical errors from cascading outwards; the area within the defensive isolation boundary is encapsulated as the isolated subgraph region to be manually reviewed.

7. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to claim 1, characterized in that, If the probability value of the topological logical conflict is less than the preset danger threshold, then the connected paths of the local area are retained, and the global topology network is generated by: verifying that the connected paths within the local area conform to the deterministic topology rules; The verified connectivity paths are solidified into the initial logical topology network; all verified regions are merged to dynamically generate the global topology network that reflects the actual physical connectivity relationships.

8. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to claim 1, characterized in that, Based on the isolated sub-graph area awaiting manual review, the manual review defense line is triggered. The global topology network is updated in combination with the feedback results of the manual review, and the final document drawing change parsing version is output, including: pushing alarm information containing the isolated sub-graph area awaiting manual review to the preset review terminal, as well as the image slice of the corresponding local area in the two-dimensional degraded document image; Receive manually inputted corrected connectivity relationships; replace the isolated subgraph regions to be manually reviewed with the corrected connectivity relationships, and recalculate the overall topology state; convert the updated global topology network into structured change data, and output the final document drawing change parsing version.

9. The intelligent parsing and change management method for chemical P&ID process drawings based on image processing according to any one of claims 1 to 8, characterized in that, The two-dimensional degraded document image is a scanned copy of a chemical pipeline and instrumentation diagram; the unstructured modification traces are manual changes and annotations made on-site. The deterministic topology rules include chemical material flow restriction rules and valve normally open and normally closed logic rules.

Images