An engineering drawing data preprocessing method and system based on image understanding

Abstract

This invention discloses a method and system for preprocessing engineering drawing data based on image understanding, comprising the following steps: S1, acquiring image data of the drawing and extracting features to form a feature representation set; S2, performing semantic segmentation on the image data to generate a region label map; S3, performing joint recognition on the feature representation set and the region label map to construct a drawing element set; S4, matching the drawing elements with an existing database to construct a mapping relationship table and marking abnormal data; S5, performing structural analysis on the drawing elements to generate a causal table; S6, performing scoring and clustering operations on the abnormal data to form a set to be corrected; S7, fusing the mapping relationship table, the causal table, and the set to be corrected to obtain the drawing preprocessing result. This invention integrates deep learning algorithms, image semantic segmentation technology, and graphic structure analysis, possessing advantages such as high processing efficiency, strong interactivity, and high data accuracy.

Description

Technical Field

[0001] This invention relates to the field of engineering drawing digitization technology, and in particular to an engineering drawing data preprocessing method and system based on image understanding. Background Technology

[0002] In engineering construction phases such as industrial design, equipment installation, and process control, P&ID drawings serve as core documents expressing equipment layout, control logic, instrument connections, and operational data interaction methods, forming a crucial foundation for achieving digital project delivery and intelligent operation and maintenance. However, existing engineering drawings typically exist in unstructured formats such as images or PDFs, lacking a semantically structured organization, making it difficult to directly participate in critical applications such as system modeling, operation and maintenance analysis, and operational control.

[0003] Current methods for digitizing drawings mostly rely on template matching, text recognition, or semi-manual annotation. These methods can only extract some static information from drawings, such as text labels or equipment tag numbers, and cannot effectively analyze graphic boundaries, equipment shapes, pipeline connections, and interlocking logic structures. Furthermore, due to inconsistent drawing standards, a wide variety of equipment symbols, and contextual dependencies in connection relationships, traditional identification methods suffer from significant shortcomings in accuracy and adaptability, making it difficult to support a comprehensive understanding and integration of equipment objects, OPC locations, and causal logic information.

[0004] Furthermore, existing technologies generally lack the ability to intelligently verify and correlate engineering drawing data with systems. When faced with frequent drawing updates, equipment adjustments, and changes in operating parameters in actual engineering projects, the drawing data cannot maintain consistency with operational databases such as tag number libraries and OPC point libraries, resulting in data lag and the risk of errors in the data used by the system. Especially in the construction of causal logical relationships and the detection of drawing anomalies, traditional methods rely on manual review, which is not only inefficient but also prone to human oversight, severely restricting the activity level and data reliability of engineering drawings.

[0005] Therefore, how to provide a method and system for preprocessing engineering drawing data based on image understanding is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for preprocessing engineering drawing data based on image understanding. This invention fully integrates deep learning algorithms, image semantic segmentation technology, and graphic structure analysis models. It describes in detail the joint identification, semantic parsing, data matching, and anomaly classification process for equipment objects, instrument objects, data point objects, and causal logic objects in equipment drawings. It constructs a structured drawing data expression method for interactive display, which has the advantages of high processing efficiency, strong semantic understanding ability, strong drawing interactivity, and high data accuracy.

[0007] An image-understanding-based engineering drawing data preprocessing method according to an embodiment of the present invention includes the following steps: S1. Obtain the image data of the engineering drawings to be processed, and use a convolutional neural network to extract features to form a feature representation set; S2. Perform semantic segmentation on the image data using the DeepLabv3+ model, and label the three types of regions—equipment, data points, and causal logic—according to preset classification rules, and generate region label maps. S3. The Mask R-CNN algorithm is used to perform joint recognition operations on the feature representation set and the region label map to extract the graphic boundaries, attribute text and spatial location of the three types of regions and construct the drawing element set. S4. Extract the drawing elements of the equipment, instruments and data point areas, and use the Hungarian algorithm to match and verify them with the existing tag number library and OPC point library to build a mapping relationship table. Mark the drawing elements that fail to match as abnormal data. S5. Perform structural analysis on the drawing elements of the causal logic region through a conditional random field, extract the triggering conditions and execution results, and generate a causal table by combining the mapping relationship table. S6. The isolated forest algorithm is used in conjunction with the causal table to perform scoring and causal clustering operations on the abnormal data, identify the anomaly type, and form a set to be corrected containing anomaly type labels. S7. Merge the mapping relationship table, causal table and the set to be corrected and perform color association operation to obtain an interactive drawing preprocessing result.

[0008] Optionally, the tag library represents a preset database of equipment and instrument tag information, the OPC point library represents an OPC data point database used for operational status interaction, the color association operation represents marking different categories of drawing elements with specific colors and performing jump associations, and the drawing preprocessing result represents a structured data output result used for drawing understanding and engineering data interaction.

[0009] Optionally, S2 specifically includes: S21. Perform size unification and boundary padding operations on the image data, and adjust the numerical range of the image color information according to a preset threshold to form standard image data, specifically including: Perform resolution adjustment on the image data, reset the image width and height to the preset standard size, and maintain the original aspect ratio of the image; Perform boundary padding on the adjustment results, and use mirror expansion to fill in missing pixels; Threshold constraint processing is performed on the pixel values ​​representing image color information in the boundary filling result. The pixel values ​​are cropped and restricted according to the preset threshold. Pixel values ​​that exceed the threshold range are mapped to a uniform numerical range to obtain standard image data. S22. Input standard image data into the DeepLabv3+ model and perform dilated convolution operation to extract contextual features at different sampling intervals and generate a semantic feature set containing multi-scale structural information. S23. Perform cross-scale feature aggregation on the semantic feature set, and perform upsampling interpolation on the aggregation result in combination with the spatial distribution of the image data to generate a semantic segmentation map with the same spatial size as the image data. S24. According to the preset classification rules, perform a category judgment operation on each pixel in the semantic segmentation map, and label the judgment result as the device instrument area, data point area or causal logic area. S25. Perform a pixel-by-pixel alignment operation between the labeled three types of regions and the image data to obtain the region label map.

[0010] Optionally, the semantic segmentation map generation process specifically includes: In the DeepLabv3+ model, feature encoding operations are performed on standard image data to generate intermediate feature maps; By employing dilated convolution kernels with different dilation rates, context-aware operations are performed on intermediate feature maps in parallel at multiple sampling intervals to extract context features at multiple scales and form a semantic feature set. Dimension alignment and feature normalization operations are performed on the context features at each scale in the semantic feature set, and pixel-wise weighted aggregation operation is performed on the normalization results according to the preset weight coefficient set to form a fused feature map; A stepwise upsampling operation is performed on the fused feature map, and a bilinear interpolation algorithm is used to obtain an initial feature map with the same spatial size as the image data. Specifically, this includes: The fused feature map is divided into multiple upsampling stages according to a preset number of upsampling layers, and the target size of each stage is set. The bilinear interpolation algorithm is used to perform spatial size expansion operation on the fused feature map. The fused feature map is interpolated according to the target size, and the difference result is numerically smoothed according to the set numerical range to generate an upsampled feature map as the input for the next upsampling stage. Use the upsampled feature map output from the last upsampling stage as the initial feature map; Edge structure smoothing and spatial coordinate compensation operations are performed on the initial feature map to correct the jagged edge effect generated during upsampling, resulting in a semantic segmentation map.

[0011] Optionally, S3 specifically includes: S31. Spatially align the feature representation set and the region label map to establish a mapping relationship between the feature representation set and the region label map, and construct joint feature data containing feature representations and region labels; S32. Based on joint feature data, the Mask R-CNN algorithm is used to perform candidate region generation operations on the equipment and instrument region, data point region and causal logic region respectively, forming a candidate region set that corresponds one-to-one with the three types of regions. S33. Perform boundary regression operation on each candidate region to determine the corresponding graphic boundary range in the engineering drawing image, specifically including: Perform a region cropping operation on the boundary range of each candidate region in the image data to extract the image sub-region corresponding to the boundary of the candidate region; Perform multi-scale convolution operations on image sub-regions to extract boundary direction features and calculate the response intensity and gradient change rate in each boundary direction; Based on the gradient change rate, the coordinate offsets of the upper and lower boundaries and the left and right boundaries are calculated respectively. The coordinate offsets are then numerically superimposed with the boundary coordinates of the corresponding candidate regions to obtain the boundary correction results. The overlap between the boundary correction result and the corresponding region in the region label map is compared. If the overlap is lower than the preset overlap threshold, the boundary range is iteratively refined, the receptive field range of the convolution kernel is updated, and the multi-scale convolution, boundary correction and overlap comparison operations are repeated. Candidate regions with an overlap degree greater than or equal to the overlap threshold are marked as boundary localization completed, and the corresponding boundary correction results are used as the graphic boundary range. S34. Within the boundary of the graphic, perform attribute text recognition on the joint feature data, extract the attribute text, and bind it to the corresponding candidate region. S35. Perform spatial location encoding on all candidate regions to generate spatial marker vectors for identifying the two-dimensional coordinate positions of the candidate regions. S36. According to the corresponding category labels in the area label map, summarize the three types of candidate area sets that have completed boundary positioning, attribute identification and spatial coding, and construct the drawing element set.

[0012] Optionally, S32 specifically includes: S321. Select the pixel range corresponding to the equipment and instrument region, data point region and causal logic region in the joint feature data, perform multi-scale sliding window scanning operation on the feature representation within the pixel range, and extract a fixed-size set of image segments. S322. Perform two-dimensional convolution and batch normalization operations on each image segment in the image segment set to generate a convolution response map, and compare the convolution response map with the region label map pixel by pixel. S323. Compare the image segments that match the three types of region labels in the comparison results, perform bounding box coordinate prediction operation, and determine the candidate boundary range of the image segment in the image data by calculating the boundary offset of the image segment in the horizontal and vertical directions. S324. Perform a fixed-region convolution operation within the candidate boundary range, calculate the feature mean and gradient distribution, and perform a category probability scoring operation based on the annotation information in the region label map. S325. Perform an overlapping region judgment operation on all image segments with candidate boundary ranges, compare image segments with a boundary overlap greater than a preset overlap threshold, retain the image segment with higher category probability score as candidate region, and form three candidate region sets corresponding to each of the three types of regions.

[0013] Optionally, S4 specifically includes: S41. Select the drawing elements labeled as equipment and instrument area and data point area from the drawing element set to form equipment and instrument element set and data point element set respectively; S42. Extract the tag text features, spatial location features, and area category features of each drawing element from the equipment and instrument element set and the data point element set respectively, and construct the equipment feature vector and the data point feature vector respectively according to the preset feature order. S43. Using the Hungarian algorithm, the feature vectors of each device are matched with the tag features in the existing tag database to generate a set of matching results between the device instrument element set and the tag database, specifically including: Read the corresponding tag text features, spatial location features, and region category features from the feature vector of each device, and perform character normalization processing on the tag text features to obtain a standard feature sequence; Extract all tag feature sets from the tag library, perform character normalization processing on the tag text in each tag feature, and perform unified coordinate system transformation processing on the spatial location information in the tag features to form a standard tag feature set; Based on the standard feature sequence and the standard position number feature set, the text difference value and spatial position difference value between each standard feature vector and the standard position number feature are calculated one by one, and the two difference values ​​are combined according to the preset weight to construct the matching cost matrix; The Hungarian algorithm is used to perform row and column normalization, zero element marking, and minimum cost path search operations on the matching cost matrix to calculate the optimal matching result. The optimal matching result that satisfies the one-to-one constraint is recorded as the valid matching result, and the results are summarized to form a set of matching results. S44. Using the Hungarian algorithm, the feature vectors of each data point are matched with the feature vectors of the existing OPC point database to generate a set of matching results between the data point feature set and the OPC point database. S45. Perform consistency verification on the two matching result sets, record the matching results that meet the preset conditions as valid mapping relationships, and summarize them to form a mapping relationship table; S47. Mark drawing elements that do not form a valid mapping relationship in the equipment and instrument element set and data point element set as abnormal data.

[0014] Optionally, S5 specifically includes: S51. Select the drawing elements marked as causal logic areas from the drawing element set to form a causal logic element set; S52. Based on the causal logic element set, a conditional random field is used to perform sequence modeling operations on the graphical structure and attribute text within the causal logic region. The causal logic elements are labeled with states according to preset state transition rules, and triggering conditions and execution results are identified and distinguished. Specifically, this includes: Extract graphical boundaries, attribute text, and spatial order information from each causal logic element to construct a causal logic sequence; Construct directed connections based on the order of drawing reading in the causal logic sequence, and generate an initial connection diagram based on the spatial position of each causal logic element; Context encoding is performed on the initial connection graph to extract the graphical pattern features and textual semantic features of each causal logical element in the adjacent nodes before and after it, and then they are fused to form context features. Input the context features into the conditional random field, and perform state probability estimation operations on the causal logic elements corresponding to each context feature according to the preset state set; Based on the probability estimation results, a state label is assigned to each causal logic element, marking it as a triggering condition or execution result; S53. Based on the status labeling results, perform causal relationship pairing operations on the triggering conditions and execution results to construct one-to-many or many-to-one associations between the triggering conditions and execution results, and generate a set of causal relationships. S54. Align the causal relationship set with the mapping relationship table, and map the triggering conditions and execution results to the corresponding equipment, instruments or data points in the drawing elements to form a causal relationship record with equipment identifiers and data point identifiers. S55. Perform a structural consistency check operation on the causal relationship records, sort the triggering conditions and execution results according to the preset logical order, and summarize the causal relationship records that pass the check to generate a causal table.

[0015] Optionally, S6 specifically includes: S61. Extract four types of features from each abnormal data: attribute text, spatial location, semantic label, and structural relationship, and construct an abnormal feature set; S62. Based on the set of abnormal features, construct multiple random trees using the isolated forest algorithm and perform anomaly scoring operations; S63. Correspond to each abnormal score with the causal relationship record in the causal table to map the causal logic characteristics corresponding to the abnormal data. S64. Perform density similarity clustering on all causal logical features, construct a causal clustering relationship graph between abnormal data and divide it into associated clusters to form multiple abnormal cluster sets; S65. For each abnormal data in the abnormal cluster set, combine the corresponding abnormal score with the distance to the cluster center, perform an abnormal type determination operation, assign an abnormal type label to the abnormal data according to the preset abnormal classification rules, and summarize to form a set to be corrected containing abnormal type labels.

[0016] An image understanding-based engineering drawing data preprocessing system according to an embodiment of the present invention includes: The feature extraction module is used to acquire image data of the engineering drawings to be processed, and to extract features using a convolutional neural network to form a feature representation set; The semantic segmentation module is used to perform semantic segmentation on image data using the DeepLabv3+ model, and to label three types of regions—equipment, data points, and causal logic—according to preset classification rules, generating region label maps. The joint recognition module is used to perform joint recognition operations on the feature representation set and the region label map using the Mask R-CNN algorithm, extracting the graphic boundaries, attribute text and spatial location of the three types of regions, and constructing a set of drawing elements. The matching and verification module is used to extract the drawing elements of equipment instruments and data point areas, and use the Hungarian algorithm to match and verify them with the existing tag number library and OPC point library to build a mapping relationship table, and mark the drawing elements that fail to match as abnormal data. The structure parsing module is used to perform structure parsing operations on the drawing elements of the causal logic region through a conditional random field, extract the triggering conditions and execution results, and generate a causal table by combining the mapping relationship table. The anomaly classification module is used to perform scoring and causal clustering operations on anomaly data using the Isolation Forest algorithm combined with a causal table, to identify anomaly types and form a set of data to be corrected containing anomaly type labels. The fusion and association module is used to merge the mapping relationship table, the cause-effect table, and the set to be corrected, and perform color association operations to obtain interactive drawing preprocessing results.

[0017] The beneficial effects of this invention are: First, this invention introduces convolutional neural networks, DeepLabv3+ semantic segmentation model and Mask R-CNN object detection algorithm to achieve accurate identification and structural analysis of equipment, instruments, data points and causal logic areas in engineering drawings. This solves the problem that traditional template matching and OCR methods are difficult to adapt to the complex and diverse structure of drawings, and improves the accuracy and robustness of drawing element extraction.

[0018] Secondly, this invention combines the Hungarian algorithm with modeling methods such as conditional random fields to construct a mapping relationship between drawing elements and the tag number library and OPC point library, and generates a causal logic table between triggering conditions and execution results. At the same time, it uses the isolated forest algorithm to identify abnormal elements in the drawings, realizing consistency verification and anomaly clustering analysis of drawing semantics, and providing technical support for automatic verification and semantic modeling of engineering drawings.

[0019] Finally, this invention constructs an interactive drawing data structure that integrates mapping relationships, causal relationships, and anomaly information. It uses color coding and jump mechanisms to visually enhance various drawing elements, thereby improving the usability and readability of engineering drawings in actual operation and maintenance systems, reducing manual identification costs, and enhancing the data activity and interactive value of drawings in the process of connecting with engineering management and intelligent systems. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an image understanding-based engineering drawing data preprocessing method proposed in this invention; Figure 2 This is a flowchart of the drawing data matching and causal logic structure generation process of an image understanding-based engineering drawing data preprocessing method proposed in this invention. Figure 3 This is a module structure diagram of an engineering drawing data preprocessing system based on image understanding proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1-2 An image understanding-based method for preprocessing engineering drawing data includes the following steps: S1. Obtain the image data of the engineering drawings to be processed, and use a convolutional neural network to extract features to form a feature representation set; S2. Perform semantic segmentation on the image data using the DeepLabv3+ model, and label the three types of regions—equipment, data points, and causal logic—according to preset classification rules, and generate region label maps. S3. The Mask R-CNN algorithm is used to perform joint recognition operations on the feature representation set and the region label map to extract the graphic boundaries, attribute text and spatial location of the three types of regions and construct the drawing element set. S4. Extract the drawing elements of the equipment, instruments and data point areas, and use the Hungarian algorithm to match and verify them with the existing tag number library and OPC point library to build a mapping relationship table. Mark the drawing elements that fail to match as abnormal data. S5. Perform structural analysis on the drawing elements of the causal logic region through a conditional random field, extract the triggering conditions and execution results, and generate a causal table by combining the mapping relationship table. S6. The isolated forest algorithm is used in conjunction with the causal table to perform scoring and causal clustering operations on the abnormal data, identify the anomaly type, and form a set to be corrected containing anomaly type labels. S7. Merge the mapping relationship table, causal table and the set to be corrected and perform color association operation to obtain an interactive drawing preprocessing result.

[0023] In this embodiment, the tag library represents a preset database of equipment and instrument tag information, the OPC point library represents an OPC data point database used for operational status interaction, the color association operation represents marking different categories of drawing elements with specific colors and performing jump associations, and the drawing preprocessing result represents the structured data output result used for drawing understanding and engineering data interaction.

[0024] In this embodiment, S2 specifically includes: S21. Perform size unification and boundary padding operations on the image data, and adjust the numerical range of the image color information according to a preset threshold to form standard image data, specifically including: Perform resolution adjustment on the image data, reset the image width and height to the preset standard size, and maintain the original aspect ratio of the image; Perform boundary padding on the adjustment results, and use mirror expansion to fill in missing pixels; Threshold constraint processing is performed on the pixel values ​​representing image color information in the boundary filling result. The pixel values ​​are cropped and restricted according to the preset threshold. Pixel values ​​that exceed the threshold range are mapped to a uniform numerical range to obtain standard image data. S22. Input standard image data into the DeepLabv3+ model and perform dilated convolution operation to extract contextual features at different sampling intervals and generate a semantic feature set containing multi-scale structural information. S23. Perform cross-scale feature aggregation on the semantic feature set, and perform upsampling interpolation on the aggregation result in combination with the spatial distribution of the image data to generate a semantic segmentation map with the same spatial size as the image data. S24. According to the preset classification rules, perform a category judgment operation on each pixel in the semantic segmentation map, and label the judgment result as the device instrument area, data point area or causal logic area. S25. Perform a pixel-by-pixel alignment operation between the labeled three types of regions and the image data to obtain the region label map.

[0025] In this embodiment, the semantic segmentation map generation process specifically includes: In the DeepLabv3+ model, feature encoding operations are performed on standard image data to generate intermediate feature maps; By employing dilated convolution kernels with different dilation rates, context-aware operations are performed on intermediate feature maps in parallel at multiple sampling intervals to extract context features at multiple scales and form a semantic feature set. Dimension alignment and feature normalization operations are performed on the context features at each scale in the semantic feature set, and pixel-wise weighted aggregation operation is performed on the normalization results according to the preset weight coefficient set to form a fused feature map; A stepwise upsampling operation is performed on the fused feature map, and a bilinear interpolation algorithm is used to obtain an initial feature map with the same spatial size as the image data. Specifically, this includes: The fused feature map is divided into multiple upsampling stages according to a preset number of upsampling layers, and the target size of each stage is set. The bilinear interpolation algorithm is used to perform spatial size expansion operation on the fused feature map. The fused feature map is interpolated according to the target size, and the difference result is numerically smoothed according to the set numerical range to generate an upsampled feature map as the input for the next upsampling stage. Use the upsampled feature map output from the last upsampling stage as the initial feature map; Edge structure smoothing and spatial coordinate compensation operations are performed on the initial feature map to correct the jagged edge effect generated during upsampling, resulting in a semantic segmentation map.

[0026] In this embodiment, the edge structure smoothing and spatial coordinate compensation operation specifically includes: A fixed-size sliding window is constructed in the initial feature map, and the first derivative between the gray values ​​of each pixel is calculated as the gradient change value within each window; Pixels whose gradient change values ​​exceed a preset gradient threshold are identified as edge regions, and an edge region set is constructed. Perform numerical adjustment operations on the pixel values ​​in the edge region set, and perform interpolation averaging based on the numerical differences between adjacent pixels; Based on the spatial dimensions of the image data and the positional mapping relationship of the initial feature map, coordinate alignment is performed on the numerically adjusted edge regions, and the aligned edge regions are mapped to the image data to obtain a semantic segmentation map.

[0027] In this embodiment, S3 specifically includes: S31. Spatially align the feature representation set and the region label map to establish a mapping relationship between the feature representation set and the region label map, and construct joint feature data containing feature representations and region labels; S32. Based on joint feature data, the Mask R-CNN algorithm is used to perform candidate region generation operations on the equipment and instrument region, data point region and causal logic region respectively, forming a candidate region set that corresponds one-to-one with the three types of regions. S33. Perform boundary regression operation on each candidate region to determine the corresponding graphic boundary range in the engineering drawing image, specifically including: Perform a region cropping operation on the boundary range of each candidate region in the image data to extract the image sub-region corresponding to the boundary of the candidate region; Perform multi-scale convolution operations on image sub-regions to extract boundary direction features and calculate the response intensity and gradient change rate in each boundary direction; Based on the gradient change rate, the coordinate offsets of the upper and lower boundaries and the left and right boundaries are calculated respectively. The coordinate offsets are then numerically superimposed with the boundary coordinates of the corresponding candidate regions to obtain the boundary correction results. The overlap between the boundary correction result and the corresponding region in the region label map is compared. If the overlap is lower than the preset overlap threshold, the boundary range is iteratively refined, the receptive field range of the convolution kernel is updated, and the multi-scale convolution, boundary correction and overlap comparison operations are repeated. Candidate regions with an overlap degree greater than or equal to the overlap threshold are marked as boundary localization completed, and the corresponding boundary correction results are used as the graphic boundary range. S34. Within the boundary of the graphic, perform attribute text recognition on the joint feature data, extract the attribute text, and bind it to the corresponding candidate region. S35. Perform spatial location encoding on all candidate regions to generate spatial marker vectors for identifying the two-dimensional coordinate positions of the candidate regions. S36. According to the corresponding category labels in the area label map, summarize the three types of candidate area sets that have completed boundary positioning, attribute identification and spatial coding, and construct the drawing element set.

[0028] In this embodiment, S32 specifically includes: S321. Select the pixel range corresponding to the equipment and instrument region, data point region and causal logic region in the joint feature data, perform multi-scale sliding window scanning operation on the feature representation within the pixel range, and extract a fixed-size set of image segments. S322. Perform two-dimensional convolution and batch normalization operations on each image segment in the image segment set to generate a convolution response map, and compare the convolution response map with the region label map pixel by pixel. S323. Compare the image segments that match the three types of region labels in the comparison results, perform bounding box coordinate prediction operation, and determine the candidate boundary range of the image segment in the image data by calculating the boundary offset of the image segment in the horizontal and vertical directions. S324. Perform a fixed-region convolution operation within the candidate boundary range, calculate the feature mean and gradient distribution, and perform a category probability scoring operation based on the annotation information in the region label map. S325. Perform an overlapping region judgment operation on all image segments with candidate boundary ranges, compare image segments with a boundary overlap greater than a preset overlap threshold, retain the image segment with higher category probability score as candidate region, and form three candidate region sets corresponding to each of the three types of regions.

[0029] In this embodiment, S4 specifically includes: S41. Select the drawing elements labeled as equipment and instrument area and data point area from the drawing element set to form equipment and instrument element set and data point element set respectively; S42. Extract the tag text features, spatial location features, and area category features of each drawing element from the equipment and instrument element set and the data point element set respectively, and construct the equipment feature vector and the data point feature vector respectively according to the preset feature order. S43. Using the Hungarian algorithm, the feature vectors of each device are matched with the tag features in the existing tag database to generate a set of matching results between the device instrument element set and the tag database, specifically including: Read the corresponding tag text features, spatial location features, and region category features from the feature vector of each device, and perform character normalization processing on the tag text features to obtain a standard feature sequence; Extract all tag feature sets from the tag library, perform character normalization processing on the tag text in each tag feature, and perform unified coordinate system transformation processing on the spatial location information in the tag features to form a standard tag feature set; Based on the standard feature sequence and the standard position number feature set, the text difference value and spatial position difference value between each standard feature vector and the standard position number feature are calculated one by one, and the two difference values ​​are combined according to the preset weight to construct the matching cost matrix; The Hungarian algorithm is used to perform row and column normalization, zero element marking, and minimum cost path search operations on the matching cost matrix to calculate the optimal matching result. The optimal matching result that satisfies the one-to-one constraint is recorded as the valid matching result, and the results are summarized to form a set of matching results. S44. Using the Hungarian algorithm, the feature vectors of each data point are matched with the feature vectors of the existing OPC point database to generate a set of matching results between the data point feature set and the OPC point database. S45. Perform consistency verification on the two matching result sets, record the matching results that meet the preset conditions as valid mapping relationships, and summarize them to form a mapping relationship table; S47. Mark drawing elements that do not form a valid mapping relationship in the equipment and instrument element set and data point element set as abnormal data.

[0030] In this embodiment, the row and column normalization, zero element marking, and minimum cost path search operations specifically include: Perform a row minimum value extraction operation on each row of the matching cost matrix, take the minimum value in each row as the row baseline value of that row, and subtract the row baseline value from each element in that row to obtain the row normalization matrix; Perform a column minimum value extraction operation on each column of the row normalization matrix, take the minimum value in each column as the column reference value, and subtract the column reference value from each element in the column to obtain the normalized cost matrix; In the normalized cost matrix, scan all positions where the value is equal to zero, and mark the corresponding row index and column index to construct a set of zero element positions; Based on the set of zero element positions, determine whether there exists a combination of zero elements in different rows and columns that covers all rows or all columns of the matrix; If none exists, record the row and column indices not covered by zero elements, calculate the minimum non-zero value in the uncovered region, subtract the minimum non-zero value from the corresponding elements of all uncovered rows, and add the minimum non-zero value to the corresponding elements of all covered columns. Iterate and update the normalized cost matrix until there is a combination of zero elements that meets the conditions, and construct the minimum cost path corresponding to the combination of zero elements as the optimal matching result.

[0031] In this embodiment, S5 specifically includes: S51. Select the drawing elements marked as causal logic areas from the drawing element set to form a causal logic element set; S52. Based on the causal logic element set, a conditional random field is used to perform sequence modeling operations on the graphical structure and attribute text within the causal logic region. The causal logic elements are labeled with states according to preset state transition rules, and triggering conditions and execution results are identified and distinguished. Specifically, this includes: Extract graphical boundaries, attribute text, and spatial order information from each causal logic element to construct a causal logic sequence; Construct directed connections based on the order of drawing reading in the causal logic sequence, and generate an initial connection diagram based on the spatial position of each causal logic element; Context encoding is performed on the initial connection graph to extract the graphical pattern features and textual semantic features of each causal logical element in the adjacent nodes before and after it, and then they are fused to form context features. Input the context features into the conditional random field, and perform state probability estimation operations on the causal logic elements corresponding to each context feature according to the preset state set; Based on the probability estimation results, a state label is assigned to each causal logic element, marking it as a triggering condition or execution result; S53. Based on the status labeling results, perform causal relationship pairing operations on the triggering conditions and execution results to construct one-to-many or many-to-one associations between the triggering conditions and execution results, and generate a set of causal relationships. S54. Align the causal relationship set with the mapping relationship table, and map the triggering conditions and execution results to the corresponding equipment, instruments or data points in the drawing elements to form a causal relationship record with equipment identifiers and data point identifiers. S55. Perform a structural consistency check operation on the causal relationship records, sort the triggering conditions and execution results according to the preset logical order, and summarize the causal relationship records that pass the check to generate a causal table.

[0032] In this embodiment, S6 specifically includes: S61. Extract four types of features from each abnormal data: attribute text, spatial location, semantic label, and structural relationship, and construct an abnormal feature set; S62. Based on the set of abnormal features, construct multiple random trees using the isolated forest algorithm and perform anomaly scoring operations; S63. Correspond to each abnormal score with the causal relationship record in the causal table to map the causal logic characteristics corresponding to the abnormal data. S64. Perform density similarity clustering on all causal logical features, construct a causal clustering relationship graph between abnormal data and divide it into associated clusters to form multiple abnormal cluster sets; S65. For each abnormal data in the abnormal cluster set, combine the corresponding abnormal score with the distance to the cluster center, perform an abnormal type determination operation, assign an abnormal type label to the abnormal data according to the preset abnormal classification rules, and summarize to form a set to be corrected containing abnormal type labels.

[0033] In this embodiment, S62 specifically includes: S621. In the set of abnormal features, perform feature dimension normalization operation on each abnormal feature, and use the normalization result as the training sample set. S622. Set the maximum tree depth and the minimum sample splitting threshold, perform random feature selection and random sample partitioning on the training sample set, and recursively generate multiple random tree structures with different splitting paths. S623. For each anomalous feature in each random tree, calculate the path depth of the anomalous feature in the random tree and convert the path depth into a local anomaly score. S624. Average the local anomaly scores of all random trees to obtain the global anomaly score for each anomaly feature. S625. Bind the global anomaly score to the corresponding anomaly data to construct an anomaly score set containing the score results.

[0034] In this embodiment, S7 specifically includes: S71. Integrate the mapping relationship table, cause-effect table, and set to be corrected to establish a unique index for each drawing element in the drawing element set and associate it with the corresponding attribute text, spatial location, and semantic category information to generate an interactive element set. S72. Set blue jump behavior indicators for the drawing elements corresponding to the data point areas in the interactive element set, so that clicking the blue indicator will jump to the corresponding drawing location. S73. Set orange labels for the drawing elements corresponding to the equipment and instrument area in the interactive element set. The orange labels are used for static display and attribute prompts in the visual interactive interface. S74. Set a green double-click trigger icon for the drawing elements corresponding to the causal logic area in the interactive element set, so as to realize the interactive response of double-clicking the green icon to pop up the causal table. S75. The set of interactive elements associated with color is fused with the image data using spatial coordinates to generate a preprocessed drawing result with interactive functions.

[0035] refer to Figure 3 An image understanding-based engineering drawing data preprocessing system includes: The feature extraction module is used to acquire image data of the engineering drawings to be processed, and to extract features using a convolutional neural network to form a feature representation set; The semantic segmentation module is used to perform semantic segmentation on image data using the DeepLabv3+ model, and to label three types of regions—equipment, data points, and causal logic—according to preset classification rules, generating region label maps. The joint recognition module is used to perform joint recognition operations on the feature representation set and the region label map using the Mask R-CNN algorithm, extracting the graphic boundaries, attribute text and spatial location of the three types of regions, and constructing a set of drawing elements. The matching and verification module is used to extract the drawing elements of equipment instruments and data point areas, and use the Hungarian algorithm to match and verify them with the existing tag number library and OPC point library to build a mapping relationship table, and mark the drawing elements that fail to match as abnormal data. The structure parsing module is used to perform structure parsing operations on the drawing elements of the causal logic region through a conditional random field, extract the triggering conditions and execution results, and generate a causal table by combining the mapping relationship table. The anomaly classification module is used to perform scoring and causal clustering operations on anomaly data using the Isolation Forest algorithm combined with a causal table, to identify anomaly types and form a set of data to be corrected containing anomaly type labels. The fusion and association module is used to merge the mapping relationship table, the cause-effect table, and the set to be corrected, and perform color association operations to obtain interactive drawing preprocessing results.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical engineering drawing data digitization process. This scenario involves a complete set of P&ID drawings, which include equipment symbols, instrument nodes, data acquisition points, and control logic interlocking information. The drawings are characterized by a large number of drawings, complex graphic structures, and interwoven logical relationships.

[0037] Traditional processing methods require engineers to analyze and manually annotate the drawings page by page, and then manually compare and supplement data with external databases. This is inefficient and prone to errors such as incorrect equipment numbering, missing data points, or incomplete identification of causal relationships due to oversights, which seriously affects the accuracy of subsequent system modeling and control strategy integration.

[0038] In this embodiment, the image understanding-based engineering drawing data preprocessing method proposed in this invention is used to perform comprehensive preprocessing operations on the set of P&ID drawings.

[0039] First, the system acquires drawing images using a high-resolution scanning device. Then, it extracts image features using a convolutional neural network and performs semantic segmentation on the drawing images using a DeepLabv3+ model, identifying three types of graphic regions: equipment and instrument areas, data point areas, and causal logic areas. The system further utilizes the Mask R-CNN algorithm to complete bounding box regression and attribute text recognition, forming a structured set of drawing elements. After recognition, the system automatically matches equipment tag numbers and data point names with the tag number database and OPC point database, respectively, establishing mapping relationships, and marking elements that fail to match as abnormal data.

[0040] Subsequently, the system performs structural parsing on the identified causal logic regions, automatically extracting triggering conditions and execution results, constructing a causal relationship table, and associating it with matched devices and data points. For data marked as anomalous, the system uses the Isolation Forest algorithm combined with the causal table information for scoring and clustering, automatically identifying and classifying suspected error types. All processing results are presented using a color-coding mechanism: device / instrument objects are marked in orange, data point objects are marked in blue (clicking them will take you to the corresponding drawing area), and causal relationship objects are marked in green (expanding to display the logical chain details).

[0041] After completing all processing, the system outputs an interactive drawing data file and compares the processing results with those obtained through manual verification. The table below shows the comparison data from the actual processing: Table 1 Comparison of AI preprocessing and manual processing effects of engineering drawings

[0042] As can be seen from the data in Table 1, this invention demonstrates significant advantages in processing speed, recognition accuracy, and intelligent interactivity. In terms of processing efficiency, this invention can reduce the average manual analysis time per drawing from over 9 minutes to less than 1.5 minutes, greatly improving overall processing efficiency.

[0043] In terms of recognition capabilities, the present invention outperforms manual processing in both device recognition accuracy and data point matching rate, with a particularly significant improvement in causal logic pair recognition.

[0044] Furthermore, the system automatically classifies and locates inconsistencies in drawings through anomaly detection and cluster analysis, providing a basis for subsequent manual review and correction, and effectively improving the integrity and reliability of drawing data. Through a color-coding mechanism, users can quickly navigate and trace logical chains, greatly enhancing the interactivity and readability of drawing data within the operations and maintenance platform.

[0045] In summary, this invention successfully solves the problems of low efficiency, poor accuracy, and lack of interactive expression in traditional P&ID drawing data processing in practical applications, and has significant engineering practical value and promotion potential.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for preprocessing engineering drawing data based on image understanding, characterized in that, Includes the following steps: S1. Obtain the image data of the engineering drawings to be processed, and use a convolutional neural network to extract features to form a feature representation set; S2. Perform semantic segmentation on the image data using the DeepLabv3+ model, and label the three types of regions—equipment, data points, and causal logic—according to preset classification rules, and generate region label maps. S3. The Mask R-CNN algorithm is used to perform joint recognition operations on the feature representation set and the region label map to extract the graphic boundaries, attribute text and spatial location of the three types of regions and construct the drawing element set. S4. Extract the drawing elements of the equipment, instruments and data point areas, and use the Hungarian algorithm to match and verify them with the existing tag number library and OPC point library to build a mapping relationship table. Mark the drawing elements that fail to match as abnormal data. S5. Perform structural analysis on the drawing elements of the causal logic region through a conditional random field, extract the triggering conditions and execution results, and generate a causal table by combining the mapping relationship table. S6. The isolated forest algorithm is used in conjunction with the causal table to perform scoring and causal clustering operations on the abnormal data, identify the anomaly type, and form a set to be corrected containing anomaly type labels. S7. Merge the mapping relationship table, causal table and the set to be corrected and perform color association operation to obtain an interactive drawing preprocessing result.

2. The method for preprocessing engineering drawing data based on image understanding according to claim 1, characterized in that, The tag library represents a preset database of equipment and instrument tag information; the OPC point library represents an OPC data point database used for operational status interaction; the color association operation represents marking different categories of drawing elements with specific colors and performing jump associations; and the drawing preprocessing result represents the structured data output result used for drawing understanding and engineering data interaction.

3. The method for preprocessing engineering drawing data based on image understanding according to claim 1, characterized in that, S2 specifically includes: S21. Perform size unification and boundary padding operations on the image data, and adjust the numerical range of the image color information according to a preset threshold to form standard image data, specifically including: Perform resolution adjustment on the image data, reset the image width and height to the preset standard size, and maintain the original aspect ratio of the image; Perform boundary padding on the adjustment results, and use mirror expansion to fill in missing pixels; Threshold constraint processing is performed on the pixel values ​​representing image color information in the boundary filling result. The pixel values ​​are cropped and restricted according to the preset threshold. Pixel values ​​that exceed the threshold range are mapped to a uniform numerical range to obtain standard image data. S22. Input standard image data into the DeepLabv3+ model and perform dilated convolution operation to extract contextual features at different sampling intervals and generate a semantic feature set containing multi-scale structural information. S23. Perform cross-scale feature aggregation on the semantic feature set, and perform upsampling interpolation on the aggregation result in combination with the spatial distribution of the image data to generate a semantic segmentation map with the same spatial size as the image data. S24. According to the preset classification rules, perform a category judgment operation on each pixel in the semantic segmentation map, and label the judgment result as the device instrument area, data point area or causal logic area. S25. Perform a pixel-by-pixel alignment operation between the labeled three types of regions and the image data to obtain the region label map.

4. The method for preprocessing engineering drawing data based on image understanding according to claim 3, characterized in that, The process of generating the semantic segmentation map specifically includes: In the DeepLabv3+ model, feature encoding operations are performed on standard image data to generate intermediate feature maps; By employing dilated convolution kernels with different dilation rates, context-aware operations are performed on intermediate feature maps in parallel at multiple sampling intervals to extract context features at multiple scales and form a semantic feature set. Dimension alignment and feature normalization operations are performed on the context features at each scale in the semantic feature set, and pixel-wise weighted aggregation operation is performed on the normalization results according to the preset weight coefficient set to form a fused feature map; A stepwise upsampling operation is performed on the fused feature map, and a bilinear interpolation algorithm is used to obtain an initial feature map with the same spatial size as the image data. Specifically, this includes: The fused feature map is divided into multiple upsampling stages according to a preset number of upsampling layers, and the target size of each stage is set. The bilinear interpolation algorithm is used to perform spatial size expansion operation on the fused feature map. The fused feature map is interpolated according to the target size, and the difference result is numerically smoothed according to the set numerical range to generate an upsampled feature map as the input for the next upsampling stage. Use the upsampled feature map output from the last upsampling stage as the initial feature map; Edge structure smoothing and spatial coordinate compensation operations are performed on the initial feature map to correct the jagged edge effect generated during upsampling, resulting in a semantic segmentation map.

5. The method for preprocessing engineering drawing data based on image understanding according to claim 1, characterized in that, S3 specifically includes: S31. Spatially align the feature representation set and the region label map to establish a mapping relationship between the feature representation set and the region label map, and construct joint feature data containing feature representations and region labels; S32. Based on joint feature data, the Mask R-CNN algorithm is used to perform candidate region generation operations on the equipment and instrument region, data point region and causal logic region respectively, forming a candidate region set that corresponds one-to-one with the three types of regions. S33. Perform boundary regression operation on each candidate region to determine the corresponding graphic boundary range in the engineering drawing image, specifically including: Perform a region cropping operation on the boundary range of each candidate region in the image data to extract the image sub-region corresponding to the boundary of the candidate region; Perform multi-scale convolution operations on image sub-regions to extract boundary direction features and calculate the response intensity and gradient change rate in each boundary direction; Based on the gradient change rate, the coordinate offsets of the upper and lower boundaries and the left and right boundaries are calculated respectively. The coordinate offsets are then numerically superimposed with the boundary coordinates of the corresponding candidate regions to obtain the boundary correction results. The overlap between the boundary correction result and the corresponding region in the region label map is compared. If the overlap is lower than the preset overlap threshold, the boundary range is iteratively refined, the receptive field range of the convolution kernel is updated, and the multi-scale convolution, boundary correction and overlap comparison operations are repeated. Candidate regions with an overlap degree greater than or equal to the overlap threshold are marked as boundary localization completed, and the corresponding boundary correction results are used as the graphic boundary range. S34. Within the boundary of the graphic, perform attribute text recognition on the joint feature data, extract the attribute text, and bind it to the corresponding candidate region. S35. Perform spatial location encoding on all candidate regions to generate spatial marker vectors for identifying the two-dimensional coordinate positions of the candidate regions. S36. According to the corresponding category labels in the area label map, summarize the three types of candidate area sets that have completed boundary positioning, attribute identification and spatial coding, and construct the drawing element set.

6. The method for preprocessing engineering drawing data based on image understanding according to claim 5, characterized in that, Specifically, S32 includes: S321. Select the pixel range corresponding to the equipment and instrument region, data point region and causal logic region in the joint feature data, perform multi-scale sliding window scanning operation on the feature representation within the pixel range, and extract a fixed-size set of image segments. S322. Perform two-dimensional convolution and batch normalization operations on each image segment in the image segment set to generate a convolution response map, and compare the convolution response map with the region label map pixel by pixel. S323. Compare the image segments that match the three types of region labels in the comparison results, perform bounding box coordinate prediction operation, and determine the candidate boundary range of the image segment in the image data by calculating the boundary offset of the image segment in the horizontal and vertical directions. S324. Perform a fixed-region convolution operation within the candidate boundary range, calculate the feature mean and gradient distribution, and perform a category probability scoring operation based on the annotation information in the region label map. S325. Perform an overlapping region judgment operation on all image segments with candidate boundary ranges, compare image segments with a boundary overlap greater than a preset overlap threshold, retain the image segment with higher category probability score as candidate region, and form three candidate region sets corresponding to each of the three types of regions.

7. The method for preprocessing engineering drawing data based on image understanding according to claim 1, characterized in that, S4 specifically includes: S41. Select the drawing elements labeled as equipment and instrument area and data point area from the drawing element set to form equipment and instrument element set and data point element set respectively; S42. Extract the tag text features, spatial location features, and area category features of each drawing element from the equipment and instrument element set and the data point element set respectively, and construct the equipment feature vector and the data point feature vector respectively according to the preset feature order. S43. Using the Hungarian algorithm, the feature vectors of each device are matched with the tag features in the existing tag database to generate a set of matching results between the device instrument element set and the tag database, specifically including: Read the corresponding tag text features, spatial location features, and region category features from the feature vector of each device, and perform character normalization processing on the tag text features to obtain a standard feature sequence; Extract all tag feature sets from the tag library, perform character normalization processing on the tag text in each tag feature, and perform unified coordinate system transformation processing on the spatial location information in the tag features to form a standard tag feature set; Based on the standard feature sequence and the standard position number feature set, the text difference value and spatial position difference value between each standard feature vector and the standard position number feature are calculated one by one, and the two difference values ​​are combined according to the preset weight to construct the matching cost matrix; The Hungarian algorithm is used to perform row and column normalization, zero element marking, and minimum cost path search operations on the matching cost matrix to calculate the optimal matching result. The optimal matching result that satisfies the one-to-one constraint is recorded as the valid matching result, and the results are summarized to form a set of matching results. S44. Using the Hungarian algorithm, the feature vectors of each data point are matched with the feature vectors of the existing OPC point database to generate a set of matching results between the data point feature set and the OPC point database. S45. Perform consistency verification on the two matching result sets, record the matching results that meet the preset conditions as valid mapping relationships, and summarize them to form a mapping relationship table; S47. Mark drawing elements that do not form a valid mapping relationship in the equipment and instrument element set and data point element set as abnormal data.

8. The method for preprocessing engineering drawing data based on image understanding according to claim 1, characterized in that, S5 specifically includes: S51. Select the drawing elements marked as causal logic areas from the drawing element set to form a causal logic element set; S52. Based on the causal logic element set, a conditional random field is used to perform sequence modeling operations on the graphical structure and attribute text within the causal logic region. The causal logic elements are labeled with states according to preset state transition rules, and triggering conditions and execution results are identified and distinguished. Specifically, this includes: Extract graphical boundaries, attribute text, and spatial order information from each causal logic element to construct a causal logic sequence; Construct directed connections based on the order of drawing reading in the causal logic sequence, and generate an initial connection diagram based on the spatial position of each causal logic element; Context encoding is performed on the initial connection graph to extract the graphical pattern features and textual semantic features of each causal logical element in the adjacent nodes before and after it, and then they are fused to form context features. Input the context features into the conditional random field, and perform state probability estimation operations on the causal logic elements corresponding to each context feature according to the preset state set; Based on the probability estimation results, a state label is assigned to each causal logic element, marking it as a triggering condition or execution result; S53. Based on the status labeling results, perform causal relationship pairing operations on the triggering conditions and execution results to construct one-to-many or many-to-one associations between the triggering conditions and execution results, and generate a set of causal relationships. S54. Align the causal relationship set with the mapping relationship table, and map the triggering conditions and execution results to the corresponding equipment, instruments or data points in the drawing elements to form a causal relationship record with equipment identifiers and data point identifiers. S55. Perform a structural consistency check operation on the causal relationship records, sort the triggering conditions and execution results according to the preset logical order, and summarize the causal relationship records that pass the check to generate a causal table.

9. The method for preprocessing engineering drawing data based on image understanding according to claim 1, characterized in that, S6 specifically includes: S61. Extract four types of features from each abnormal data: attribute text, spatial location, semantic label, and structural relationship, and construct an abnormal feature set; S62. Based on the set of abnormal features, construct multiple random trees using the isolated forest algorithm and perform anomaly scoring operations; S63. Correspond to each abnormal score with the causal relationship record in the causal table to map the causal logic characteristics corresponding to the abnormal data. S64. Perform density similarity clustering on all causal logical features, construct a causal clustering relationship graph between abnormal data and divide it into associated clusters to form multiple abnormal cluster sets; S65. For each abnormal data in the abnormal cluster set, combine the corresponding abnormal score with the distance to the cluster center, perform an abnormal type determination operation, assign an abnormal type label to the abnormal data according to the preset abnormal classification rules, and summarize to form a set to be corrected containing abnormal type labels.

10. An image-understanding-based engineering drawing data preprocessing system, comprising executing the image-understanding-based engineering drawing data preprocessing method according to any one of claims 1 to 9, characterized in that, include: The feature extraction module is used to acquire image data of the engineering drawings to be processed, and to extract features using a convolutional neural network to form a feature representation set; The semantic segmentation module is used to perform semantic segmentation on image data using the DeepLabv3+ model, and to label three types of regions—equipment, data points, and causal logic—according to preset classification rules, generating region label maps. The joint recognition module is used to perform joint recognition operations on the feature representation set and the region label map using the Mask R-CNN algorithm, extracting the graphic boundaries, attribute text and spatial location of the three types of regions, and constructing a set of drawing elements. The matching and verification module is used to extract the drawing elements of equipment instruments and data point areas, and use the Hungarian algorithm to match and verify them with the existing tag number library and OPC point library to build a mapping relationship table, and mark the drawing elements that fail to match as abnormal data. The structure parsing module is used to perform structure parsing operations on the drawing elements of the causal logic region through a conditional random field, extract the triggering conditions and execution results, and generate a causal table by combining the mapping relationship table. The anomaly classification module is used to perform scoring and causal clustering operations on anomaly data using the Isolation Forest algorithm combined with a causal table, to identify anomaly types and form a set of data to be corrected containing anomaly type labels. The fusion and association module is used to merge the mapping relationship table, the cause-effect table, and the set to be corrected, and perform color association operations to obtain interactive drawing preprocessing results.

Images