Electric power design drawing identification and review method and system

Abstract

The invention discloses an electric power design drawing identification and review method and system, and the method comprises the following steps: S1, obtaining an electric power design drawing, and constructing a primitive library and a rule library; s2, performing primitive recognition on the electric power design drawing based on an edge recognition method and a deep learning method to obtain a corresponding digital image; s3, in OCR recognition, a weak supervised learning method is used, deep learning and a computer vision processing method are combined, texts in the electric power design drawing are recognized, and text information is obtained; s4, comparing the digital image with a primitive library to obtain position and type information of each electrical primitive in the electric power design drawing; and S5, comparing the position, type information and text information of each electrical primitive with a rule base to obtain an identification result, post-processing the obtained identification result, and judging and modifying the result through priori knowledge.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power drawing review, and more particularly to a method and system for identifying and reviewing electric power design drawings. Background Art

[0002] Drawing review refers to the review of the contents of construction drawings involving public interests, public safety and mandatory standards for engineering construction in accordance with relevant laws and regulations, aiming to improve the quality of design and discover possible design or construction problems in advance. During the review process, the relevant participating units and individuals will review the design drawings one by one according to the review points stipulated by the state, local or enterprise.

[0003] Currently, there are two main ways to review drawings: offline review and online review. In the offline review method, all design drawings that need to be reviewed will be printed out using professional equipment before the review. Relevant participating units and individuals will gather together at the agreed time and place in the form of a meeting to review the printed design drawings, and the reviewers must be physically together. In the online review method, all design drawings that need to be reviewed will be uploaded online before the review. Relevant participating units and individuals can review online drawings in a relatively flexible way, and the reviewers do not have to be physically together.

[0004] However, the two existing review methods still have defects. The offline review method has high requirements for personnel concentration, and it is necessary to search for the content to be reviewed in physical books or documents according to the profession and type of each drawing, so there are problems of high cost, long time and low efficiency. The online review method has solved the efficiency problem of offline review to a certain extent, but it is still necessary to find the chapters related to the drawings in the electronic list of review points.

[0005] The publication number is CN111047274A, and the name is "Method and system for reviewing engineering design drawings". It establishes a first association relationship between drawing feature information and review content, and obtains the pending feature information of the pending drawing and historical matching information of drawings of the same type; based on the pending feature information, the first association relationship and the historical matching information are called to generate a pending review list and display it; in response to a confirmation request for the pending review list, a qualified review list is generated for reviewers to review the pending drawings.

[0006] The publication number is CN111859503A, and the name is a drawing review method, electronic device and graphics server, the method comprising: obtaining a first operation instruction for an electronic drawing, and forming a target local area on the electronic drawing according to the first operation instruction, the target local area being used to represent a local area in the electronic drawing that does not meet the standard requirements in the architectural design specification; determining at least one first graphic element of the target local area; generating a review rule list according to the graphic element type of at least one first graphic element; obtaining a second operation instruction for the review rule list, and determining a target preset review rule from the review rule list according to the second operation instruction; determining a review result of at least one first graphic element according to the target preset review rule.

[0007] Therefore, the current review method mainly focuses on how to conduct online review and partial review. Although it eliminates the need to review physical materials, it is still necessary to find the review points related to the drawings in the electronic list. It requires a good electronic document management system and network environment support, and requires all relevant personnel to gather at the same time and place. It has poor flexibility, and the costs of printing drawings, personnel transportation and meeting time are high, and the review cycle is long.

[0008] Therefore, how to provide a method and system for automatically reviewing electrical drawings is an urgent problem that those skilled in the art need to solve. Summary of the invention

[0009] In view of this, the present invention provides a method and system for identifying and reviewing electric power design drawings, which aims to solve the problem of online review and, to a certain extent, the efficiency problem of offline review. However, it still needs to solve the technical problem of finding the chapters related to the drawings in the electronic list of review points.

[0010] In order to achieve the above object, the present invention adopts the following technical solution:

[0011] A method for identifying and reviewing electric power design drawings comprises the following steps:

[0012] S1. Obtain power design drawings and build a graphic element library and rule library;

[0013] S2. Performing graphic element recognition on the electric power design drawings based on edge recognition methods and deep learning methods to obtain corresponding digital images;

[0014] S3. In OCR recognition, the weakly supervised learning method is used, combined with deep learning and computer vision processing methods, to recognize the text in the power design drawings and obtain text information;

[0015] S4, comparing the digital image with the graphic element library to obtain the location and type information of each electrical graphic element in the power design drawing;

[0016] S5. Compare the position, type information and text information of each electrical graphic element with the rule base to obtain the recognition result, post-process the obtained recognition result, and judge and modify the result through prior knowledge.

[0017] Further, the graphic element library constructed in S1 specifically includes a grounding switch, a current transformer, a disconnector, a lightning arrester, a circuit breaker, a three-phase live display, a bushing, a fast grounding switch, a voltage transformer, a fast disconnector, and a cable terminal;

[0018] The rule base specifically includes various index parameters and constraint relationships of primary equipment, secondary equipment, and intelligent secondary equipment.

[0019] Furthermore, the S2 is specifically:

[0020] S21, preprocessing the acquired power design drawings, specifically graying the power design drawings, smoothing and denoising the grayed images, binarizing the denoised grayed images to convert them into binary images, performing morphological operations on the binary images, including corrosion or expansion, segmenting the processed binary images to obtain multiple region images, and post-processing the multiple region images, including removing small objects, filling holes, and smoothing boundaries;

[0021] S22, using Faster RCNN network architecture for electrical component identification and training;

[0022] S23. Through the deep learning network, use the overall attractive wireframe parser to vectorize the refined straight line structure diagram to obtain a preliminary vectorization result, improve the preliminary vectorization result, screen the edges, cluster the points, remove duplicate and redundant information, and obtain a standard vector diagram result.

[0023] Furthermore, the S21 is specifically:

[0024] Grayscale the power design drawings:

[0025] I gray =0.2989×R+0.5870×G+0.1140×B

[0026] Among them, R, G, B are the red, green, and blue channel values ​​of the image;

[0027] Smooth and denoise the grayscale image based on Gaussian filtering:

[0028] Determine the size and standard deviation of the Gaussian kernel, create the Gaussian kernel, use the Gaussian function to calculate the value of each element in the kernel, calculate the Gaussian value at each position in the kernel based on the standard deviation,

[0029]

[0030] Where x, y are the positions relative to the center of the kernel, σ is the standard deviation, and the normalized Gaussian kernel is completed;

[0031] Slide the Gaussian kernel onto each pixel of the image, calculate the weighted sum of all pixel values ​​under the kernel, and get the new pixel value. For each pixel (i, j), the new pixel value is:

[0032]

[0033] Where k is half of the kernel size, G(m,n) is the value of the Gaussian kernel, and I(i+m,j+n) is the pixel value of the original image;

[0034] The grayscale histogram of the new image is calculated, and all possible thresholds are traversed based on the Otsu algorithm. For each threshold, the image is divided into a foreground part and a background part. For each threshold, the intra-class variance and the inter-class variance of the foreground and the background are calculated. The intra-class variance is used to judge the deviation of the pixel value inside the foreground or the background from its average value, and the inter-class variance is used to judge the separation between the foreground and the background. The threshold that maximizes the inter-class variance is obtained, and all pixels in the image that are less than the threshold are set to black, and all pixels that are greater than or equal to the threshold are set to white; morphological operations are performed on the binary image, including corrosion or expansion, and the processed binary image is segmented to obtain multiple regional images. The multiple regional images are post-processed, including removing small objects, filling holes, and smoothing boundaries.

[0035] Furthermore, the S22 is specifically:

[0036] The VGG16 network is used as the basic network, and the deep learning method is used to extract features of power grid graph elements. The shared convolution and RPN weights are initialized using a large data set pre-trained model. The RPN is trained end-to-end to generate RegionProposals. The same pre-trained model is used to initialize the shared convolution, lock the RPN weights, and train the RCNN network with the Proposals obtained by RPN. The loss values ​​of the classification network and regression network are calculated, and backpropagation is performed to optimize the network parameters. The trained shared convolution and RCNN are used to fix the shared convolution layer, continue to train the RPN, fine-tune the RPN, and use the tuned shared convolution and RPN to continue to train and fine-tune the RCNN. After training, the final complete Faster-RCNN model is obtained, and electrical components are identified through the final complete Faster-RCNN model.

[0037] Furthermore, the S3 is specifically:

[0038] The pre-trained text detection model is used to detect the image to be recognized. By detecting the area of ​​each text character and the relationship between the text characters, the text area box at the entire word level is predicted. For vertical text, character cutting is used to obtain single character text, and for horizontal text, text lines are used directly. Recognition is performed through the CNN+BiLSTM+CTC model to obtain text information.

[0039] Further, specifically:

[0040] Using the CNN algorithm, the feature map of the image is extracted through the convolution layer. The RPN network generates candidate regions on the feature map, and the features extracted from the candidate regions on the feature map are pooled to obtain a fixed-size feature map for classification and bounding box regression. The area of ​​each text character in the image is predicted, and based on the relationship between these areas and the text characters, the text area box at the entire word level is predicted. The text detection model uses weak supervision training to learn character features for each word-level annotation area. For vertical text, character cutting technology is used to obtain single character text. The blank space between characters on the vertical projection of the image row is used to segment the single character image. For horizontal text, the text row is used directly to output the final detection result, including the bounding box coordinates of each word or text row, as well as the text content.

[0041] Furthermore, the S4 is specifically:

[0042] Establish a vector template for power components, use correlation measurement to identify the black and white image and the components in it, and perform template matching; use target detection technology to obtain the location and type information of each electrical element in the wiring diagram and save it to an XML file;

[0043] The contour tracking algorithm is used to extract the outer contour of the connection line of the wiring diagram after the preprocessing operation. According to the acquired position of each electrical graphic element and the extracted outer contour coordinate information of each connection line, the connection relationship between each connection line and the electrical graphic element is analyzed.

[0044] An electric power design drawing recognition and review system includes a drawing acquisition module, a graphic element and rule base module, a recognition module, and a comparison module;

[0045] The drawing acquisition module is used to acquire power design drawings and perform preprocessing;

[0046] The graphic element and rule library module is used to store graphic elements and review rules;

[0047] The recognition module is used to extract graphic element information and text information from the power design drawings;

[0048] The comparison module is used to compare the graphic element information and text information raised by the recognition module with the stored graphic elements and the review rules to obtain the review result.

[0049] Furthermore, a data maintenance module is provided, which is used to ensure data consistency throughout the system, avoid data redundancy and conflict, improve data accuracy through verification and validation mechanisms, regularly update data, clean up obsolete or useless data, monitor data usage and performance indicators, and generate data reports.

[0050] The above technical solution includes at least the following technical effects:

[0051] The present invention can automatically identify and review power design drawings, reduce manual intervention, and improve work efficiency. It uses deep learning and computer vision technology to improve the accuracy of graphic element recognition and text recognition, reduce human errors and omissions, and improve the quality of design drawings. By constructing a graphic element library and a rule library, it achieves standardization of recognition and review, ensuring that the design drawings meet industry standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0053] Figure 1 The present invention is a schematic diagram of the process structure of a method for identifying and reviewing electric power design drawings. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] See attached Figure 1 The embodiment of the present invention discloses a method for identifying and reviewing electric power design drawings, comprising the following steps:

[0056] S1. Obtain power design drawings and build a graphic element library and rule library;

[0057] S2. Performing graphic element recognition on the electric power design drawings based on edge recognition methods and deep learning methods to obtain corresponding digital images;

[0058] S3. In OCR recognition, the weakly supervised learning method is used, combined with deep learning and computer vision processing methods, to recognize the text in the power design drawings and obtain text information;

[0059] S4, comparing the digital image with the graphic element library to obtain the location and type information of each electrical graphic element in the power design drawing;

[0060] S5. Compare the position, type information and text information of each electrical graphic element with the rule base to obtain the recognition result, post-process the obtained recognition result, and judge and modify the result through prior knowledge.

[0061] To further optimize the above technical solution, the graphic element library constructed in S1 specifically includes grounding switch, current transformer, disconnector, lightning arrester, circuit breaker, three-phase live display, bushing, fast grounding switch, voltage transformer, fast disconnector, and cable terminal;

[0062] The rule base specifically includes various indicator parameters and constraint relationships of primary equipment, secondary equipment, and intelligent secondary equipment.

[0063] To further optimize the above technical solution, S2 is specifically:

[0064] S21, preprocessing the acquired power design drawings, specifically graying the power design drawings, smoothing and denoising the grayed images, binarizing the denoised grayed images to convert them into binary images, performing morphological operations on the binary images, including corrosion or expansion, segmenting the processed binary images to obtain multiple region images, and post-processing the multiple region images, including removing small objects, filling holes, and smoothing boundaries;

[0065] S22, using FasterRCNN network architecture for electrical component identification and training;

[0066] S23. Through the deep learning network, use the overall attractive wireframe parser to vectorize the refined straight line structure diagram to obtain a preliminary vectorization result, improve the preliminary vectorization result, screen the edges, cluster the points, remove duplicate and redundant information, and obtain a standard vector diagram result.

[0067] To further optimize the above technical solution, S21 is specifically as follows:

[0068] Grayscale the power design drawings:

[0069] I gray =0.2989×R+0.5870×G+0.1140×B

[0070] Among them, R, G, B are the red, green, and blue channel values ​​of the image;

[0071] Smooth and denoise the grayscale image based on Gaussian filtering:

[0072] Determine the size and standard deviation of the Gaussian kernel, create the Gaussian kernel, use the Gaussian function to calculate the value of each element in the kernel, calculate the Gaussian value at each position in the kernel based on the standard deviation,

[0073]

[0074] Where x, y are the positions relative to the center of the kernel, σ is the standard deviation, and the normalized Gaussian kernel is completed;

[0075] Slide the Gaussian kernel onto each pixel of the image, calculate the weighted sum of all pixel values ​​under the kernel, and get the new pixel value. For each pixel (i, j), the new pixel value is:

[0076]

[0077] Where k is half of the kernel size, G(m,n) is the value of the Gaussian kernel, and I(i+m,j+n) is the pixel value of the original image;

[0078] Calculate the grayscale histogram of the new image, traverse all possible thresholds based on the Otsu algorithm, for each threshold, divide the image into foreground and background, for each threshold, calculate the intra-class variance and inter-class variance of the foreground and background, the intra-class variance is used to judge the deviation of the pixel value inside the foreground or background from its average value, and the inter-class variance is used to judge the separation between the foreground and background, obtain the threshold that maximizes the inter-class variance, set all pixels in the image that are less than the threshold to black, and set all pixels that are greater than or equal to the threshold to white; perform morphological operations on the binary image, including corrosion or expansion, segment the processed binary image to obtain multiple regional images, and post-process the multiple regional images, including removing small objects, filling holes, and smoothing boundaries.

[0079] To further optimize the above technical solution, S22 is specifically:

[0080] The VGG16 network is used as the basic network, and the deep learning method is used to extract features of power grid graph elements. The shared convolution and RPN weights are initialized using a large data set pre-trained model. The RPN is trained end-to-end to generate RegionProposals. The same pre-trained model is used to initialize the shared convolution, lock the RPN weights, and train the RCNN network with the Proposals obtained by RPN. The loss values ​​of the classification network and regression network are calculated, and backpropagation is performed to optimize the network parameters. The trained shared convolution and RCNN are used to fix the shared convolution layer, continue to train the RPN, fine-tune the RPN, and use the tuned shared convolution and RPN to continue to train and fine-tune the RCNN. After training, the final complete Faster-RCNN model is obtained, and electrical components are identified through the final complete Faster-RCNN model.

[0081] To further optimize the above technical solution, S3 is specifically:

[0082] The pre-trained text detection model is used to detect the image to be recognized. By detecting the area of ​​each text character and the relationship between the text characters, the text area box at the entire word level is predicted. For vertical text, character cutting is used to obtain single character text, and for horizontal text, text lines are used directly. Recognition is performed through the CNN+BiLSTM+CTC model to obtain text information.

[0083] To further optimize the above technical solution, specifically:

[0084] Using the CNN algorithm, the feature map of the image is extracted through the convolution layer. The RPN network generates candidate regions on the feature map, and the features extracted from the candidate regions on the feature map are pooled to obtain a fixed-size feature map for classification and bounding box regression. The area of ​​each text character in the image is predicted, and based on the relationship between these areas and the text characters, the text area box at the entire word level is predicted. The text detection model uses weak supervision training to learn character features for each word-level annotation area. For vertical text, character cutting technology is used to obtain single character text. The blank space between characters on the vertical projection of the image row is used to segment the single character image. For horizontal text, the text row is used directly to output the final detection result, including the bounding box coordinates of each word or text row, as well as the text content.

[0085] To further optimize the above technical solution, S4 is specifically:

[0086] Establish a vector template for power components, use correlation measurement to identify the black and white image and the components in it, and perform template matching; use target detection technology to obtain the location and type information of each electrical element in the wiring diagram and save it to an XML file;

[0087] The contour tracking algorithm is used to extract the outer contour of the connection line of the wiring diagram after the preprocessing operation. According to the acquired position of each electrical graphic element and the extracted outer contour coordinate information of each connection line, the connection relationship between each connection line and the electrical graphic element is analyzed.

[0088] An electric power design drawing recognition and review system includes a drawing acquisition module, a graphic element and rule base module, a recognition module, and a comparison module;

[0089] The drawing acquisition module is used to acquire power design drawings and perform preprocessing;

[0090] The graphic element and rule library module is used to store graphic elements and review rules;

[0091] The recognition module is used to extract graphic element information and text information from the power design drawings;

[0092] The comparison module is used to compare the graphic element information and text information raised by the recognition module with the stored graphic elements and review rules to obtain the review result.

[0093] In order to further optimize the above technical solution, a data maintenance module is also set up. The data maintenance module is used to ensure the consistency of data in the entire system, avoid data redundancy and conflict, ensure the accuracy of data through verification and validation mechanisms, update data regularly, clean up outdated or useless data, monitor data usage and performance indicators, and generate data reports.

[0094] Step S1: Obtain power design drawings and build a graphic element library and rule library

[0095] 1.1 Obtaining drawings

[0096] Collect electronic files of power design drawings and support multiple formats such as DWG, PDF, JPEG, etc.

[0097] Make sure the drawings are clear and of high enough resolution to facilitate subsequent processing.

[0098] 1.2 Element Library Construction

[0099] Define the categories and properties of electrical elements according to the standards and specifications of the electrical industry.

[0100] Create a library of graphics elements, including but not limited to grounding switches, current transformers, disconnectors, etc.

[0101] Define a unique identifier and set of properties for each primitive.

[0102] 1.3 Rule Base Construction

[0103] Build a rule base based on safety specifications and operating standards for power design.

[0104] The rule library contains layout rules, connection rules, and parameter constraints for electrical entities.

[0105] The rule base should support the addition, deletion and modification of rules to adapt to the update of specifications.

[0106] (1) Equipment parameter rule base

[0107] The various index parameters and constraints of primary equipment, secondary equipment, and intelligent secondary equipment are organized into structured data and solidified into the rule base. The specific equipment types are as follows:

[0108] Primary equipment: transformers, voltage transformers, current transformers, circuit breakers, fuses, shunt capacitors, switchgear, busbars, a total of 26,324 equipment types.

[0109] Secondary equipment: relay protection and automatic devices, power supply systems, automation systems and equipment, electricity consumption information collection, measurement and control and online monitoring systems, safe and stable automatic control devices, low-voltage panels (cabinets), boxes, operation and inspection monitoring systems, power grid dispatching technical support systems, distribution automation, data network equipment, a total of 2,152 types of equipment.

[0110] Intelligent secondary equipment: intelligent substation relay protection and automatic devices, intelligent substation automation systems and equipment, intelligent substation relay protection devices, intelligent substation electricity consumption information collection, intelligent substation power supply systems, intelligent substation measurement and control and online monitoring systems, intelligent substation safety and stability automatic control devices, etc., a total of 232 types of equipment.

[0111] (2) Expert Experience Rule Base

[0112] Expert experience is organized, analyzed, and quantified, and knowledge extraction and knowledge modeling are performed on this basis to form structured knowledge data. The quantifiable part needs to support structured queries. The expert review experience includes: preferences for equipment usage scenarios, matching rules for various equipment models, experience values ​​for equipment parameter selection, and usage taboos for various scenarios.

[0113] Create corresponding field description information for the quantifiable experience of experts, establish structured data based on the actual content, and the non-quantifiable part needs to be graphed into knowledge, that is, the descriptive language of experts is converted into elements such as ontology, object, relationship predicate and inheritance relationship, and the attributes of each element are determined.

[0114] Step S2: Performing graphic element recognition on the power design drawing based on edge recognition method and deep learning method to obtain the corresponding digital image

[0115] 2.1 Preprocessing

[0116] Grayscale: Convert color drawings into grayscale images to reduce computational complexity.

[0117] Denoising: Apply algorithms such as Gaussian filtering to reduce image noise and improve image quality.

[0118] Binarization: Use the Otsu algorithm to automatically determine the optimal threshold and convert the grayscale image into a binary image.

[0119] Morphological operation: perform erosion and dilation operations on binary images to highlight primitive features.

[0120] 2.2 Element Recognition

[0121] Faster R-CNN recognition: Use the pre-trained Faster R-CNN model to perform primitive recognition on the pre-processed images.

[0122] Model training: If necessary, fine-tune the model using annotated drawing data to adapt to specific drawing styles and primitive features.

[0123] Result Extraction: Extract the location and category information of entities from the model output.

[0124] The identifiable graphic elements include: grounding switch, current transformer, disconnector, lightning arrester, circuit breaker, three-phase live display, bushing, fast grounding switch, voltage transformer, fast disconnector, cable terminal.

[0125] Step S3: In OCR recognition, a weakly supervised learning method is used, combined with deep learning and computer vision processing methods, to recognize text in the power design drawings and obtain text information;

[0126] By applying target detection technology and OCR technology, the system can intelligently identify the graphics and text content in CAD drawings and locate their coordinate positions. The system performs consistency data comparison and basic normative review based on expert review experience and review points. After data annotation and model training of 557 wiring diagrams, the system can recognize 11 types of graphics information including disconnectors, current transformers, circuit breakers, etc., with an average graphics recognition accuracy of nearly 82%. In order to address the problem of low graphics recognition accuracy due to differences in drawing habits, the system will continue to optimize the model to adapt to more diverse graphics expressions.

[0127] 3.1 Text Detection

[0128] Use a pre-trained text detection model such as EAST or CTPN to identify text regions in images.

[0129] 3.2 Character Segmentation

[0130] For vertical text, the text line is segmented into individual characters by character segmentation technique.

[0131] For horizontal text, the text lines are directly used for recognition.

[0132] 3.3 Text Recognition

[0133] The CNN+BiLSTM+CTC model is used to identify text areas and extract text information.

[0134] Step S4: Compare the digital image with the graphic element library to obtain the location and type information of each electrical graphic element in the power design drawing

[0135] 4.1 Template Matching

[0136] The power component vector template is used to match the recognition result to determine the location and type of the graphic element.

[0137] 4.2 Object Detection Technology

[0138] The target detection technology is used to obtain the location and type information of each electrical element in the wiring diagram and save it to an XML file.

[0139] Step S5: Compare the location, type information and text information of each electrical graphic element with the rule base, obtain the recognition result, post-process the obtained recognition result, and judge and modify the result based on prior knowledge

[0140] 5.1 Rule Comparison

[0141] Compare the identified element information with the constraints in the rule base to check the compliance of the design drawings.

[0142] 5.2 Post-processing

[0143] Post-process the recognition results, such as removing small objects, filling holes, smoothing boundaries, etc.

[0144] 5.3 Result determination

[0145] The recognition results are judged and modified based on prior knowledge to ensure the accuracy of the final results.

[0146] Step S6: Data maintenance

[0147] 6.1 Data Consistency

[0148] Ensure data consistency across the entire system to avoid data redundancy and conflicts.

[0149] 6.2 Data Accuracy

[0150] Ensure data accuracy through verification and validation mechanisms.

[0151] 6.3 Data Update

[0152] Update data regularly and clean up outdated or useless data.

[0153] 6.4 Data Monitoring

[0154] Monitor data usage and performance indicators and generate data reports.

[0155] Morphological operations are a technique in image processing that operates based on the shape of an image and is mainly used for binary images. These operations usually involve a structural element (also called a kernel or brush), which is a small binary matrix that is convolved with the image to achieve a specific effect. The following are the specific processes of several common morphological operations:

[0156] 1. Erosion

[0157] Purpose: Reduce the boundaries of objects in the image, eliminate small white noise, and separate adjacent objects.

[0158] process:

[0159] Choose a structure element, usually a 3x3 or 5x5 matrix.

[0160] Align the center of the structure element with each pixel in the image.

[0161] If the structure element is completely contained within the foreground (white area), the pixel is retained; otherwise, it is turned into the background (black).

[0162] The result is that the boundaries of objects in the image shrink inwards.

[0163] 2. Dilation

[0164] Purpose: Increase the boundaries of objects in the image, fill small holes inside objects, and connect adjacent objects.

[0165] process:

[0166] Select a structural element.

[0167] Align the center of the structure element with each pixel in the image.

[0168] If the structure element intersects the foreground, the pixel is changed to foreground; otherwise, it remains unchanged.

[0169] The result is that the boundaries of objects in the image expand outward.

[0170] 3. Opening

[0171] Purpose: Erosion followed by expansion is used to eliminate small white noise, separate adjacent objects, and maintain the shape of the object.

[0172] process:

[0173] First, perform an erosion operation on the image.

[0174] Then the eroded image is dilated.

[0175] The result is that small objects in the image are removed, while larger objects are retained.

[0176] 4. Closing

[0177] Purpose: Dilation followed by corrosion, used to fill small holes inside objects, connect adjacent objects, and reduce gaps between objects.

[0178] process:

[0179] First, dilate the image.

[0180] Then the dilated image is eroded.

[0181] The result is that small holes inside objects in the image are filled, while small gaps between objects are connected.

[0182] 5. Morphological Gradient

[0183] Purpose: To highlight the boundaries of objects.

[0184] process:

[0185] The image is first dilated and then eroded, or first eroded and then dilated.

[0186] Compute the difference between the dilation and erosion results.

[0187] The result is that boundaries in the image are emphasized.

[0188] 6. Top-hat and Black-hat transformations

[0189] Purpose: Used to highlight areas that are lighter or darker than their surroundings, respectively.

[0190] process:

[0191] Top Hat: Subtract the result of the opening operation from the original image.

[0192] Through big data, artificial intelligence and other technical means, we can realize intelligent and automated research and judgment on the technical specifications, drawings and other contents of important materials such as primary equipment, secondary equipment, and intelligent secondary equipment in primary procurement, automatically generate audit reports, and put forward modification opinions for expert reference. We can realize the auxiliary audit capability of technical specifications based on artificial intelligence, with an accuracy rate of ≥95% for equipment identification on drawings, reducing the workload of audit experts by 30%-40%. Users include audit experts of primary equipment, secondary equipment, and intelligent secondary equipment, and realize business support for the technical specification audit link of the bidding process, reducing the workload of experts and improving work efficiency.

[0193] Since the implementation of the project, we have developed an intelligent review module for technical specifications based on general material design specifications, procurement standards and usage experience. We have used knowledge graphs and image recognition technology to automatically extract 187 main information such as material categories, component configurations, and electrical component symbols in technical parameters and design drawings, intelligently identify technical specification problems and generate review reports, and accurately review all categories of solidified technical specifications and self-compiled technical specifications for major categories. Since its application in May 2023, a total of 10,337 national recruitment plans have been intelligently reviewed, including 10,129 solidified national recruitment plans and 208 self-compiled national recruitment plans. We have completed intelligent review of 4,061 technical specifications, automatically identified 1,428 problems, including 1,053 solidified plan problems and 375 self-compiled plan problems. The average review time has been shortened by 55%, effectively improving the digitalization and lean management of procurement.

[0194] A method and system for identifying and reviewing electric power design drawings designed according to the present invention can automatically identify and review electric power design drawings compared to the prior art, reduce manual intervention, and improve work efficiency. It uses deep learning and computer vision technology to improve the accuracy of graphic element recognition and text recognition, reduce human errors and omissions, and improve the quality of design drawings. By constructing a graphic element library and a rule library, it realizes the standardization of identification and review, ensuring that the design drawings meet industry standards.

[0195] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0196] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying and reviewing electric power design drawings, characterized in that: The following steps are involved: S1. Obtain power design drawings and build a graphic element library and rule library; S2. Performing graphic element recognition on the electric power design drawings based on edge recognition methods and deep learning methods to obtain corresponding digital images; S3. In OCR recognition, the weakly supervised learning method is used, combined with deep learning and computer vision processing methods, to recognize the text in the power design drawings and obtain text information; S4, comparing the digital image with the graphic element library to obtain the location and type information of each electrical graphic element in the power design drawing; S5. Compare the position, type information and text information of each electrical graphic element with the rule base to obtain the recognition result, post-process the obtained recognition result, and judge and modify the result through prior knowledge.

2. A method for identifying and reviewing electric power design drawings according to claim 1, characterized in that: The graphic element library constructed in S1 specifically includes a grounding switch, a current transformer, a disconnector, a lightning arrester, a circuit breaker, a three-phase live display, a bushing, a fast grounding switch, a voltage transformer, a fast disconnector, and a cable terminal; The rule base specifically includes various index parameters and constraint relationships of primary equipment, secondary equipment, and intelligent secondary equipment.

3. The method for identifying and reviewing electric power design drawings according to claim 1, characterized in that: The S2 is specifically: S21, preprocessing the acquired power design drawings, specifically graying the power design drawings, smoothing and denoising the grayed images, binarizing the denoised grayed images to convert them into binary images, performing morphological operations on the binary images, including corrosion or expansion, segmenting the processed binary images to obtain multiple region images, and post-processing the multiple region images, including removing small objects, filling holes, and smoothing boundaries; S22, using Faster RCNN network architecture for electrical component identification and training; S23. Through the deep learning network, use the overall attractive wireframe parser to vectorize the refined straight line structure diagram to obtain a preliminary vectorization result, improve the preliminary vectorization result, screen the edges, cluster the points, remove duplicate and redundant information, and obtain a standard vector diagram result.

4. A method for identifying and reviewing electric power design drawings according to claim 3, characterized in that: The S21 is specifically: Grayscale the power design drawings: I gray =0.2989×R+0.5870×G+0.1140×B Among them, R, G, B are the red, green, and blue channel values ​​of the image; Smooth and denoise the grayscale image based on Gaussian filtering: Determine the size and standard deviation of the Gaussian kernel, create the Gaussian kernel, use the Gaussian function to calculate the value of each element in the kernel, calculate the Gaussian value at each position in the kernel based on the standard deviation, Where x, y are the positions relative to the center of the kernel, σ is the standard deviation, and the normalized Gaussian kernel is completed; Slide the Gaussian kernel onto each pixel of the image, calculate the weighted sum of all pixel values ​​under the kernel, and get the new pixel value. For each pixel (i, j), the new pixel value is: Where k is half of the kernel size, G(m,n) is the value of the Gaussian kernel, and I(i+m,j+n) is the pixel value of the original image; The grayscale histogram of the new image is calculated, and all possible thresholds are traversed based on the Otsu algorithm. For each threshold, the image is divided into a foreground part and a background part. For each threshold, the intra-class variance and the inter-class variance of the foreground and the background are calculated. The intra-class variance is used to judge the deviation of the pixel value inside the foreground or the background from its average value, and the inter-class variance is used to judge the separation between the foreground and the background. The threshold that maximizes the inter-class variance is obtained, and all pixels in the image that are less than the threshold are set to black, and all pixels that are greater than or equal to the threshold are set to white; morphological operations are performed on the binary image, including corrosion or expansion, and the processed binary image is segmented to obtain multiple regional images. The multiple regional images are post-processed, including removing small objects, filling holes, and smoothing boundaries.

5. The method for identifying and reviewing electric power design drawings according to claim 3, characterized in that: The S22 is specifically: The VGG16 network is used as the basic network, and the deep learning method is used to extract features of power grid graph elements. The shared convolution and RPN weights are initialized using a large data set pre-trained model. The RPN is trained end-to-end, and Region Proposals are generated. The same pre-trained model is used to initialize the shared convolution, lock the RPN weights, and train the RCNN network with the Proposals obtained by RPN. The loss values ​​of the classification network and regression network are calculated, and back propagation is performed to optimize the network parameters. The trained shared convolution and RCNN are used, the shared convolution layer is fixed, and the RPN is continued to be trained. The RPN is fine-tuned, and the tuned shared convolution and RPN are used to continue training and fine-tuning the RCNN. After training, the final complete Faster-RCNN model is obtained, and electrical components are identified through the final complete Faster-RCNN model.

6. The method for identifying and reviewing electric power design drawings according to claim 1, characterized in that: The S3 is specifically: The pre-trained text detection model is used to detect the image to be recognized. By detecting the area of ​​each text character and the relationship between the text characters, the text area box at the entire word level is predicted. For vertical text, character cutting is used to obtain single character text, and for horizontal text, text lines are used directly. Recognition is performed through the CNN+BiLSTM+CTC model to obtain text information.

7. A method for identifying and reviewing electric power design drawings according to claim 6, characterized in that: Specifically: Using the CNN algorithm, the feature map of the image is extracted through the convolution layer. The RPN network generates candidate regions on the feature map, and the features extracted from the candidate regions on the feature map are pooled to obtain a fixed-size feature map for classification and bounding box regression. The area of ​​each text character in the image is predicted, and based on the relationship between these areas and the text characters, the text area box at the entire word level is predicted. The text detection model uses weak supervision training to learn character features for each word-level annotation area. For vertical text, character cutting technology is used to obtain single character text. The blank space between characters on the vertical projection of the image row is used to segment the single character image. For horizontal text, the text row is used directly to output the final detection result, including the bounding box coordinates of each word or text row, as well as the text content.

8. The method for identifying and reviewing electric power design drawings according to claim 1, characterized in that: The S4 is specifically: Establish a vector template for power components, use correlation measurement to identify the black and white image and the components in it, and perform template matching; use target detection technology to obtain the location and type information of each electrical element in the wiring diagram and save it to an XML file; The contour tracking algorithm is used to extract the outer contour of the connection line of the wiring diagram after the preprocessing operation. According to the acquired position of each electrical graphic element and the extracted outer contour coordinate information of each connection line, the connection relationship between each connection line and the electrical graphic element is analyzed.

9. A power design drawing recognition and review system, characterized in that: It includes drawing acquisition module, graphic element and rule base module, recognition module and comparison module; The drawing acquisition module is used to acquire power design drawings and perform preprocessing; The graphic element and rule library module is used to store graphic elements and review rules; The recognition module is used to extract graphic element information and text information from the power design drawings; The comparison module is used to compare the graphic element information and text information raised by the recognition module with the stored graphic elements and the review rules to obtain the review result.

10. The electric power design drawing recognition and review system according to claim 9, characterized in that: A data maintenance module is also provided, which is used to ensure data consistency throughout the system, avoid data redundancy and conflict, improve data accuracy through verification and validation mechanisms, regularly update data, clean up obsolete or useless data, monitor data usage and performance indicators, and generate data reports.

Images