Electric power design CAD drawing intelligent processing system based on AI large model

Abstract

The invention relates to the technical field of electric power design assistance, and discloses an electric power design CAD drawing intelligent processing system based on an AI large model, and the system comprises an API gateway module which receives an examination request, and generates and distributes a request instruction; the task service scheduling module is used for receiving the request instruction and scheduling CAD analysis service, symbol recognition service, topology analysis service and rule engine service; the CAD analysis service analyzes the drawing file, recognizes graphic data and converts the graphic data into a digital model and a data structure; the symbol recognition service recognizes marking symbols in the graphic elements and the data structure and converts the marking symbols into business objects; the topology analysis service reads a business object, and establishes a spatial relationship and a connection logic; and the rule engine service verifies the digital model, the data structure, the graphic element, the annotation symbol, the business object, the spatial relationship and the connection logic, outputs a verification report, constructs an automatic drawing review system, and improves the efficiency and accuracy of electric power engineering design review.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power design assistance, in particular to an AI large model-based power design CAD drawing intelligent processing system. BACKGROUND

[0002] A power engineering CAD drawing is a core technical document for engineering design, construction and operation and maintenance management, and its standardization and accuracy directly determine the quality of power engineering construction, system operation safety and investment benefit. Power engineering involves multiple professional fields such as power transformation, power transmission and power distribution, and the drawing contains a large amount of key information such as equipment parameters, wiring logic and size annotations. Any omission or error may cause equipment failure, construction conflict or even safety accidents, so strict checking of the CAD drawing is an indispensable key link in the whole life cycle management of power engineering.

[0003] The power engineering drawing has significant professional specificity, and its symbol system is diverse and covers multiple professional types such as electrical, structural and civil engineering. The topology of each system is complex, and a series of industry standards must be strictly followed. The consistency and compliance of the drawing are extremely high. At present, the checking of the CAD drawing in the industry mainly relies on professional checking engineers with rich experience, which is carried out by manually checking page by page and verifying the standard provisions, and a lot of effort is needed to check the details of the drawing and the professional connection problems.

[0004] With the rapid development of China's power industry towards intelligence, large capacity and cross-regional direction, the scale of power engineering design continues to expand, the demand for multi-professional collaborative design increases, and the project delivery cycle is increasingly tight. The traditional manual checking mode has been difficult to adapt to the current design management requirements of high efficiency and high quality. Manual checking is not only tedious and time-consuming, but also easily affected by subjective factors such as the experience and energy of the checking personnel, which may lead to potential errors or judgment deviations. It is also difficult to keep up with the dynamic updates of industry standards in real time, so it is urgent to build an automatic drawing checking system. SUMMARY

[0005] In order to build an automatic drawing checking system, the application provides an AI large model-based power design CAD drawing intelligent processing system, which adopts the following technical solutions:

[0006] An AI large model-based power design CAD drawing intelligent processing system includes the following modules:

[0007] An API gateway module is used to receive a drawing checking request, analyze the checking request to generate and distribute a request instruction;

[0008] A task service scheduling module is used to receive the request instruction and schedule a CAD analysis service, a symbol recognition service, a topology analysis service and a rule engine service in response to the request instruction.

[0009] The CAD analysis service is configured to obtain a drawing file corresponding to a request instruction, analyze the drawing file based on a preset drawing database, identify graphic data in the drawing file, and convert the graphic data into a digital model and a data structure matched in the drawing database;

[0010] The symbol recognition service is configured to identify graphic elements in the digital model and identify annotation symbols in the data structure based on a preset large recognition model, and convert the identified graphic elements and annotation symbols into business objects after association;

[0011] The topology analysis service is configured to read the business objects and establish spatial relationships and connection logics between the business objects based on a preset multi-modal large model according to contents of the business objects;

[0012] The rule engine service is configured to be synchronously enabled with the CAD analysis service, the symbol recognition service, and the topology analysis service, and perform real-time checking on the digital model, the data structure, the graphic elements, the annotation symbols, the business objects, the spatial relationships, and the connection logics based on a preset industry specification rule library, and output a checking report according to a checking result.

[0013] By adopting the above technical solutions, the request of the user enters the system through the API gateway, which is a unified entrance for traffic and is responsible for receiving and distributing the request. The request is transmitted to the task scheduling service of the coordination layer, and the task scheduler starts the four professional services of the core service layer in sequence for processing. The CAD analysis service analyzes the drawing file and converts it into a digital model and a data structure that can be understood by the system. The symbol recognition service identifies the graphic elements and annotation symbols in the drawing and converts them into business objects. The topology analysis service analyzes the spatial relationships and connection logics. The rule engine service performs checking and logical judgment according to the industry specification rule library. The system can automatically analyze, recognize symbols, analyze topology, and check specification compliance. The AI large model is introduced to build an automated and accurate drawing review system, which improves the efficiency and accuracy of power engineering design review.

[0014] Optionally, the API gateway module further includes the following submodules:

[0015] The authentication submodule is connected with the user management module, obtains login information generated after login of the user management module, obtains a review request and authentication information corresponding to the review request according to the login information, and matches the authentication information with permission information in a cloud database. If the matching is passed, the function authorization is performed by calling the authorization submodule, otherwise, the authentication failure information is returned to the user management module;

[0016] The authorization submodule is connected with the authentication submodule, and is configured to match a corresponding authorization instruction according to the login information, analyze the file information obtained from the review request, obtain a plurality of service versions associated with the review request, and select a corresponding service version according to the authorization instruction.

[0017] The balancing submodule is connected with the authorization submodule, and is configured to calculate a required design load according to the file information, and distribute the design load to a calculation unit according to the service version.

[0018] The monitoring submodule is connected with the balancing submodule, and is configured to obtain running data of the distributed calculation unit, and send a normal heartbeat signal if the running data is within a set data range, or send an abnormal heartbeat signal otherwise.

[0019] By using the above technical solution, the API gateway, as a single entry for all client requests, can effectively manage and route external requests, and can also provide authentication, authorization, load balancing, monitoring and other functions.

[0020] Optionally, the balancing submodule further includes the following components:

[0021] The load identification component extracts a file type and a file data volume from the file information, matches a file characteristic value from a set type table according to the file type, and matches a quantity characteristic value from a set quantity table according to the file data volume.

[0022] The load calculation component calculates a load characteristic value according to the file characteristic value and the quantity characteristic value, and calculates a design load according to the load characteristic value and a set load reference value.

[0023] The load adjustment component obtains a current load and a remaining load of the system, calculates a fluctuation characteristic value of the current load in a set reference period, calculates a load ratio of the remaining load and the current load, calculates a distribution coefficient according to the load ratio and the fluctuation characteristic value, calculates a calculation load according to the distribution coefficient and the design load, and distributes the calculation unit according to the calculation load.

[0024] By using the above technical solution, the file characteristics are matched with the quantitative load state, the distribution coefficient is determined in combination with the load fluctuation law and the remaining resource condition, the calculation unit distribution is highly matched with the drawing processing demand, the system can maintain stable operation when facing different types and data volumes of power CAD drawings, and the professional demand of power engineering drawing processing is matched.

[0025] Optionally, the task service scheduling module further includes the following submodules:

[0026] The accounting submodule calculates the utilization rate of the allocated computing unit, calculates the change trend of the utilization rate in a set reference period, and if the utilization rate is greater than a set reference value and the change trend is greater than a set reference trend, calculates a utilization value according to the utilization rate and the reference value, calculates a trend value according to the change trend and the reference trend, and calculates an accounting value according to the utilization value and the trend value.

[0027] The new creation submodule allocates a new computing unit as a temporary operation unit according to the accounting value if the accounting value is less than a set first accounting differentiation value, or allocates a hardware acceleration unit as the temporary operation unit according to the accounting value if the accounting value is not less than the set first accounting differentiation value; wherein the larger the accounting value is, the larger the temporary unit is, and the smaller the accounting value is, the smaller the temporary unit is.

[0028] By adopting the above technical solution, based on the distributed computing power scheduling framework and the hardware acceleration computing capability, it is beneficial to maintain stable and efficient operation in a high-load scenario.

[0029] Optionally, the task service scheduling module further includes the following submodules:

[0030] The accounting submodule calculates the utilization rate of the allocated computing unit, calculates the change trend of the utilization rate in a set reference period, and if the utilization rate is greater than a set reference value and the change trend is greater than a set reference trend, calculates a utilization value according to the utilization rate and the reference value, calculates a trend value according to the change trend and the reference trend, and calculates an accounting value according to the utilization value and the trend value.

[0031] The new creation submodule allocates a hardware acceleration unit as a temporary operation unit according to the accounting value if the accounting value is less than a set second accounting differentiation value, or newly creates an audit request according to the accounting value if the accounting value is not less than the set second accounting differentiation value, groups the uncalculated drawing files into a new file in the smallest file unit as the smallest unit, and returns the file information of the new file to the API gateway module, and the API gateway module allocates a new computing unit as a new creation computing unit according to the file information of the new file; wherein the larger the accounting value is, the larger the temporary unit is, and the smaller the accounting value is, the smaller the temporary unit is.

[0032] By adopting the above technical solution, based on the distributed new creation task scheduling framework and the hardware acceleration computing capability, it is beneficial to maintain stable and efficient operation in a high-load scenario.

[0033] Optionally, the symbol recognition service further includes the following submodules:

[0034] The symbol display submodule is configured to create a new symbol display window, generate a reduced image block corresponding to the graphic element and the annotation symbol when the graphic element and the annotation symbol are recognized, and display the reduced image block in the symbol display window.

[0035] The content comparison submodule is configured to extract a reduced image of a graphic element and a label in the drawing file, display the reduced image in the symbol display window and beside the reduced image block;

[0036] The result aggregation submodule is configured to acquire the input information corresponding to the reduced image through the symbol display window, mark the input information beside the frame of the symbol display window, generate a graphic link corresponding to the reduced image block and the reduced image when the reduced image block and the reduced image in the symbol display window are covered, and load the graphic link on the input information.

[0037] By using the above technical solution, the symbol recognition service can generate a symbol and display the symbol for manual review at the same time.

[0038] Optionally, the topology analysis service further includes the following submodules:

[0039] The content submodule is configured to create an object window, generate a display graphic corresponding to the spatial relationship and the connection logic when the spatial relationship and the connection logic are established, and display the display graphic in the object window;

[0040] The logic submodule is configured to extract a reduced image of the associated relationship of the business object in the drawing file, display the reduced image in the object window and beside the display graphic;

[0041] The relationship submodule is configured to acquire the input information corresponding to the display graphic through the object window, mark the input information in the object window, and generate a multi-modal large model early warning prompt when the display graphic in the object window is covered.

[0042] By using the above technical solution, the topology analysis service can introduce a multi-level verification and manual review mechanism to minimize false positives and false negatives.

[0043] Optionally, the rule engine service further includes the following submodules:

[0044] The rule management submodule is configured to acquire an update instruction in real time, identify a content address from the update instruction, acquire a rule supplement file through the content address, update the rule supplement file to the industry standard rule library, and re-verify a target that is being verified in real time using the industry standard rule library;

[0045] The rule editing submodule is configured to acquire a rule input instruction, identify rule content from the rule input instruction, update the rule content to the industry standard rule library, and re-verify a target that is being verified in real time using the industry standard rule library;

[0046] The log tracking submodule is configured to highlight the modified information if the information updated to the industry standard rule library is the modified information, and save the information updated to the industry standard rule library to a change log.

[0047] By adopting the technical scheme, the adaptability of the rule engine service is improved through rule version management, a visual rule editor and a change log tracking.

[0048] Optionally, the rule engine service further comprises the following sub-modules:

[0049] A result analysis sub-module is configured to obtain a result of the verification, and generate compliance content if the result is in compliance with the specification, or generate error content if the result is not in compliance with the specification;

[0050] A result comparison sub-module is configured to obtain a result of the verification before the update of the rule content and a result of the verification after the update of the rule content, and provide a specification adjustment prompt if the two results are different.

[0051] A report generation sub-module is configured to generate a compliance report according to the compliance content, generate an error report according to the error content, and add the specification adjustment prompt and the result corresponding to the specification adjustment prompt into the compliance report or the error report.

[0052] By adopting the technical scheme, after the judgment of the rule engine service, the system is divided into two directions in the auxiliary layer; if an error is found, the system automatically generates an error report to clearly indicate the non-compliance; if all is in compliance with the specification, the system automatically generates a compliance report.

[0053] Optionally, the rule engine service further comprises the following sub-modules:

[0054] A review sub-module is configured to wait for a review signal in response to the generation of the compliance report or the error report, and import the compliance report or the error report into a review container, create a review display window, and display the review container through the review display window if the review signal is obtained.

[0055] A decision sub-module is connected with the review sub-module, and is configured to load an error editing control on the review display window, record editing content of the editing control, and generate a decision file.

[0056] A feedback sub-module is connected with the decision sub-module, and is configured to receive the decision file and provide a feedback operation interface; the feedback operation interface waits for modified content corresponding to the decision file; and the feedback operation interface is configured to judge the correspondence between the modified content and the decision file, and generate a feedback error warning if the modified content does not correspond to the decision file.

[0057] By adopting the technical scheme, regardless of the type of the report, the report is finally submitted to the artificial review interface for final review, and the corresponding decision is obtained: confirming the error and modifying according to the report, or objecting to the judgment of the system, changing the drawing or submitting feedback according to the review result, and forming a review closed loop. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a system block diagram of an AI large model-based power design CAD drawing intelligent processing system.

[0059] Figure 2 is a module block diagram of an API gateway.

[0060] Figure 3 is a flowchart of a technical route of an AI large model-based power design CAD drawing intelligent processing system.

[0061] Figure 4 is a technology selection and relationship diagram of a user layer in a technology stack.

[0062] Figure 5 is a technology selection and relationship diagram of a business and access layer in a technology stack.

[0063] Figure 6 is a technology selection and relationship diagram of an AI and data layer in a technology stack.

[0064] Figure 7 is a technology selection and relationship diagram of an infrastructure layer in a technology stack. DETAILED DESCRIPTION

[0065] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings.

[0066] In the description of the present specification, the description referring to the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the described embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0067] Traditional manual review is time-consuming and cannot meet the demand of multiple projects in parallel and high-intensity output. Manual review is difficult to cover complex and variable specification items comprehensively, and is prone to missed detection, false detection and poor specification compliance. It is highly dependent on professional technical personnel, and the consistency is difficult to guarantee, and the labor cost is high. In order to improve these problems, the embodiment of the present application discloses an AI large model-based power design CAD drawing intelligent processing system, referring to Figure 1 , including a user management module, an API gateway module and a task service scheduling module.

[0068] The user management module is used for user login and management. The user uploads a CAD drawing file requiring review through the user management module. The format of the CAD drawing file is DWG or PDF. For example, a user, i.e., a requester, submits a drawing review request. An internal drawing is directly transmitted to the review large model through a plug-in of CAD. An external drawing (in DWG / PDF format) is transmitted to the review large model for automatic review, and then transmitted to a reviewer. After checking, the reviewer returns a review report to the OA system. The OA system sends review information to the requester. If the drawing is correct, the requester archives the drawing that has passed the review through the OA system. If modification is required, the requester modifies the drawing according to the review information.

[0069] The API gateway module is used for receiving a review request of a drawing, analyzing the review request, generating and distributing a request instruction. The API gateway module further includes an authentication submodule, an authorization submodule and a balancing submodule.

[0070] The authentication submodule is connected with the user management module, obtains login information generated after login of the user management module, obtains a review request and authentication information corresponding to the review request according to the login information, and matches the authentication information with permission information in a cloud database. If the matching is passed, the authorization submodule is called for function authorization. Otherwise, authentication failure information is returned to the user management module.

[0071] The authorization submodule is connected with the authentication submodule, and is used for matching a corresponding authorization instruction according to the login information, analyzing the review request to obtain file information, obtaining a plurality of service versions associated with the review request, and selecting a corresponding service version according to the authorization instruction. Different service versions correspond to different computing processing capabilities.

[0072] The balancing submodule is connected with the authorization submodule, and is used for calculating a design load required according to the file information, and distributing the design load to a calculation unit according to the service version. The calculation unit is a calculation unit to which a corresponding amount of processing capability or computing resource is allocated. The balancing submodule further includes a load identification component, a load calculation component and a load adjustment component.

[0073] The load identification component extracts a file type and a file data amount from the file information, matches a file characteristic value from a set type table according to the file type, and matches a quantity characteristic value from a set quantity table according to the file data amount.

[0074] The load calculation component calculates a load characteristic value by multiplication according to the file characteristic value and the quantity characteristic value, and calculates the design load by multiplication according to the load characteristic value and a set load reference value.

[0075] The load adjustment component obtains the current load and the remaining load of the system, calculates a fluctuation characteristic value of the current load in a set reference period, calculates a load ratio of the remaining load and the current load, calculates a distribution coefficient according to the load ratio and the fluctuation characteristic value, and calculates a calculation load according to the distribution coefficient and the design load.

[0076] The fluctuation characteristic value is used to quantify the fluctuation degree of the current load of the system in the set reference period, and the calculation formula of the fluctuation characteristic value is: wherein Li is the current load value at the i th sampling time point in the reference period, and the unit is an algorithm unit, 1 algorithm unit corresponding to the basic algorithm requirement of 1 MB-DWG format drawing; is the average value of the current load of all sampling points in the reference period, and n is the sampling number in the reference period.

[0077] The distribution coefficient is a weight coefficient for balancing the system load pressure and the drawing processing requirement, and the value range is 0 < distribution coefficient < 1, reflecting the support ability of the remaining load to the design load, and the calculation formula is: distribution coefficient = load ratio x fluctuation correction coefficient; wherein: load ratio = remaining load / current load; fluctuation correction coefficient = 1-(fluctuation characteristic value / maximum fluctuation threshold), if the fluctuation characteristic value > maximum fluctuation threshold, the fluctuation correction coefficient is 0.1.

[0078] The specific distribution mode of the calculation unit is:

[0079] Calculation load = design load x distribution coefficient. The cloud platform algorithm power distribution is based on resource pooling, and the physical algorithm power is integrated into a shared resource pool, and then the idle calculation unit is preferentially distributed according to the calculation load.

[0080] The monitoring submodule is connected with the balancing submodule, obtains the running data of the distributed calculation unit, and if the running data is within the set data range, a normal heartbeat signal is sent, otherwise an abnormal heartbeat signal is sent.

[0081] As shown in Figure 2 The API gateway is a single entry for all client requests, which can effectively manage and route external requests, and can also provide authentication, authorization, load balancing, logging and monitoring functions. Matching file features and quantifying load state, combined with load fluctuation law and remaining resource situation to determine distribution coefficient, make calculation unit distribution and drawing processing requirement highly consistent, keep the system running smoothly when facing different types and data volume of power CAD drawings, and be beneficial to matching the professional requirements of power engineering drawing processing.

[0082] The task service scheduling module is configured to receive a request instruction, and schedule a CAD analysis service, a symbol recognition service, a topology analysis service, and a rule engine service in response to the request instruction. The task service scheduling module serves as a command center of the process, and is responsible for orderly coordinating the calling sequence and execution logic of the various services in the rear, thereby ensuring that the task can be efficiently and continuously performed.

[0083] The task service scheduling module further includes a calculation submodule and a new creation submodule.

[0084] In this embodiment, the calculation submodule calculates the utilization rate of the allocated computing unit, the utilization rate being the proportion of working time to total time, calculates the change trend of the utilization rate in a set reference period, the change trend being the change percentage; if the utilization rate is greater than a set reference value and the change trend is greater than a set reference trend, then a utilization value is calculated using division according to the utilization rate and the reference value, a trend value is calculated using division according to the change trend and the reference trend, and a calculation value is calculated using multiplication according to the utilization value and the trend value. The reference value can be 50%, and the reference trend can be 10%.

[0085] The new creation submodule determines whether the calculation value is less than a set first calculation distinguishing value, and if so, allocates a new computing unit as a temporary operation unit according to the calculation value; otherwise, allocates a hardware acceleration unit as a temporary operation unit according to the calculation value; wherein the greater the calculation value, the greater the temporary unit, and the smaller the calculation value, the smaller the temporary unit.

[0086] In addition, the new creation submodule determines whether the calculation value is less than a set second calculation distinguishing value, and if so, allocates a hardware acceleration unit as a temporary operation unit according to the calculation value; otherwise, creates a review request according to the calculation value, groups the uncalculated drawing files into a new file in the smallest file unit as the smallest unit, and returns the file information of the new file to the API gateway module, which allocates a new computing unit as a new creation computing unit according to the file information of the new file; wherein the greater the calculation value, the greater the temporary unit, and the smaller the calculation value, the smaller the temporary unit.

[0087] The first calculation distinguishing value is a calculation value threshold for distinguishing between “regular computing power expansion” and “hardware acceleration computing power expansion”, and is a system preset fixed value with a value of 1.6 and a calculation value range of 1-5. The calculation value reflects the overall tightness of the utilization rate of the computing unit + the change trend:

[0088] Calculation value < 1.6: the utilization rate of the computing unit does not reach a high level of tension, and a new regular general computing unit (such as a CPU computing node) is added as a temporary operation unit;

[0089] Calculation value ≥ 1.6: the computing unit is close to saturation, and a GPU / TPU hardware acceleration unit is called as a temporary operation unit.

[0090] The second accounting distinction value is the threshold value for distinguishing between "hardware-accelerated computing power expansion" and "task splitting and redistribution". It is a fixed value preset by the system, with a value of 2.8, and the overall range of the accounting value is 1.6-5.

[0091] If the calculated value is less than 2.8, the hardware acceleration unit can meet the computing power requirements and can be directly allocated.

[0092] If the calculated value is ≥2.8, the hardware acceleration unit cannot meet the requirements. The unprocessed drawings will be split into "single drawing / single layer" (the smallest file unit), and a new review request will be generated and returned to the API gateway, which will then reallocate other independent computing units.

[0093] The CAD parsing service retrieves the drawing file corresponding to the requested command. Based on a pre-defined drawing database, it parses the drawing file, identifies the graphic data within, and converts it into a matching digital model and data structure within the database. Specifically, the CAD parsing service decodes DWG format binary drawing files using a professional CAD parsing engine (such as LibreDWG or ODA-SDK), extracting geometric elements (such as line segments, arcs, block references, etc.), layer structures, and attribute information (such as color, line type, and text annotations), and converts this information into structured, machine-readable JSON data. This process achieves a mapping from unstructured graphic data to a semantic digital model.

[0094] The resulting digital model is an object-oriented drawing information model described in JSON format. Each graphic element corresponds to a JSON object, which contains the following key fields:

[0095] {"entity_id":"entity_001",

[0096] "deftype": "ARC"

[0097] “layer”: “Electrical layer”

[0098] “geometry”:{

[0099] "center": {"x": 100, "y": 200},

[0100] "radius": 50,

[0101] "start_angle": 0,

[0102] “end_angle”:90},

[0103] "properties": {

[0104] "color": "red",

[0105] "linetype": "solid",

[0106] "handle": "ABC123",

[0107] "owner": "BLOCK_01"}.

[0108] Here, "deftype" is used to distinguish the type of graphical elements (such as ARC, LINE, TEXT, etc.), facilitating subsequent symbol recognition modules for differential processing; "handle" and "owner" are unique identifiers, providing a basis for subsequent topology analysis and graph element association; all graph elements are organized by layer and type, forming a hierarchical data structure, facilitating query and traversal, and facilitating subsequent recognition and differential processing.

[0109] The symbol recognition service is used to identify graphical elements in the digital model and identify annotation symbols in the data structure based on the pre-set recognition large model, and then associate the recognized graphical elements and annotation symbols and convert them into business objects. The symbol recognition service receives the JSON data output by the CAD analysis service and completes the conversion through the following three steps:

[0110] Graph clustering: based on spatial position and graphical features, adjacent geometric elements (such as lines, arcs, circles) are aggregated into potential symbols;

[0111] Text association: identify nearby text annotations (such as "10A", "Circuit Breaker") and bind them to the corresponding symbols;

[0112] Semantic mapping: according to the clustering results and annotation content, map to the pre-defined business object types (such as "Circuit Breaker", "Transformer").

[0113] The output is an enhanced JSON structure, example:

[0114] { "symbols":

[0115] [ { "symbol_id": "symbol_081",

[0116] "type": "Circuit Breaker",

[0117] "confidence": 0.95,

[0118] "entities": ["entity_001", "entity_005"],

[0119] "bbox": {"xmin": 100, "ymin": 200, "xmax": 150, "ymax": 250},

[0120] "attributes": {"rating": "10A", "manufacturer": "Siemens"}]},

[0121] "remaining_entities": ["entity_002", "entity_003"]}.

[0122] GraphSAGE (Graph-Sample-and-Aggregate) model, which belongs to a kind of graph neural network, is suitable for inductive learning and can process large-scale graph metadata (such as thousands of nodes); through neighbor sampling to control memory consumption, it is suitable for high-concurrency scenarios; it can directly reason about unseen graphs, adapting to production environment requirements.

[0123] In order to enable the symbol recognition service to generate symbols and display symbols for manual review, the symbol recognition service further includes a symbol display submodule, a content comparison submodule, and a result summary submodule.

[0124] The symbol display submodule creates a symbol display window, and when a graphic element and a label symbol are recognized, a reduced image corresponding to the graphic element and the label symbol is generated, and the reduced image is displayed in the symbol display window.

[0125] The content comparison submodule is used to extract a reduced image of a graphic element and a label symbol in a drawing file, display the reduced image in the symbol display window, and display the reduced image next to the reduced image.

[0126] The result summary submodule obtains the entered information corresponding to the reduced image through the symbol display window, marks the entered information next to the border of the symbol display window, and when the reduced image in the symbol display window is covered by a new reduced image, a graphic link corresponding to the covered reduced image is generated, and the graphic link is loaded on the entered information.

[0127] The topology analysis service is used to read business objects and establish spatial relationships and connection logic between business objects based on a pre-set multi-modal large model according to the content of the business objects. In order to enable the topology analysis service to introduce a multi-level verification and manual review mechanism and minimize false positives and false negatives, the topology analysis service further includes a content submodule, a logic submodule, and a relationship submodule.

[0128] The "preset multi-modal large model" in this application is a power CAD dedicated model based on LLM large model of multiple input data, which integrates visual + text dual modalities. For example:

[0129] Training data: 100,000+ annotated power CAD drawings (power transmission / distribution professionals), industry standard texts such as "Electric Power Engineering Drawing Standards", and power equipment spatial relationship annotation datasets;

[0130] Model structure: bottom layer is visual encoder (extracts drawing graph coordinates / size / shape) + text encoder (extracts device name / parameter of annotation symbol), middle layer is cross-modal fusion layer, and top layer is topological relationship output layer. Among them, cross-modal fusion is not an independent module, but is implicitly realized through the aggregation mechanism of GraphSAGE. The specific process is as follows:

[0131] Node construction: geometric primitives and adjacent text annotations are taken as graph nodes, and preliminary connections are constructed based on spatial relationships;

[0132] Feature aggregation: GraphSAGE aggregates the features of each node and its neighbors (including graph and text nodes) through multiple iterations;

[0133] Semantic fusion: the model automatically learns the correlation pattern between graph combination and text description in the feature space, realizing deep fusion of geometric and text information;

[0134] Output structured objects: finally output business objects (Symbol) with semantic labels, complete the mapping from multi-modal data to business semantics.

[0135] Optimization method: through pre-training + power scene fine-tuning, adapt to the recognition of spatial features and connection logic of power equipment.

[0136] The steps of establishing spatial relationship and connection logic using multi-modal large model are as follows:

[0137] Step 1: the model reads the business objects output by the symbol recognition service, extracts visual features (X / Y / Z coordinates of device graph, size) and text features (device name / parameter / number).

[0138] Step 2: encode the features to generate business object feature vectors, including spatial coordinates, device type, and parameter sub-vectors. Feature encoding refers to the process of converting graph and text information into machine-processable vector representation during symbol recognition and topology analysis. The specific steps are as follows:

[0139] Feature extraction: extract geometric features (coordinates, shape), text features (device name, parameter), and attribute features (layer, color) from JSON data;

[0140] Vectorization encoding: map various features into high-dimensional vectors using pre-trained encoders (e.g., PyTorch-based visual and text encoders);

[0141] Fusion into business object feature vector: concatenate or weight the sub-vectors of spatial coordinates, device types, and parameters to form a unified feature vector for subsequent topology correlation calculation.

[0142] This process is automatically completed by the multi-modal large model in the system without human intervention.

[0143] Step 3: Calculate the spatial correlation between business objects: spatial correlation = spatial distance weight × device type matching degree, spatial distance weight = 1 / (1+d), d is the normalized value of pixel distance; device type matching degree is set according to specifications, such as transformer-circuit breaker matching degree 0.9.

[0144] Step 4: Filter business object pairs with correlation ≥0.7, and combine power connection rules, such as "circuit breaker must be connected to transformer outlet", to establish connection logic, such as "transformer A-circuit breaker B-bus C".

[0145] Device type matching degree is a pre-defined numerical table based on power engineering design specifications, reflecting the connection logic strength between different device types. Its setting basis includes:

[0146] National standards: such as "GB / T 4728 Electrical diagrams - Graphical symbols";

[0147] Industry specifications: such as "DL / T 5352 Technical regulations for power system design";

[0148] Engineering experience library: typical connection system in historical drawings.

[0149] Example matching degree table (part) as follows:

[0150] Device A Device B Matching degree Explanation Transformer Circuit breaker 0.9 Typical outgoing line connection relationship Circuit breaker Busbar 0.8 Common combination of power distribution connection Switch Load 0.7 Strong functional association Transformer Capacitor cabinet 0.4 Weak association, depending on specific design

[0151] System implementation:

[0152] Matching degree table is stored in industry specification rule library in the form of rule file;

[0153] Support dynamic update through rule management submodule;

[0154] In the topology analysis process, the GAT model combines the table with the spatial distance weight to calculate the final correlation and outputs graph structure data containing connection relationships.

[0155] Step 5: Spatial relationships (such as "circuit breaker B is located 50 cm to the right of transformer A") and connection logic can be output in JSON format for rule engine verification.

[0156] The content submodule is used to create an object window, generate display graphics corresponding to the spatial relationship and connection logic when establishing the spatial relationship and connection logic, and display the display graphics in the object window.

[0157] The logic submodule is used to extract reduced images of business objects having a correlation relationship in a drawing file, display the reduced images in the object window, and locate the reduced images next to the display graphics.

[0158] The relationship submodule obtains the entered information corresponding to the display graphics through the object window, marks the entered information in the object window, and generates a multi-modal large model early warning prompt when the display graphics in the object window are covered.

[0159] The rule engine service is used to be enabled synchronously with the CAD analysis service, the symbol recognition service, and the topology analysis service, and to perform real-time checking on the digital model, the data structure, the graphic element, the annotation symbol, the business object, the spatial relationship, and the connection logic based on a preset industry specification rule library, and to output a checking report according to the checking result. The rule engine service further includes a rule management submodule, a rule editing submodule, a log tracking submodule, a result analysis submodule, a result comparison submodule, a report generation submodule, a review submodule, a decision submodule, and a feedback submodule.

[0160] The rule management submodule is used to obtain an update instruction in real time, identify a content address from the update instruction, obtain a rule supplement file through the content address, update the rule supplement file to the industry specification rule library, and recheck a target that is being checked in real time using the industry specification rule library.

[0161] The rule editing submodule is used to obtain a rule entry instruction, identify rule content from the rule entry instruction, update the rule content to the industry specification rule library, and recheck a target that is being checked in real time using the industry specification rule library.

[0162] The log tracking submodule highlights modified information if the information updated to the industry specification rule library is the modified information, and saves the information updated to the industry specification rule library to a change log.

[0163] The result analysis submodule is used to obtain a result of the checking, and generate compliance content if the result is compliant, or generate error content otherwise.

[0164] The result comparison submodule is used to obtain a result of a pre-update rule content checking and a result of a post-update rule content checking after a target is rechecked, and perform a specification adjustment prompt if the results are different.

[0165] The report generation submodule generates a compliance report according to the compliance content, generates an error report according to the error content, and adds the specification adjustment prompt and the result corresponding to the specification adjustment prompt into the compliance report or the error report.

[0166] The review submodule is configured to wait for a review signal in response to the generation of the compliance report or the error report, the review signal being manually input, and if the review signal is obtained, the compliance report or the error report is imported into a review container, a review display window is newly created, and the review container is displayed through the review display window.

[0167] The decision submodule is connected with the review submodule and is configured to load an error editing control on the review display window and record editing content of the editing control to generate a decision file.

[0168] The feedback submodule is connected with the decision submodule and is configured to receive the decision file and provide a feedback operation interface, the feedback operation interface waiting for uploading of modification content corresponding to the decision file, judging the correspondence between the modification content and the decision file, and generating a feedback error warning if the modification content does not correspond to the content of the decision file.

[0169] The adaptability of the rule engine service is improved through rule version management, a visual rule editor and a change log tracking. After the judgment of the rule engine service, the system is divided into two directions in the auxiliary layer; if an error is found, the system will automatically generate an error report and clearly point out the non-compliance; if all the specifications are met, the system will automatically generate a compliance report. Regardless of the report generated, it will eventually be submitted to the manual review interface for final review to obtain the corresponding decision: confirm the error and modify according to the report, or object to the judgment of the system, change the drawing or submit feedback according to the review result, and form a review closed loop.

[0170] The request of the user enters the system through the API gateway, which is the unified entrance of traffic and is responsible for the reception and distribution of requests; the request is transmitted to the task scheduling service of the coordination layer, and the task scheduler starts the four professional services of the core service layer in turn for processing; the CAD analysis service analyzes the drawing file and converts it into a digital model and data structure that can be understood by the system; the symbol recognition service recognizes the graphic elements and annotation symbols in the drawing and converts them into business objects; the topology analysis service analyzes the spatial relationship and connection logic; the rule engine service checks and logically judges according to the industry specification rule library; it can automatically analyze, recognize symbols, analyze topology and check specification compliance; an AI large model is introduced to build an automatic and accurate drawing review system, which improves the efficiency and accuracy of power engineering design review.

[0171] This embodiment constructs a CAD drawing intelligent review system based on artificial intelligence and rule engine, realizes automatic analysis, symbol recognition, topology analysis and specification compliance check of CAD drawings. By integrating high-performance computing hardware and full-stack optimized software architecture, it can support multi-concurrent drawing processing capability, compatible with multiple national and industry design specifications, realize the whole process automation from drawing uploading, intelligent analysis to error report generation. Achieve the core goal of improving review efficiency, reducing labor cost, enhancing design specification compliance and engineering quality reliability.

[0172] Accuracy is the priority, ensuring the accuracy of the review task is the primary goal. Combined with deep learning-based symbol recognition and topology analysis algorithms, and introducing multi-level verification and manual review mechanism, to minimize false positives and false negatives, to ensure the credibility and practicality of the review results. High performance and high concurrency are realized, supporting multi-task parallel processing to meet enterprise-level throughput requirements. The system relies on distributed task scheduling framework and GPU accelerated computing capability to ensure stable and efficient operation performance under high load scenarios. It is maintainable, and rules and business logic should be easy to update and maintain to respond to specification changes. Through rule version management, visual rule editor and change log tracking, maintenance complexity is reduced, ensuring that the system can quickly respond to specification iterations, reducing manual intervention and downtime.

[0173] As shown in Figure 3 , the technical stack of the specific implementation content of this embodiment mainly includes four levels: user layer, business and interface layer, data and AI layer, and infrastructure layer. As shown in Figure 4 , the user layer includes a front-end interaction platform and user roles. The front-end interaction platform includes a Web front-end and a mobile APP, and the user roles include designers, review experts and administrators. As shown in Figure 5 , the business and interface layer includes API gateway & proxy, API service and task scheduling. API gateway & proxy includes NGINX and OAuth2 authentication, API service includes FastAPI and RESTful interface, and task scheduling includes Celery+Redis and task monitoring. As shown in Figure 6 , the data and AI layer includes CAD parsing engine, device recognition engine, AI framework & reasoning, rule engine and post-processing fusion. The CAD parsing engine includes LibreDWG and ODA-SDK, the device recognition engine includes component model extraction or AI detection-legend comparison, the AI framework & reasoning includes PyTorch+TorchVisions and multi-modal LLM enhancement, the rule engine includes Drools rule engine and topology graph construction, and the post-processing fusion includes virtual block generator and manual review workstation. As shown in Figure 7As shown, the infrastructure layer includes an operating system and container platforms and data stores. The operating system is Ubuntu, the container platforms include Kubernetes and Docker containers, and the data stores include MySQL, Redis cache, a vector database, and file storage.

[0174] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that variations, modifications, substitutions and changes can be made to the above-described embodiments by those skilled in the art in the scope of the present application.

Claims

1. A smart processing system for power design CAD drawings based on AI large-scale models, characterized in that, Includes the following modules: The API gateway module is used to receive review requests for drawings, parse the review requests, generate and distribute request instructions; The task service scheduling module is used to receive request instructions and, in response to the request instructions, schedule CAD parsing service, symbol recognition service, topology analysis service, and rule engine service. The CAD parsing service is used to obtain drawing files corresponding to the request instructions. It parses the drawing files based on a preset drawing database, identifies the graphic data in the drawing files, and converts the graphic data into matching digital models and data structures in the drawing database. The symbol recognition service is used to identify graphic elements in digital models and labeled symbols in data structures based on a preset large recognition model, and then the identified graphic elements and labeled symbols are associated and converted into business objects. The topology analysis service is used to read business objects and establish spatial relationships and connection logic between business objects based on the content of the business objects according to the preset multimodal large model; The rules engine service is used to synchronize with the CAD parsing service, symbol recognition service, and topology analysis service. Based on a preset industry standard rule base, it performs real-time verification of digital models, data structures, graphic elements, annotation symbols, business objects, spatial relationships, and connection logic, and outputs a verification report based on the verification results.

2. The intelligent processing system for power design CAD drawings based on AI large model as described in claim 1, characterized in that, The API gateway module also includes the following sub-modules: The authentication submodule connects to the user management module, obtains the login information generated after logging in to the user management module, retrieves the review request and the corresponding authentication information based on the login information, and matches the authentication information with the permission information in the cloud database; if the match is successful, it calls the authorization submodule to authorize the function; otherwise, it returns an authentication failure message to the user management module. The authorization submodule, connected to the authentication submodule, is used to match the corresponding authorization instructions based on the login information, parse the review request to obtain file information, obtain multiple service versions related to the review request, and select the corresponding service version according to the authorization instructions. The load balancing submodule, connected to the authorization submodule, is used to calculate the required design load based on the file information and distribute the design load to calculation units according to the service version. The monitoring submodule connects to the balancing submodule to obtain the operating data of the allocated computing units. If the operating data is within the set data range, it sends a normal heartbeat signal; otherwise, it sends an abnormal heartbeat signal.

3. The intelligent processing system for power design CAD drawings based on AI large model as described in claim 2, characterized in that, The equalization submodule also includes the following components: The load identification component extracts the file type and file data volume from the file information, matches the file feature value from the set type table according to the file type, and matches the quantity feature value from the set quantity table according to the file data volume. The load calculation component calculates the load characteristic value based on the file characteristic value and the quantity characteristic value, and calculates the design load based on the load characteristic value and the set load reference value; The load adjustment component obtains the current load and remaining load of the system, calculates the fluctuation characteristic value of the current load within a set reference period, calculates the load ratio of the remaining load to the current load, calculates the allocation coefficient based on the load ratio and fluctuation characteristic value, calculates the calculation load based on the allocation coefficient and the design load, and allocates calculation units according to the calculation load.

4. The intelligent processing system for power design CAD drawings based on AI large model according to claim 3, characterized in that, The task service scheduling module also includes the following sub-modules: The accounting submodule calculates the utilization rate of the allocated computing units and calculates the trend of the utilization rate over a set reference period. If the utilization rate is greater than the set benchmark value and the trend of the change is greater than the set benchmark trend, the utilization value is calculated based on the utilization rate and the benchmark value, the trend value is calculated based on the trend of the change and the benchmark trend, and the accounting value is calculated based on the utilization value and the trend value. When creating a new submodule, if the calculated value is less than the set first calculation distinction value, a new calculation unit is allocated as a temporary calculation unit based on the calculated value. Otherwise, hardware acceleration units are allocated as temporary computing units based on the calculated value; the larger the calculated value, the larger the temporary unit, and the smaller the calculated value, the smaller the temporary unit.

5. The intelligent processing system for power design CAD drawings based on AI large model according to claim 3, characterized in that, The task service scheduling module also includes the following sub-modules: The accounting submodule calculates the utilization rate of the allocated computing units and calculates the trend of the utilization rate over a set reference period. If the utilization rate is greater than the set benchmark value and the trend of the change is greater than the set benchmark trend, the utilization value is calculated based on the utilization rate and the benchmark value, the trend value is calculated based on the trend of the change and the benchmark trend, and the accounting value is calculated based on the utilization value and the trend value. If a newly created submodule has a calculated value less than the set second calculation distinction value, a hardware acceleration unit is allocated as a temporary computing unit based on the calculated value. Otherwise, a new review request is created based on the calculated value. The smallest file unit is used as the smallest unit. The uncalculated drawing files are collected into a new file. The file information of the new file is returned to the API gateway module. The API gateway module allocates a new calculation unit as the new calculation unit based on the file information of the new file. The larger the calculated value, the larger the temporary unit, and the smaller the calculated value, the smaller the temporary unit.

6. The intelligent processing system for power design CAD drawings based on AI large model according to claim 1, characterized in that, The symbol recognition service also includes the following sub-modules: The symbol display submodule is used to create a new symbol display window. When a graphic element and a label symbol are detected, a reduced-size block corresponding to the graphic element and the label load is generated and displayed in the symbol display window. The content comparison submodule is used to extract the reduced images of graphic elements and annotation loads in the drawing file, and display the reduced images in the symbol display window next to the reduced blocks; The results summary submodule is used to obtain the input information corresponding to the scaled-down image through the symbol display window, mark the input information next to the border of the symbol display window, and generate graphic links corresponding to the scaled-down image when the scaled-down tile and scaled-down image in the symbol display window are covered, and load the graphic links onto the input information.

7. The intelligent processing system for power design CAD drawings based on AI large model according to claim 1, characterized in that, The topology analysis service also includes the following sub-modules: The content submodule is used to create new object windows. When establishing spatial relationships and connection logic, it generates display graphics corresponding to the spatial relationships and connection logic and displays the display graphics in the object window. The logic submodule is used to extract the scaled-down images of business objects that have relationships in the drawing file, and display the scaled-down images in the object window next to the displayed graphic. The Relationships submodule is used to obtain the input information corresponding to the displayed graphic through the object window, mark the input information in the object window, and generate a multimodal large model warning prompt when the displayed graphic in the object window is covered.

8. The intelligent processing system for power design CAD drawings based on AI large model according to claim 1, characterized in that, The rules engine service also includes the following sub-modules: The rule management submodule is used to obtain update instructions in real time, identify the content address from the update instructions, obtain the rule supplementary file through the content address, update the rule supplementary file to the industry standard rule library, and re-verify the targets that are currently being verified in real time using the industry standard rule library. The rule editing submodule is used to obtain rule input instructions, identify rule content from the rule input instructions, update the rule content to the industry standard rule library, and re-validate targets that are currently being validated in real time using the industry standard rule library; In the log tracking submodule, if the information updated to the industry standard rule base is modified information, the modified information will be highlighted. The information updated to the industry standard rule base will be saved to the change log.

9. The intelligent processing system for power design CAD drawings based on AI large model according to claim 8, characterized in that, The rules engine service also includes the following sub-modules: The results analysis submodule is used to obtain the verification results. If the results are compliant with the specifications, compliant content is generated; otherwise, error content is generated. The result comparison submodule is used to obtain the result of the verification before the rule content is updated and the result of the verification after the rule content is updated after the target is re-verified. If the results are different, the module will prompt for standard adjustment. The report generation submodule generates compliance reports based on compliant content and error reports based on erroneous content, and adds regulatory adjustment prompts and corresponding results to the compliance report or error report.

10. The intelligent processing system for power design CAD drawings based on AI large model according to claim 9, characterized in that, The rules engine service also includes the following sub-modules: The review submodule is used to wait for a review signal in response to the generation of a compliance report or error report. If a review signal is obtained, the compliance report or error report is imported into the review container, a new review display window is created, and the review container is displayed through the review display window. The decision submodule, connected to the review submodule, is used to load error editing controls on the review display window, record the edited content of the editing controls, and generate a decision file; The feedback submodule, connected to the decision submodule, is used to receive decision documents and provide a feedback operation interface; The feedback operation interface waits for the uploaded modification content corresponding to the decision document; it then checks the correspondence between the modification content and the decision document. If the modification content does not correspond to the content of the decision document, a feedback error warning is generated.

Images