Control cabinet wiring planning method and system based on multi-modal analysis and knowledge constraint

Abstract

This invention provides a control cabinet wiring planning method and system based on multimodal analysis and knowledge constraints, belonging to the technical field of digital process planning. The method includes acquiring multi-source drawing data of the control cabinet, extracting global topology features and text annotation features, determining the local pixel coordinates of each terminal of the control cabinet based on the global topology features, and determining an initial wiring task set; determining process rule sub-graphs based on the initial wiring task set, constraining the process rule sub-graphs to obtain numerical constraint indicators, and constructing a dynamic process constraint feature matrix; determining structured process reasoning prompts based on the initial wiring task set and the dynamic process constraint feature matrix, and determining a digital process instruction list based on the physical wiring sequence chain and wiring table to complete the control cabinet wiring planning. This application achieves the technical effect of alleviating the low level of process standardization in electrical control cabinet assembly operations.

Description

Technical Field

[0001] This invention relates to the field of digital process planning technology, and in particular to a control cabinet wiring planning method and system based on multimodal analysis and knowledge constraints. Background Technology

[0002] In the field of industrial automation control, the assembly and wiring of electrical control cabinets are crucial to the overall product quality and electrical safety. Currently, wiring process planning primarily relies on skilled electricians using their experience to manually interpret diagrams and make on-site decisions. In actual operation, workers must repeatedly review multi-source, heterogeneous design drawings, such as two-dimensional electrical schematics and assembly layout diagrams, and manually translate the abstract electrical logic connections into the wiring routes within the cabinet's three-dimensional physical space.

[0003] However, in engineering practice, there is a lack of efficient digital mapping methods for connecting multi-source design drawings with physical space. When operators read component wiring points with deflection angles and multi-scale markings, the cognitive load is extremely high, which can easily lead to terminal mapping errors or omissions in cross-page connection logic. The wiring process lacks standardized and rigid rules for supervision. Manual wiring makes it difficult to accurately and in real time verify the physical isolation distance between high-voltage power lines and low-voltage signal lines. Furthermore, it is impossible to accurately predict the actual stacking and filling degree inside the cable tray during the design stage, which leads to frequent physical conflicts such as electromagnetic interference, cable tray congestion, and excessively small cable bending radii during the assembly process.

[0004] However, the existing technology has the following problems: reliance on manual drawing interpretation leads to inconsistent wiring standards and a high error rate, and robots in automated wiring systems have difficulty understanding the logic of design drawings, resulting in a low level of process standardization in electrical control cabinet assembly operations. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a control cabinet wiring planning method based on multimodal analysis and knowledge constraints, so as to alleviate the technical problems of inconsistent wiring standards and high error rate caused by relying on manual drawing interpretation, and the difficulty for robots in automated wiring systems to autonomously understand the logic of design drawings, which leads to a low level of process standardization in electrical control cabinet assembly operations.

[0006] In a first aspect, embodiments of this application provide a control cabinet wiring planning method based on multimodal analysis and knowledge constraints, the method comprising: Acquire multi-source drawing data of the control cabinet, extract the global topology features and text annotation features of the multi-source drawing data, determine the local pixel coordinates of each wiring terminal of the control cabinet based on the global topology features, and determine the initial wiring task set. The multi-source drawing data includes electrical schematic diagram, general assembly layout diagram and wiring table. Based on the initial wiring task set, a process rule subgraph is determined, and the process rule subgraph is constrained to obtain numerical constraint indices, and a dynamic process constraint feature matrix is ​​constructed. Based on the initial wiring task set and the dynamic process constraint feature matrix, structured process reasoning prompts are determined. Based on the structured process reasoning prompts and the dynamic process constraint feature matrix, the physical wiring sequence timing chain is determined. Based on the physical wiring sequence timing chain and the wiring table, a digital process instruction list is determined to complete the control cabinet wiring planning.

[0007] In one implementation of the first aspect, the steps of acquiring multi-source drawing data of the control cabinet, extracting global topology features and text annotation features from the multi-source drawing data, determining the local pixel coordinates of each wiring terminal of the control cabinet based on the global topology features, and determining the initial wiring task set include: Global topological features and text annotation features of the multi-source drawing data are extracted based on a fully convolutional network. The local pixel coordinates of each terminal are determined based on the global topology features. A logic mapping matrix is ​​constructed based on the electrical schematic diagram and the assembly layout diagram, and an initial wiring task set is determined according to the logic mapping matrix.

[0008] In one implementation of the first aspect, the process of determining the local pixel coordinates of each terminal based on the global topology features includes: Based on the global topology features, the response heatmap of the wiring terminals is determined using a local key point localization algorithm. The local pixel coordinates of each terminal are determined based on the response heatmap of the terminal.

[0009] In one implementation of the first aspect, the process of constructing a logic mapping matrix based on the electrical schematic diagram and the assembly layout diagram, and determining the initial wiring task set according to the logic mapping matrix, includes: Extract the logical network topology diagram of the electrical schematic diagram, the logical network topology diagram including a set of logical terminals determined based on the text annotation features and a set of logical connected edges determined based on the set of logical terminals; Based on the local pixel coordinates of the terminal, the physical entity terminals identified in the assembly layout diagram are constructed as a set of physical space points; A comprehensive alignment confidence score is determined based on the set of logical terminals and the set of physical space points, and a logical mapping matrix is ​​determined based on the comprehensive alignment confidence score. Based on the logical mapping matrix, the set of logically connected edges is translated into wiring tasks, and an initial wiring task set is determined according to the wiring tasks.

[0010] In one implementation of the first aspect, the steps of determining a process rule subgraph based on the initial wiring task set, constraining the process rule subgraph to obtain numerical constraint indices, and constructing a dynamic process constraint feature matrix include: Determine the process rule sub-graph corresponding to the wiring tasks in the initial wiring task set; The process rule subgraph is converted into numerical constraint indices based on the compilation mapping function. Acquire three-dimensional point cloud data inside the control cabinet, and construct a dynamic process constraint feature matrix based on the three-dimensional point cloud data.

[0011] In one implementation of the first aspect, the process of completing the control cabinet wiring plan includes: The structured process reasoning prompts are determined based on the initial wiring task set and the dynamic process constraint feature matrix. Based on the structured process reasoning prompts and the dynamic process constraint feature matrix, the initial wiring task set is sorted to obtain the physical wiring sequence time chain. Based on the physical wiring sequence timing chain and the wiring table, the wiring task is calculated, and a digital process instruction list is output according to the numerical constraint index to complete the control cabinet wiring plan.

[0012] In one implementation of the first aspect, the method further includes adjusting the digital process instruction list to obtain adjustment actions; The fine-grained features of the adjustment action are recorded using an action capture operator, and the fine-grained features are arranged and wrapped to obtain a dataset of differences before and after the process correction. The process scheme containing the difference dataset before and after the process modification is determined based on the bottom-line verification algorithm, and the final executable process scheme is obtained based on the determination result.

[0013] One implementation of the first aspect further includes determining a short text of process reflection summary based on the digital process instruction list and the final executable process scheme; Using a pre-defined industrial text embedding model, the short text of the process reflection summary is projected onto a pre-defined high-dimensional semantic space to obtain an incremental empirical feature vector; The incremental experience feature vectors are input into the long-term experience vector database. Incremental experience feature vectors with similarity scores greater than a preset threshold are retrieved from the long-term experience vector database to complete the update of the long-term experience vector database.

[0014] Secondly, embodiments of this application provide a control cabinet wiring planning system based on multimodal analysis and knowledge constraints, the system comprising: The multimodal parsing module is used to acquire multi-source drawing data of the control cabinet, extract the global topology features and text annotation features of the multi-source drawing data, determine the local pixel coordinates of each wiring terminal of the control cabinet based on the global topology features, and determine the initial wiring task set. The multi-source drawing data includes electrical schematic diagrams, assembly layout diagrams and wiring tables. The process knowledge management center, which is connected to the multimodal parsing module, is used to determine the process rule subgraph based on the initial wiring task set, constrain the process rule subgraph to obtain numerical constraint indicators, and construct a dynamic process constraint feature matrix. The path planning brain, which is connected to the multimodal parsing module and the process knowledge management center respectively, is used to determine the structured process reasoning prompt words based on the initial wiring task set and the dynamic process constraint feature matrix, determine the physical wiring sequence chain based on the structured process reasoning prompt words and the dynamic process constraint feature matrix, and determine the digital process instruction list based on the physical wiring sequence chain and the wiring table to complete the control cabinet wiring planning. A virtual rendering module, connected to the path planning brain, is used to adjust the digital process instruction list to obtain adjustment actions. It uses an action capture operator to record the fine-grained features of the adjustment actions, arranges and wraps the fine-grained features to obtain a dataset of differences before and after process correction, and judges the process scheme containing the dataset of differences before and after process correction based on the bottom-line verification algorithm. The final executable process scheme is obtained based on the judgment result. The process reflection module, connected to the virtual rendering module, is used to determine a short text of process reflection summary based on the digital process instruction list and the final executable process scheme. Using a preset industrial text embedding model, the short text of process reflection summary is projected onto a preset high-dimensional semantic space to obtain incremental experience feature vectors. The incremental experience feature vectors are input into a long-term experience vector database. Incremental experience feature vectors with similarity scores greater than a preset threshold are retrieved from the long-term experience vector database to complete the update of the long-term experience vector database.

[0015] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the control cabinet wiring planning method based on multimodal analysis and knowledge constraints provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a control cabinet wiring planning system based on multimodal analysis and knowledge constraints provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, it is a flowchart illustrating the control cabinet wiring planning method based on multimodal analysis and knowledge constraints provided in an embodiment of the present invention.

[0021] To facilitate understanding of this embodiment, a detailed description of the control cabinet wiring planning method based on multimodal analysis and knowledge constraints disclosed in this embodiment of the invention will be provided first, including: Step S1: Obtain multi-source drawing data of the control cabinet, extract the global topology features and text annotation features of the multi-source drawing data, determine the local pixel coordinates of each wiring terminal of the control cabinet based on the global topology features, and determine the initial wiring task set. The multi-source drawing data includes electrical schematic diagram, general assembly layout diagram and wiring table. Step S2: Determine the process rule subgraph based on the initial wiring task set, constrain the process rule subgraph to obtain numerical constraint indicators, and construct a dynamic process constraint feature matrix; Step S3: Determine structured process reasoning prompts based on the initial wiring task set and the dynamic process constraint feature matrix; determine the physical wiring sequence timing chain based on the structured process reasoning prompts and the dynamic process constraint feature matrix; determine the digital process instruction list based on the physical wiring sequence timing chain and the wiring table; and complete the control cabinet wiring planning. Specifically, embodiments of the present invention further include adjusting the digital process instruction list to obtain adjustment actions; using an action capture operator to record fine-grained features of the adjustment actions, arranging and wrapping the fine-grained features to obtain a dataset of differences before and after process correction; judging the process scheme containing the dataset of differences before and after process correction based on a bottom-line verification algorithm, and obtaining the final executable process scheme based on the judgment result.

[0022] Specifically, embodiments of the present invention further include determining a short text of process reflection summary based on the digital process instruction list and the final executable process scheme; projecting the short text of process reflection summary onto a preset high-dimensional semantic space using a preset industrial text embedding model to obtain an incremental experience feature vector; inputting the incremental experience feature vector into a long-term experience vector database, and retrieving incremental experience feature vectors with similarity scores greater than a preset threshold from the long-term experience vector database to complete the update of the long-term experience vector database.

[0023] Specifically, this invention extracts global topology and text annotations to determine terminal pixel coordinates and generate an initial wiring task set. Based on the task set, a process rule subgraph is constructed, constrained to obtain numerical indicators, and a dynamic process constraint feature matrix is ​​formed. Then, a physical wiring sequence chain is generated through structured reasoning prompts and the feature matrix, producing a digital process instruction list. Subsequently, action capture and bottom-line verification are used to record and judge the adjustment actions in a fine-grained manner to obtain an executable solution. Finally, the semantics of the process reflection summary are embedded and stored in a long-term experience vector database. Similar experiences are retrieved to achieve knowledge updates, automatically aligning spatial coordinates with electrical logic, reducing errors in manual drawing interpretation. The dynamic constraint matrix and structured reasoning work together to generate the optimal wiring sequence that meets electrical specifications and avoids cable interference, improving wiring efficiency and process rationality. The long-term experience database transforms manually corrected experience into searchable high-dimensional semantic vectors, possessing the ability for self-evolution of process knowledge, thereby improving the process standardization level of electrical control cabinet assembly operations.

[0024] Please see Figure 2 As shown, it is a structural schematic diagram of the control cabinet wiring planning system based on multimodal analysis and knowledge constraints provided in an embodiment of the present invention.

[0025] This invention also provides a control cabinet wiring planning system based on multimodal analysis and knowledge constraints, comprising: The multimodal parsing module is used to acquire multi-source drawing data of the control cabinet, extract the global topology features and text annotation features of the multi-source drawing data, determine the local pixel coordinates of each wiring terminal of the control cabinet based on the global topology features, and determine the initial wiring task set. The multi-source drawing data includes electrical schematic diagrams, assembly layout diagrams and wiring tables. The process knowledge management center, which is connected to the multimodal parsing module, is used to determine the process rule subgraph based on the initial wiring task set, constrain the process rule subgraph to obtain numerical constraint indicators, and construct a dynamic process constraint feature matrix. The path planning brain, which is connected to the multimodal parsing module and the process knowledge management center respectively, is used to determine the structured process reasoning prompt words based on the initial wiring task set and the dynamic process constraint feature matrix, determine the physical wiring sequence chain based on the structured process reasoning prompt words and the dynamic process constraint feature matrix, and determine the digital process instruction list based on the physical wiring sequence chain and the wiring table to complete the control cabinet wiring planning. A virtual rendering module, connected to the path planning brain, is used to adjust the digital process instruction list to obtain adjustment actions. It uses an action capture operator to record the fine-grained features of the adjustment actions, arranges and wraps the fine-grained features to obtain a dataset of differences before and after process correction, and judges the process scheme containing the dataset of differences before and after process correction based on the bottom-line verification algorithm. The final executable process scheme is obtained based on the judgment result. The process reflection module, connected to the virtual rendering module, is used to determine a short text of process reflection summary based on the digital process instruction list and the final executable process scheme. Using a preset industrial text embedding model, the short text of process reflection summary is projected onto a preset high-dimensional semantic space to obtain incremental experience feature vectors. The incremental experience feature vectors are input into a long-term experience vector database. Incremental experience feature vectors with similarity scores greater than a preset threshold are retrieved from the long-term experience vector database to complete the update of the long-term experience vector database.

[0026] Specifically, step S1 in this embodiment of the invention includes the following sub-steps: Step S101: Feature-level preprocessing of drawing data. The multi-modal parsing module receives multi-source drawing data from external input. The multi-source drawing data includes electrical schematic diagrams, general assembly layout diagrams, and structured wiring tables.

[0027] Understandably, due to differences in resolution, size, and scanning quality of the drawing files input from different source companies, the multimodal analysis module uses the image pyramid algorithm to perform multi-scale Gaussian filtering and downsampling processing on the input multi-source drawing data to construct a multi-scale spatial feature representation.

[0028] Specifically, the multimodal parsing module of this embodiment calls a preset fully convolutional network (FCN) to extract the global topological features and text annotation features of two-dimensional CAD drawings from multi-source drawing data.

[0029] To ensure low latency and deterministic feature representation when processing large-format industrial drawings, this embodiment of the invention employs a backbone network architecture based on an improved version of VGG-16 (Visual Geometry Group 16-layer network). The fully convolutional network comprises 13 convolutional layers, 3 max-pooling layers, and 3 upsampling deconvolutional layers. The core network layers and hyperparameters are configured as follows: the first two layers are convolutional layers with 3 input channels, 64 output channels, a 3×3 kernel size, and a stride of 1. At the end of the backbone network, the fully connected layer is removed and replaced with a transposed convolutional layer with an 8-span, adaptive order, to achieve 1:1 pixel-level alignment of input and output resolution. The overall learning rate of the fully convolutional network during training is preset to a certain value. The batch size is configured to 8, and gradient descent is performed using the Adam optimizer.

[0030] The multi-channel output of the fully convolutional network in this embodiment of the invention includes two sets of parallel pixel feature maps. The first channel outputs a binary segmentation mask of the cable and conductor region with a probability value greater than 0.85, i.e., a binary segmentation image, which is used to characterize the global topological structure features of the two-dimensional CAD drawing. The second channel output includes a text annotation feature response map containing the character outline edges, which is used to characterize the text annotation features.

[0031] In step S102, the multimodal analysis module performs orientation frame recognition on components such as circuit breakers, contactors, switching power supplies, and PLC modules in the overall assembly layout drawing, thereby completing the component positioning and wiring point analysis at the drawing level.

[0032] It is understandable that the rotation sensing detection of components and wiring points in step S102 is designed to address the characteristics of compact component layout and deflection interference in the overall assembly layout drawing of the electrical control cabinet. This is to achieve accurate sensing and positioning of components and wiring points on the drawing and to overcome the problem that general vision algorithms cannot directly extract the center position of the wiring terminals in the drawing.

[0033] Specifically, step S102 in this embodiment of the invention includes the following sub-steps: Step S1021: In this embodiment of the invention, for a specific industrial scenario where the PLC module is long and narrow and the relay group is densely and compactly arranged in the general assembly layout of the electrical control cabinet, the YOLO-obb rotating target detection network has been improved in a specific domain to ensure that the deflection angle is accurately calculated in the dense non-standard drawing symbols.

[0034] The deep neural network architecture of the YOLO-obb rotating target detection network in this embodiment of the invention includes: a feature extraction backbone network with 52 convolutional layers built based on CSP-DarkNet (Cross-Stage Local Dark Network); a neck network with 16 feature fusion layers built based on a path aggregation network; and three rotation detection decoupling heads containing classification and regression branches. On the shallow feature map of the Feature Pyramid Network (FPN), the aspect ratio set of the default square anchor boxes is expanded from the conventional {1:1, 1:2, 2:1} to {1:5, 5:1, 1:8, 8:1} specifically adapted to narrow electrical components. The network input resolution is normalized to 1024×1024 pixels, the initial training learning rate is preset to 0.01, and the weight decay coefficient is configured as follows: .

[0035] To address the issue of missed detections caused by overlapping bounding boxes when components are densely arranged, this invention optimizes the loss function from the standard Smooth-L1 loss function to a Bayesian uncertainty-constrained rotating bounding box loss function based on rotation angle awareness. Represented as: , in, For classifying losses, For bounding box regression loss, and These are the predicted and actual oriented bounding boxes, respectively. and These are the rotation angles of the predicted and actual bounding boxes, respectively. These are the preset balance coefficients.

[0036] It is understood that the embodiments of the present invention suppress the positioning ambiguity when recognizing overlapping high-density terminal blocks by adding a cosine penalty term for the angle difference to the loss function.

[0037] The YOLO-obb rotating target detection network of this embodiment outputs an oriented bounding box containing five parameters. Represented as: , in, The center pixel coordinates of the component. This represents the width of the component's bounding box. This represents the height of the component's bounding box. This represents the relative deflection angle of the oriented bounding box relative to the reference horizontal axis.

[0038] In step S1022, in order to purify electrical topology connectivity from the binary segmentation mask of the cable conductor region extracted from the fully convolutional network, this embodiment of the invention introduces a connectivity algorithm based on graph theory skeleton extraction.

[0039] In this embodiment of the invention, the multimodal parsing module obtains the first channel data output by the fully convolutional network, namely the binary segmentation mask of the cable conductor region, and calls the central axis transformation operator to refine the image of the solid line pixel segment, extracting the center line grid skeleton with a pixel width of 1 that represents the electrical wiring; then, the graph topology scanner finds the endpoints (Degree=1) and bifurcation intersections (Degree≥3) in the center line grid skeleton, and maps the pixel connected regions to discrete geometric topological edges to filter out the interference of redundant frame lines and text shadows in the background of the drawing.

[0040] The multimodal parsing module in this embodiment of the invention calls the built-in local key point localization algorithm. The execution process of the local key point localization algorithm is as follows: taking the center pixel coordinates of the component as the origin, an inverse rotation transformation matrix is ​​constructed based on the relative deflection angle. The oriented bounding box is cropped and corrected to an orthogonal rectangular region image. Then, the orthogonal rectangular region image is input into a stacked hourglass network to extract spatial feature maps containing multi-scale context; finally, based on a... The convolutional layer outputs a set of terminal block response heatmaps ,in Indicates the first Spatial probability distribution of each terminal block.

[0041] The multimodal analysis module of this invention calculates the position of each terminal in the digital pixel coordinate system by finding the maximum point in the terminal response heatmap. Local pixel coordinates of the terminal .

[0042] Specifically, this invention performs multi-scale preprocessing on multi-source drawings using an image pyramid and an improved fully convolutional network to extract cable topology segmentation masks and text annotation features. It introduces the YOLO-obb rotating target detection network, optimized for elongated components, and utilizes an angle-aware Bayesian uncertainty loss function to accurately calculate the oriented bounding boxes of connection points. Then, based on central axis transformation and graph topology scanning, it extracts the centerline skeleton representing electrical connectivity from the segmentation mask, filtering out non-electrical interference. Finally, it outputs a terminal response heatmap through inverse rotation transformation and a stacked hourglass network to obtain the local pixel coordinates of each connection terminal, generating an initial wiring task set. The combination of multi-scale pyramids and fully convolutional networks eliminates information attenuation caused by differences in enterprise drawing resolution and size, achieving precise pixel-level alignment of topology and text features. This ensures consistency between the mapping of wiring logic and physical coordinates, improves the detection rate and deflection calculation accuracy of high-density terminal blocks, and enhances the extraction accuracy and robustness of terminal pixel coordinates, thereby improving the standardization level of electrical control cabinet assembly operations.

[0043] Step S103: In order to seamlessly bind the abstract connection relationship of the electrical schematic diagram with the actual physical spatial location of the assembly layout diagram and eliminate heterogeneous gaps between drawings, the multimodal analysis module constructs a logical mapping matrix in this embodiment of the invention. .

[0044] The data structure of the logical mapping matrix in this embodiment of the invention is defined as a matrix of size . A two-dimensional Boolean matrix, where, This represents the total number of all abstract logic terminals extracted from the electrical schematic. This represents the total number of all physical entity terminals identified in the assembly layout drawing. The initial state of the two-dimensional Boolean matrix is ​​an all-zero matrix, meaning that for any element... It continuously calls a two-dimensional Boolean matrix to trace the source of the electrical circuit corresponding to any physical intervention action on the electrical schematic diagram.

[0045] The logical mapping matrix construction and cross-drawing semantic alignment process in this embodiment of the invention are executed collaboratively by the following sub-steps: Step S1031: Extract the logic network topology diagram from the electrical schematic diagram. In this embodiment of the invention, to transform the disordered lines of the electrical schematic diagram into a graph data structure that allows for relational reasoning by a computer and to explicitly represent all hidden electrical control loops, the multimodal parsing module extracts the logic network topology diagram from the electrical schematic diagram. Represented as: , in, ， , in, For logic terminal set, It is a set of logically connected edges.

[0046] The process of acquiring and constructing the logical terminal set in this embodiment of the invention is as follows: The multimodal parsing module acquires the text annotation feature response map output by the fully convolutional network, and uses the EAST directed text line extraction operator to locate all component identification characters (such as "KM1") and terminal number characters (such as "A1") in the electrical schematic diagram; according to the preset electrical symbol geometric neighborhood connectivity rules, the component identification characters and terminal number characters whose literal distance is less than a preset pixel threshold are concatenated to form a unique logical terminal string (such as "KM1-A1"). Each unique logical terminal string serves as a vertex in the logical terminal set. .

[0047] In this embodiment of the invention, the preset pixel threshold is 45 pixels, but the above value is not limited to this, and those skilled in the art can adjust the value according to actual needs.

[0048] The process of obtaining and constructing the logically connected edge set in this embodiment of the invention is as follows: The multimodal parsing module performs binarization processing on the non-text pixel areas in the electrical schematic diagram, calls the flooding fill algorithm to trace the connected paths of all black connecting thin line pixels, and when the flooding fill algorithm detects that a pixel wire path can continuously connect two different logical terminal strings in space, it determines that there is an electrical logic path between these two vertices and generates an undirected edge. Write it into the logically connected edge set.

[0049] Step S1032: Constructing a physical space point set. In this embodiment of the invention, in order to transform the graphical assembly layout diagram into a discrete geometric data volume containing clear spatial physical coordinates and physical orientations, the multimodal analysis module constructs a physical space point set from the physical entity terminals identified in the assembly layout diagram. Represented as: , in, For the first A physical point in space.

[0050] Each physical space point in the physical space point set of this invention corresponds to a real physical entity terminal on the overall assembly layout diagram.

[0051] The construction process of the physical space point set in this embodiment of the invention is as follows: the multimodal parsing module takes the local pixel coordinates of the terminal calculated by the local key point localization algorithm and encapsulates each identified physical terminal into a structured data node containing three-dimensional spatial coordinates, orientation and associated component model features.

[0052] In this embodiment of the invention, physical terminal encapsulation can be achieved by packaging data into JSON key-value pair objects. The packaging process is a common technique in the field and will not be described in detail here.

[0053] Step S1033: Calculate the alignment confidence score and construct the logical mapping matrix. In this embodiment of the invention, in order to establish accurate entity associations in multi-source drawing data, the multimodal parsing module calls the text string fuzzy matching algorithm and the context semantic association algorithm to compare the logical terminal set with the physical space point set. The execution process of the text string fuzzy matching algorithm is as follows: The multimodal parsing module extracts text labels from logic terminal strings in the electrical schematic (e.g., "KM1-A1") and text labels from physical space points in the assembly layout (e.g., "KM1:A1"). It then calls a text string fuzzy matching algorithm based on edit distance to calculate the literal similarity score between the two text labels. Represented as: , in, This represents the standard edit distance between two text tags. Indicates the character length of the text label. The text label for the logical terminal string. Text labels for physical spatial points.

[0054] To prevent errors in literal matching due to partial character corruption or text truncation, this invention employs a contextual semantic association algorithm for collaborative verification. The multimodal parsing module extracts the circuit information (such as the wire number "No.101" of the current circuit and other component types "circuit breaker / QF1" in the logical network topology of the electrical schematic diagram) from the vertices of the logical terminal string, constructing a logical semantic context vector. Simultaneously, the set of adjacent components physically connected by wire slots in the overall assembly layout diagram is extracted to construct a physical space context vector. The context semantic association score is output by calculating the cross-attention product of the logical semantic context vector and the physical space context vector. .

[0055] The alignment processor in this embodiment of the invention performs a weighted fusion of the literal similarity score and the contextual semantic association score to obtain a comprehensive alignment confidence score. : , in, This is the preset weight adjustment coefficient.

[0056] The alignment processor in this embodiment of the invention traverses all logical and physical terminal pairs to find the one that maximizes the overall alignment confidence score and exceeds a preset confidence threshold. The pairing relationship is determined, and once a successful pairing is determined, the corresponding number in the logical mapping matrix is... line, number The element of the column is forcibly assigned the value 1, that is: .

[0057] Step S1034: Generate an initial wiring task set. In this embodiment of the invention, after the logic mapping matrix is ​​filled and constructed, the multimodal parsing module translates the logical connected edge set of the logical network topology diagram of the electrical schematic diagram into wiring tasks in the physical space based on the index of all elements with a value of 1 in the logic mapping matrix. For any logical connected edge in the logical connected edge set, the alignment processor retrieves the row index of the corresponding vertex in the logic mapping matrix and extracts the corresponding column index in reverse, thereby accurately locking the three-dimensional physical space coordinates of the corresponding source and destination terminals in the physical space point set.

[0058] The multimodal parsing module of this invention, based on the wiring tasks in the translated physical space, assembles and outputs a structured initial wiring task set. : , in, For the first A separate wiring task.

[0059] The multimodal parsing module of this invention expresses each independent wiring task as a tuple structure: ， in, For wire number, The physical coordinates of the source terminal. The physical coordinates of the terminal are: The source terminal is a set of logical attributes that include current level and signal type, and the sink terminal is a set of logical attributes that include current level and signal type.

[0060] Specifically, this invention constructs a logical mapping matrix to align the abstract logical terminals of the electrical schematic diagram with the physical entity terminals of the assembly layout diagram across drawings. This enables precise binding between abstract electrical logic and physical spatial location, eliminating the high error rate and cognitive burden of manual diagram translation. The dual-path alignment algorithm, which integrates literal similarity and topological context semantics, is highly robust to character corruption, ambiguous identification, and non-standard naming on the drawings. Even if some text is missing, reliable matching can still be achieved through loop neighbor relationships, reducing the mapping failure rate and avoiding assembly rework such as insufficient cable length or incorrect terminal insertion due to coordinate mismatch. This improves the response speed and traceability of process design, thereby enhancing the standardization level of electrical control cabinet assembly operations.

[0061] Specifically, step S2 in this embodiment of the invention includes the following sub-steps: Step S201: Retrieve the associated subgraph of the electrical assembly knowledge graph. In this embodiment of the invention, a pre-constructed global electrical assembly knowledge graph is deployed within the process knowledge management center. The global electrical assembly knowledge graph stores heterogeneous knowledge from multiple sources, including the electrical properties of components, physical assembly process specifications, and electromagnetic compatibility standards.

[0062] Understandably, step S201 transforms a vast amount of fragmented industry cabling standards and expert tacit experience into structured rules that can be automatically invoked by computers, thereby reducing the knowledge retrieval blind spots of large models. In this embodiment of the invention, the process knowledge management center inputs the component models, circuit rated voltages, and maximum carrying currents from the initial wiring task set into the retrieval slots of the global electrical assembly knowledge graph. The process knowledge management center then invokes a depth-first jump search algorithm to perform an extended search on the global electrical assembly knowledge graph, extracting process rule subgraphs associated with the wiring tasks. .

[0063] Step S202: Numerical constraint conversion of process specifications. In this embodiment of the invention, the process knowledge management center calls the built-in compiler mapping function. The textual semantic descriptions in the process rule sub-graphs are transformed into computable numerical constraint indices.

[0064] The specific implementation process of the compilation mapping function in this embodiment of the invention is as follows: the control quantities in the process text are defined as discrete technical feature items in advance. When the process rule subgraph is input, the compilation processor parses the electrical dependency chain in the process rule subgraph through the symbolic relation mapping operator and automatically compiles and translates the text rules into calculable numerical constraint indicators.

[0065] The numerical constraint indicators in this invention include electromagnetic compatibility spacing constraints, slot capacity constraints, and terminal physical constraints.

[0066] The electromagnetic compatibility spacing constraint in this embodiment of the invention is that the process knowledge management center outputs the minimum physical isolation distance threshold for laying strong and weak current cables based on the voltage level difference between the source terminal and the destination terminal. .

[0067] In this embodiment of the invention, when the loop potential difference is greater than 220V, the minimum physical isolation distance threshold is set to 50mm; when the loop potential difference is less than or equal to 24V, the minimum physical isolation distance threshold is set to 15mm.

[0068] In this embodiment of the invention, the capacity constraint of the cable trays is that the process knowledge management center retrieves the physical cross-sectional area of ​​the cable trays within the control cabinet and outputs the maximum allowable volumetric filling rate limit value. .

[0069] In this embodiment of the invention, the maximum volume filling rate limit is set to 60%, that is, the sum of the cross-sectional areas of all cables in the cable tray shall not exceed 0.60 of the total theoretical cross-sectional area of ​​the cable tray.

[0070] In this embodiment of the invention, the physical constraint of the terminals is that the process knowledge management center outputs a standard tightening torque threshold based on the material of the terminal block and the bolt specifications. And the range of conductor cross-sectional area matching.

[0071] In this embodiment of the invention, for M4 specification copper combination screw terminals, the standard tightening torque threshold is set to 1.2 N·m.

[0072] Step S203: Construction of dynamic environmental interference matrix. In order to enable the generative process inference model to have spatial anti-collision and environmental adaptive perception capabilities when generating wiring process schemes, this embodiment of the invention needs to obtain the actual physical space obstacle boundaries inside the control cabinet. To this end, the process knowledge management center first obtains the three-dimensional point cloud data of the physical environment inside the control cabinet.

[0073] In this embodiment of the invention, the three-dimensional point cloud data is defined as a discrete, unstructured, high-dimensional spatial point set representing the three-dimensional coordinates of all spatial surfaces inside the control cabinet to be wired.

[0074] The specific acquisition process of the three-dimensional point cloud data in this embodiment of the invention is as follows: a three-dimensional vision sensor is deployed in the digital offline planning operation space of the control cabinet, including at least one fixed binocular camera placed on the front and upper side of the control cabinet, or an airborne high-precision depth camera deployed on the end effector of the robotic arm.

[0075] Before the wiring scheme planning is initiated, the multimodal analysis module of this invention controls the three-dimensional vision sensor network to perform omnidirectional multi-view optical image acquisition on the physical shell inside the control cabinet to be assembled, the metal mounting back plate, the fixed electrical components (such as intermediate relays, contactors, circuit breakers, PLC modules), and the fixed guide rails and rigid cable trays.

[0076] The multimodal parsing module of this invention inputs the multi-view local raw point clouds collected by binocular cameras or depth cameras into the internally deployed point cloud registration algorithm. It uses the feature point matching operator to calculate the spatial transformation matrix, thereby stitching and transforming the local point clouds under each viewpoint into the global physical three-dimensional Cartesian coordinate system of the control cabinet. After performing voxel downsampling feature dimensionality reduction and noise reduction smoothing, it finally outputs high-resolution three-dimensional point cloud data of the internal surface of the control cabinet.

[0077] In this embodiment of the invention, after acquiring the three-dimensional point cloud data, the process knowledge management center calls the built-in spatial geometric topology partitioning algorithm to identify the boundary of obstacle point clouds belonging to guide rail fixing bolt protrusions, grounding copper busbars, structural supports, and heat dissipation areas of heat-generating components in the three-dimensional point cloud data. By extracting the minimum bounding boxes of these obstacle point cloud boundaries in a three-dimensional rectangular coordinate system, the spatial no-entry zones inside the control cabinet are constructed and defined. The spatial forbidden zone represents a set of absolutely rigid geometric boundaries that cannot be traversed by cable process trajectories.

[0078] In this embodiment of the invention, the process knowledge management center performs real-time statistics on the volume occupancy data of the planned cables in each segment of the cable tray. In order to achieve a quantitative and real-time representation of the cabling environment, the process knowledge management center constructs a dynamic process constraint feature matrix.

[0079] The data structure of the dynamic process constraint feature matrix in this embodiment of the invention is defined as a three-dimensional spatial voxel mesh matrix. Represented as: , in, The granularity of the discrete mesh in the length, width, and height directions of the three-dimensional physical space inside the control cabinet are respectively defined. The number of feature channels is [number]. Each feature channel stores the electromagnetic radiation intensity feature score of the current spatial grid point, the volume occupancy of the current slot segment, and the physical passability feature score of the obstacle.

[0080] The dynamic update logic and state transition process of the dynamic process constraint feature matrix in this embodiment of the invention are as follows: In the initial state, the volume occupancy channels of the cable tray mesh are all zero. When the process instruction for each cable is successfully planned, the process knowledge management center performs real-time accumulation and dynamic update of the volume occupancy features and electromagnetic radiation intensity features in each segment of the cable tray through the state transition equation, as shown below: ， , in, and These are the real-time fill rate scores for the slot segments to which the grid points belong before and after the update. To plan the physical diameter of the conductor, This refers to the standardized cross-sectional area of ​​the corresponding cable tray segment; and These are the cumulative electromagnetic radiation intensity values ​​of the grid points before and after the update, respectively. The rated current intensity of the conductor. The spatial geometric distance from a point in the grid space to the centerline of the traverse line. These are radiation weighting coefficients related to signal properties.

[0081] In this embodiment of the invention, in order to perform extreme hard boundary verification of high-voltage interference, the rated current intensity is set to 32A for the power bus inside the control cabinet; and the rated current intensity is set to 2A for the control low-voltage circuit.

[0082] This invention achieves the time-series dynamic decay update of cable tray space occupancy and interference field through the state transition equation, providing accurate environmental input for subsequent cable obstacle avoidance and isolation planning.

[0083] Specifically, this invention utilizes a global electrical assembly knowledge graph, employing a depth-first extended search based on the component and circuit parameters in the initial task set as slot locations. It extracts associated process rule subgraphs and then automatically translates the textual semantics of these subgraphs into calculable numerical indicators such as electromagnetic compatibility spacing, cable tray capacity, and terminal physical constraints using a compilation mapping function. Based on 3D point cloud data, a dynamic process constraint feature matrix is ​​constructed, incorporating multiple channels including electromagnetic radiation intensity, cable tray volume occupancy, and obstacle passability. The matrix is ​​then dynamically updated in real-time using state transition equations after each cable planning step. Collision detection and electromagnetic compatibility verification provide precise constraint criteria, avoiding interference from mixed strong and weak current wiring and the risk of cable tray overload. Dynamic tracking of cable tray filling rate and the cumulative effect of electromagnetic radiation enables the wiring planning to possess realistic physical environment perception and adaptive obstacle avoidance capabilities. This achieves global sequential optimization of multi-cable wiring, reduces rework caused by improper wiring sequence, improves the success rate of initial wiring and assembly efficiency, and ultimately enhances the standardization level of electrical control cabinet assembly operations.

[0084] Specifically, step S3 in this embodiment of the invention includes the following sub-steps: Step S301: Combining the generative reasoning prompts of the process specifications, the path planning brain of this embodiment of the invention is equipped with an autoregressive generative process reasoning algorithm that has been incrementally supervised and fine-tuned on a dataset in the field of electrical engineering. The generative process reasoning algorithm is built on an autoregressive Transformer (autoregressive attention network) architecture based on the self-attention mechanism.

[0085] The context prompt word group leader module of this embodiment first obtains the initial wiring task set, the dynamic process constraint feature matrix, and the historical wiring experience vector retrieved from the long-term experience vector database.

[0086] The contextual hint word group long module of this embodiment of the invention calls the preset... The (Reasoning and Acting) framework combines the features of each independent wiring task in the initial wiring task set (i.e., wire number, source terminal physical coordinates, destination terminal physical coordinates, and logical attribute set), the dynamic process constraint feature matrix, and the historical successful wiring experience vectors into a vector-level lossless concatenation to generate structured process reasoning prompts that simultaneously satisfy both physical and process logic boundaries. .

[0087] After receiving the structured process reasoning prompt, the generative process reasoning algorithm of this embodiment of the invention performs the following reasoning decision and high-dimensional solution process: The algorithm forcibly introduces a penalty weight mask matrix into the underlying self-attention mechanism layer. The minimum negative infinity blocking operator is applied directly at the probability distribution solution level to the generation probability of illegal cable tray segments that violate the spatial isolation distance between strong and weak currents or the cable tray capacity limit. The blockade forces the underlying autoregressive probability prediction network to automatically focus and select the digital process instruction list that has the highest compliance probability and contains the nine-tuple standardized physical process parameters as the output. .

[0088] This deterministic discrete solution process effectively eliminates the "logic illusion" problem of general large language models, ensuring the absolute compliance and high executability of industrial planning instructions.

[0089] Step S302: The wiring instruction sequence logic decision satisfies the physical specifications. The timing logic decision module of the path planning brain in this embodiment of the invention receives the structured process reasoning prompts. The timing logic decision module performs physical attribute classification on the tasks in the initial wiring task set based on the slot filling rate and electromagnetic isolation requirements in the dynamic process constraint feature matrix.

[0090] In this embodiment of the invention, the timing logic decision module calls the process priority weighting operator to perform sorting processing on the initial wiring task set, and arranges and outputs a timing chain of physical wiring sequence. .

[0091] The weighting logic of the process priority weighting operator in this embodiment of the invention is as follows: the routing timing of large-diameter power lines is superior to that of low-voltage signal lines; the routing timing of large-section conductors is superior to that of small-section conductors; and the routing timing of long-distance cables is superior to that of short-distance cables. For any two wiring tasks... and The formula for calculating priority weights is: , in, These are, respectively, the area-weighted coefficient, the distance-weighted coefficient, and the signal attribute penalty coefficient. The cross-sectional area of ​​the conductor is... To estimate the span distance, This is the encoded value for the signal attribute.

[0092] In this embodiment of the invention, the timing logic decision module constructs the physical wiring sequence timing chain according to the priority weights from largest to smallest.

[0093] In this embodiment of the invention, an area weighting coefficient is set. Distance weighting coefficient Signal attribute penalty coefficient .

[0094] Step S303: Direct generation of digital wiring table and process list. The process reasoning algorithm of the path planning brain in this embodiment of the invention solves the wiring task one by one according to the physical wiring sequence time chain. When the process reasoning algorithm solves and generates the current cable routing slot sequence, it calls the rule verification operator in real time and projects the current candidate spatial path onto the dynamic process constraint feature matrix for safety margin measurement.

[0095] To force generative inference algorithms to generate instructions that fully comply with electrical safety regulations, this invention employs a penalty weight embedding mechanism based on attention masks. When the rule verification operator detects that the real-time filling rate of the spatial voxel grid points of the current candidate slot segment exceeds the maximum volumetric filling rate limit, the penalty weight embedding mechanism is implemented. ,Right now Or, the cumulative electromagnetic radiation intensity value of the grid point where the weak signal line is located is detected. When the radiation exceeds the safe radiation threshold, the rule verification operator triggers a violation interception.

[0096] The process inference algorithm of this invention implements the specific embedding of penalty weights by forcibly introducing a penalty weight mask matrix into the self-attention mechanism layer that generates large models: , in, These are the query matrix, key matrix, and value matrix of the inference network, respectively. For feature scaling.

[0097] The assignment logic of the penalty weight mask matrix in this embodiment of the invention satisfies: , This invention, through the embedding of a minimal negative infinity blocking operator, completely blocks and eliminates the probability of the large model selecting the non-compliant slot segment at the probability distribution solution level, thereby forcibly driving the process inference algorithm to automatically switch and focus on the backup compliant slot segment with the highest solution probability.

[0098] The process reasoning module of this embodiment directly outputs a digital process instruction list, wherein each line of process instruction in the digital process instruction list... Expressed in the form of a nine-tuple: ， in, For the specified suggested path index sequence along the slot segment, The cross-sectional area of ​​the conductor is... For wire color standard, For the casing wire number text, The standard cut length for cables that includes end allowance. This refers to the bolt tightening torque.

[0099] Specifically, this invention constructs a structured prompt word by concatenating an initial wiring task set, a dynamic process constraint feature matrix, and a historical wiring experience vector. A process priority weighting operator sorts tasks such as large-diameter, long-distance power lines, generating a physical wiring sequence chain. When calculating each cable tray segment, a penalty weighting mechanism based on attention masks is introduced to block candidate tray segments that violate tray filling rate or electromagnetic safety thresholds with negative infinity, forcing the model to focus on outputting compliant paths. Finally, a digital process instruction list is generated. The structured prompt word integrates task geometric coordinates, real-time environmental voxel states, and historical success experiences, enabling the generated model to simultaneously possess physical space perception, standard constraint memory, and past wisdom reference. The wiring sequence automatically arranged by the priority weighting operator avoids on-site problems such as cables laid earlier being squeezed by later cables and strong / weak current cross-interference, reducing rework rates, compressing manual translation and on-site adjustment time, improving the first-time success rate of control cabinet wiring and the level of production line automation, thereby enhancing the standardization level of electrical control cabinet assembly operations.

[0100] Specifically, step S4 in this embodiment of the invention includes the following sub-steps: Step 401: Transparent display of the process decision-making chain. In this embodiment of the invention, the virtual rendering module obtains the digital process instruction list and the three-dimensional geometric model of the control cabinet. The virtual rendering module calls the spatial analytical geometry operator to convert the suggested path index sequence along the slot segment into a spatial continuous spline curve and overlays it on the three-dimensional geometric model of the control cabinet.

[0101] In this embodiment of the invention, the virtual rendering module dynamically displays specific specification entries in the global electrical assembly knowledge graph associated with the current rendering path within a designated floating area of ​​the display interface, thereby realizing the white-box output of the path planning brain decision link.

[0102] Step S402: Fine-grained data collection of auditor correction behavior. The human interaction interface of the virtual rendering module in this embodiment of the invention captures the adjustment actions of the human auditor on the digital process instruction list. The adjustment actions include changing the wire routing groove section, manually forcibly isolating specific lines, adjusting the cable quota cut length, and modifying the crimp terminal model.

[0103] In this embodiment of the invention, the data acquisition module of the virtual rendering module calls the action capture operator to record the fine-grained features of the adjustment action, wherein the fine-grained features include the unique code of the operated cable. Action operator type, initial recommended parameter values ​​before modification and the actual execution parameter values ​​after manual modification The data acquisition module arranges and wraps the fine-grained features to generate a dataset showing the differences before and after the process correction. .

[0104] Step S403: Automated logic verification for correcting conflicts. In this embodiment of the invention, the bottom-line verification module of the virtual rendering module receives the dataset of differences before and after the process correction, and the bottom-line verification module retrieves a preset set of red-line rules.

[0105] The red-line rule set in this embodiment of the invention includes electrical short-circuit rules, cable tray physical capacity limit rules, and anti-dead-bend safety radius rules. The bottom-line verification algorithm performs matrix judgment on the process scheme containing the difference dataset before and after the process modification. If no rule in the red-line rule set is triggered, the bottom-line verification algorithm outputs a pass signal and generates the final executable process scheme. .

[0106] In order to solve the disconnect between theoretical offline planning process and actual physical operation on site caused by mechanical deformation and assembly tolerance, this invention introduces a flexible deformation error compensation mechanism based on on-site visual sensor network before the final executable process plan is issued to the on-site automated wiring equipment.

[0107] The execution process of on-site visual registration and path deformation compensation in this embodiment of the invention is as follows: An airborne binocular camera is configured on the end effector of the automated wiring equipment. When the wiring operation is started, the airborne binocular camera performs local three-dimensional feature point scanning on the terminal blocks and wire troughs that are actually assembled in the current control cabinet to obtain the current on-site registration geometric point cloud data.

[0108] The error processing operator of the virtual rendering module in this embodiment of the invention compares the on-site registered geometric point cloud data with the theoretical model. When it is determined that there is a spatial geometric tolerance between the actual assembly position of the current physical cabinet and the theoretical spatial pose, the operator performs a comparison. At that time, the error processing operator uses the spatial geometric tolerance as the control boundary condition and calls the local elastic deformation compensation algorithm. The local elastic deformation compensation algorithm uses the spline curve control point flexible perturbation function to adaptively adjust the continuous physical space routing path calibrated in the final executable process plan: , in, This is the spline curve of the actual physical trace after error compensation. As basis functions, This refers to the spatial displacement vector of the control point calculated based on the aforementioned geometric tolerances.

[0109] This invention, through adaptive compensation of physical deformation of the path, ensures that the wiring trajectory sent to the automated robot can accurately conform to the subtle tolerances inside the non-standard control cabinet, eliminating the potential for damage to the robot due to collisions with component boundaries.

[0110] Specifically, this invention visualizes and renders the generated digital process instruction list in a 3D model of the control cabinet, and displays relevant specification items in a floating manner to achieve white-box decision-making. An action capture operator records adjustment actions in real time, generating heterogeneous datasets. A baseline verification module is invoked to perform automated matrix judgment on the corrected scheme. Upon approval, a final executable process scheme is generated. Granular adjustment action capture structurally stores the implicit experience of manual corrections as difference data. This data not only forms the basis for single scheme optimization but can also flow back to knowledge graphs and vector databases, transforming each manual intervention into reusable incremental experience for the system. This drives the continuous evolution of process reasoning capabilities, improves the efficiency and accuracy of process review, increases the first-time success rate of automated wiring, and enhances equipment operational reliability, thereby improving the standardization level of electrical control cabinet assembly operations.

[0111] Specifically, step S5 in this embodiment of the invention includes the following sub-steps: Step S501: Reflective process summary generation with corrected causes. The reflective summary algorithm of the process reflection module in this embodiment of the invention obtains the independently output digital process instruction list, the final executable process scheme, and the physical context features of the internal components of the control cabinet.

[0112] The reflective summarization algorithm of this invention utilizes a built-in logical difference comparator to calculate the wire number between the digitized process instruction list and the final executable process scheme. The semantic distance is calculated below. The reflective summarization algorithm, combined with a component attribute database, identifies the physical attributes of components surrounding the modification site and invokes natural language generation operators to output a structured short text of process reflective summaries. .

[0113] In order to achieve a deeper and more generalized closed-loop evolution of industrial knowledge, this embodiment of the invention synchronously incorporates the multimodal feedback signals of physical operations generated at the physical assembly site into the generation of the short text of the process reflection summary.

[0114] The flow direction and parameter correction logic of the multimodal feedback signal in this embodiment of the invention are as follows: The joint servo sensor and torque monitoring operator inside the automated wiring equipment collect physical characteristic quantities in real time during the physical execution process of wiring. The physical characteristic quantities include the actual tightening resistance torque feedback value of the terminal screw. And the axial reaction tension feedback value when the wire is inserted. .

[0115] The data feedback unit of the process reflection module in this embodiment of the invention transmits the actual fastening resistance torque feedback value and the axial reaction tension feedback value as multi-modal feedback signals back to the reflection summarization algorithm in real time. The reflection summarization algorithm compares the data with theoretical prior parameters (such as standard torque thresholds). When a persistent deviation in physical execution is detected (such as an excessively large actual crimping torque due to terminal block batch tolerance), a "parameter drift calibration rule" is automatically added to the short text of the process reflection summary.

[0116] This invention, through the reverse flow of hardware and software signals, enables the adaptive evolution of the planning system to go beyond the "path position" level and penetrate deeper into the dynamic correction of the underlying "material physical properties and assembly mechanics prior parameters".

[0117] Step S502: Vectorized representation of process reflection summary. In this embodiment of the invention, the process reflection module calls a preset industrial text embedding model to receive the structured short text of process reflection summary.

[0118] The industrial text embedding model of this invention projects the short text of the process reflection summary into a preset high-dimensional semantic space, performs matrix multiplication and weighted calculation, and outputs a fixed-length dense feature vector as an incremental empirical feature vector. : , in, , The dimension is the pre-defined high-dimensional semantic space.

[0119] Step S503: Incremental update of the long-term experience vector library. In this embodiment of the invention, the process reflection module appends the incremental experience feature vector to the long-term experience vector database of the process knowledge management center.

[0120] In this embodiment of the invention, the management module of the long-term experience vector database calls the HNSW graph index incremental update operator to reconstruct the local index structure. When performing subsequent control cabinet wiring process planning, the context cue word assembler of the path planning brain calculates the context feature vector of the new wiring planning task. The cosine similarity calculation formula is called, and the cosine similarity is used as a similarity score to perform a search in a long-term empirical vector database: , in, For the first in the long-term empirical vector database A vector of historical experience.

[0121] In this embodiment of the invention, incremental empirical feature vectors with similarity scores greater than a preset threshold are extracted and simultaneously input into the structured process reasoning prompts in step S301, thereby achieving closed-loop evolution of process planning capabilities.

[0122] Specifically, this invention compares the differences between the digital process instruction list and the final executable process scheme using a reflective summarization algorithm. It generates short reflective summaries of the process by combining the physical properties of components, and simultaneously incorporates physical feedback signals transmitted in real-time from automated wiring equipment. Subsequently, an industrial text embedding model is used to project the reflective summaries into a high-dimensional semantic space, generating incremental experience feature vectors. Finally, these incremental vectors are written into a long-term experience vector database, and highly relevant historical experiences are retrieved using cosine similarity. Structured reasoning prompts are injected to achieve a closed-loop evolution of process planning capabilities, calibrate implicit physical deviations caused by terminal block batch tolerances, and dynamically correct underlying process prior parameters. This continuously improves the rationality and compliance of the initial planning, shortens the process design cycle for new models and cabinet structures, and reduces the frequency of on-site manual intervention and rework costs, thereby improving the standardization level of electrical control cabinet assembly operations.

[0123] Please see Figure 3 As shown, it is a schematic diagram of an electronic device provided in an embodiment of the present invention. Figure 3 The electronic device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0124] Specifically, the electronic device 300 includes a processor 310, a memory 320, and a communication interface 330, which are interconnected and communicate with each other via a communication bus 340 and / or other forms of connection mechanism (not shown).

[0125] The memory 320 includes one or more (only one is shown in the figure), which may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The processor 310 and other possible components may access the memory 320 to read and / or write data therein.

[0126] Processor 310 includes one or more (only one is shown in the figure), which can be an integrated circuit chip with signal processing capabilities. The processor 310 described above can be a general-purpose processor, including a central processing unit (CPU), a microcontroller unit (MCU), a network processor (NP), or other conventional processors; it can also be a special-purpose processor, including a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0127] The communication interface 330 includes one or more (only one is shown in the figure) and can be used to communicate directly or indirectly with other devices to exchange data. For example, the communication interface 330 can be an Ethernet interface; it can be a mobile communication network interface, such as an interface for 3G, 4G, or 5G networks; or it can be other types of interfaces with data transmission and reception functions.

[0128] One or more computer program instructions can be stored in the memory 320. The processor 310 can read and run these computer program instructions to implement the control cabinet wiring planning method based on multimodal analysis and knowledge constraints provided in the embodiments of this application, as well as other desired functions.

[0129] Understandable. Figure 3 The structure shown is for illustrative purposes only; the electronic device 300 may also include components that are more advanced than those shown. Figure 3 The more or fewer components shown, or having the same Figure 3 The different configurations shown. Figure 3 The components shown can be implemented using hardware, software, or a combination thereof. For example, electronic device 300 can be a single server (or other device with computing power), a combination of multiple servers, a cluster of a large number of servers, etc., and can be either a physical device or a virtual device.

[0130] This application also provides a computer-readable storage medium storing computer program instructions. These instructions are read and executed by a computer processor to perform the control cabinet wiring planning method based on multimodal analysis and knowledge constraints provided in this application. For example, the computer-readable storage medium can be implemented as... Figure 3 The memory 320 in the electronic device 300.

[0131] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present application should be included within the scope of protection of the present application.

Claims

1. A multi-modal analysis and knowledge constraint based control cabinet wiring planning method, characterized in that, The method comprises the following steps: obtaining multi-source drawing data of a control cabinet, extracting global topological structure features and text annotation features of the multi-source drawing data, determining terminal local pixel coordinates of each terminal of the control cabinet based on the global topological structure features, and determining an initial wiring task set, wherein the multi-source drawing data comprises an electrical schematic diagram, an assembly layout diagram and a wiring table; determining a process rule subgraph based on the initial wiring task set, obtaining a numerical constraint index by constraining the process rule subgraph, and constructing a dynamic process constraint feature matrix; determining a structured process reasoning prompt word based on the initial wiring task set and the dynamic process constraint feature matrix, determining a physical wiring sequence time sequence chain according to the structured process reasoning prompt word and the dynamic process constraint feature matrix, determining a digital process instruction list according to the physical wiring sequence time sequence chain and the wiring table, and completing control cabinet wiring planning.

2. The multi-modal parsing and knowledge constraint based control cabinet wiring planning method according to claim 1, characterized in that, The step of obtaining multi-source drawing data of a control cabinet, extracting global topological structure features and text annotation features of the multi-source drawing data, determining terminal local pixel coordinates of each terminal of the control cabinet based on the global topological structure features, and determining an initial wiring task set comprises: extracting global topological structure features and text annotation features of the multi-source drawing data based on a full convolutional network; determining terminal local pixel coordinates of each terminal based on the global topological structure features; constructing a logical mapping matrix based on the electrical schematic diagram and the assembly layout diagram, and determining an initial wiring task set according to the logical mapping matrix.

3. The multi-modal parsing and knowledge constraint based control cabinet wiring planning method according to claim 1, wherein, The process of determining terminal local pixel coordinates of each terminal based on the global topological structure features comprises: determining a terminal response heat map based on the global topological structure features and using a local key point positioning algorithm; determining terminal local pixel coordinates of each terminal based on the terminal response heat map.

4. The multi-modal parsing and knowledge constraint based control cabinet wiring planning method according to claim 1, wherein, The process of constructing a logical mapping matrix based on the electrical schematic diagram and the assembly layout diagram, and determining an initial wiring task set according to the logical mapping matrix comprises: extracting a logical network topology graph of the electrical schematic diagram, wherein the logical network topology graph comprises a logical terminal set determined based on the text annotation features and a logical connected edge set determined according to the logical terminal set; constructing physical space point sets from physical entity terminals identified in the assembly layout diagram based on the terminal local pixel coordinates; determining a comprehensive alignment confidence score based on the logical terminal set and the physical space point sets, and determining a logical mapping matrix according to the comprehensive alignment confidence score; translating the logical connected edge set into wiring tasks based on the logical mapping matrix, and determining an initial wiring task set according to the wiring tasks.

5. The multi-modal parsing and knowledge constraint based control cabinet wiring planning method according to claim 1, wherein, The steps of determining a process rule subgraph based on the initial wiring task set, obtaining a numerical constraint index by constraining the process rule subgraph, and constructing a dynamic process constraint feature matrix comprise: determining a process rule subgraph corresponding to a wiring task of the initial wiring task set; translating the process rule subgraph into a numerical constraint index based on a compilation mapping function; Acquire three-dimensional point cloud data inside the control cabinet, and construct a dynamic process constraint feature matrix based on the three-dimensional point cloud data.

6. The multi-modal parsing and knowledge constraint based control cabinet wiring planning method according to claim 5, characterized in that, The process of completing the control cabinet wiring planning includes: determining a structured process reasoning prompt word based on the initial wiring task set and the dynamic process constraint feature matrix; sequencing the initial wiring task set based on the structured process reasoning prompt word and the dynamic process constraint feature matrix to obtain a physical wiring sequence time sequence chain; solving the wiring task based on the physical wiring sequence time sequence chain and the wiring table, and outputting a digital process instruction list according to the numerical constraint index to complete the control cabinet wiring planning.

7. The multi-modal parsing and knowledge constraint based control cabinet wiring planning method according to any one of claims 1 to 6, characterized in that, It also includes adjusting the digital process instruction list to obtain an adjustment action; record the fine-grained features of the adjustment action using a motion capture operator, arrange and package the fine-grained features to obtain a process modification before and after difference data set; determine the process scheme containing the process modification before and after difference data set based on a bottom line checking algorithm, and obtain a final executable process scheme according to the determination result.

8. The multi-modal parsing and knowledge constraint based control cabinet wiring planning method according to claim 7, characterized in that, It also includes determining a process reflection abstract short text based on the digital process instruction list and the final executable process scheme; project the process reflection abstract short text to a preset high-dimensional semantic space using a preset industrial text embedding model to obtain an incremental experience feature vector; input the incremental experience feature vector into a long-term experience vector database, retrieve incremental experience feature vectors with a similarity score greater than a preset threshold in the long-term experience vector database, and update the long-term experience vector database.

9. A control cabinet wiring planning system based on multi-modal analysis and knowledge constraint, adopting the control cabinet wiring planning method based on multi-modal analysis and knowledge constraint in any one of claims 1-8, characterized in that, It includes: a multi-modal analysis module for acquiring multi-source drawing data of the control cabinet, extracting global topological structure features and text annotation features of the multi-source drawing data, determining terminal local pixel coordinates of each wiring terminal of the control cabinet based on the global topological structure features, and determining an initial wiring task set, wherein the multi-source drawing data includes an electrical schematic diagram, an assembly layout diagram, and a wiring table; a process knowledge management center connected to the multi-modal analysis module, configured to determine a process rule subgraph based on the initial wiring task set, constrain the process rule subgraph to obtain a numerical constraint index, and construct a dynamic process constraint feature matrix; a path planning brain connected to the multi-modal analysis module and the process knowledge management center, configured to determine a structured process reasoning prompt word based on the initial wiring task set and the dynamic process constraint feature matrix, determine a physical wiring sequence time sequence chain based on the structured process reasoning prompt word and the dynamic process constraint feature matrix, and determine a digital process instruction list based on the physical wiring sequence time sequence chain and the wiring table to complete the control cabinet wiring planning. a virtual rendering module connected with the path planning brain, configured to adjust the digital process instruction list to obtain an adjustment action, record fine-grained features of the adjustment action by using a motion capture operator, package and arrange the fine-grained features to obtain a process modification before and after difference data set, determine a process scheme containing the process modification before and after difference data set based on a bottom line verification algorithm, and obtain a final executable process scheme according to a determination result; a process reflection module connected with the virtual rendering module, configured to determine a process reflection abstract short text based on the digital process instruction list and the final executable process scheme, project the process reflection abstract short text to a preset high-dimensional semantic space by using a preset industrial text embedding model to obtain an incremental experience feature vector, input the incremental experience feature vector into a long-term experience vector database, and retrieve an incremental experience feature vector with a similarity score greater than a preset threshold in the long-term experience vector database to complete updating of the long-term experience vector database.

Images