A process flow arrangement system based on mechanical CAD drawing recognition

Abstract

The application relates to the technical field of intelligent manufacturing and computer-aided process planning, and discloses a process flow arrangement system based on mechanical CAD drawing recognition, which comprises CAD drawing analysis and feature recognition, manufacturing information structured storage, a process knowledge base, an intelligent process flow arrangement engine and a man-machine interaction module. The system adopts a double-channel parallel processing drawing of vector analysis and visual detection, performs feature space alignment and consistency checking, calculates feature recognition risk factors and generates a reconstructed feature vector; the intelligent arrangement engine dynamically adjusts processing method matching weights according to the risk factors, constructs a process sequence optimization model containing virtual risk cost, and adaptively supplements forced detection and allowance compensation constraints according to the risk level. The application realizes adaptive fault tolerance and dynamic optimization of a process scheme by converting recognition uncertainty into a quantitative risk index and integrating the quantitative risk index into process decision.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and computer-aided process planning technology, specifically a process flow arrangement system based on mechanical CAD drawing recognition. Background Technology

[0002] In the process of digital transformation in manufacturing, automated process design relies on the accurate extraction of manufacturing features from mechanical drawings. Traditional computer-aided process planning systems mostly use vector analysis methods based on geometric topology to obtain part information. Vector analysis methods require drawings to strictly conform to drafting standards and have a complete closed geometric topology. Faced with non-standard drawing, arbitrary labeling positions, and broken topological elements in historical drawings, a single vector analysis mode is difficult to accurately reconstruct manufacturing semantics. Although emerging deep learning-based visual inspection methods can recognize semantic symbols, they have limitations in the accuracy of geometric parameter extraction. Existing technologies lack a mechanism to fuse and verify precise vector geometric information with visual semantic information, making it impossible to quantitatively evaluate the reliability of feature recognition results.

[0003] In the process decision-making stage, existing systems typically employ static rule-based matching methods. These systems assume accurate feature recognition results and execute logical judgments based on deterministic parameters. When there are deviations in upstream feature recognition, the static matching logic fails to perceive the uncertainty of the input data and continues to output the process plan according to the predetermined rules. This rigid decision-making mechanism results in a lack of fault tolerance in process method selection, making it difficult to adapt to fluctuations in input data quality.

[0004] In terms of process sequencing and parameter optimization, traditional algorithm models primarily set minimizing processing time and cost as the objective function. The model construction process does not consider the potential execution risks arising from the uncertainty of feature recognition. Existing technologies cannot automatically adjust the process route structure based on the reliability of the identified data, and lack the ability to generate adaptive constraints for low-confidence features, such as forced detection, margin compensation, and process separation. In scenarios with fuzzy feature definitions, the generated process specifications often suffer from overcutting, positioning benchmark deviations, and high scrap rates, failing to meet the robustness requirements of intelligent manufacturing for process solutions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a process flow arrangement system based on mechanical CAD drawing recognition. It solves the problems of existing computer-aided process planning technologies, which heavily rely on the accuracy of a single feature extraction mode, lack a quantitative evaluation mechanism for data uncertainty caused by non-standard drawing annotations, and are unable to dynamically adjust processing strategies and compensation constraints based on the reliability of feature recognition. These problems result in low fault tolerance and high execution risks for process procedures generated from non-standard drawings.

[0006] To achieve the above objectives, the present invention provides an adaptive process flow arrangement system based on dual verification of graphic semantics and visual features. This system includes a CAD drawing parsing and feature recognition module, a manufacturing information structured storage module, a process knowledge base module, an intelligent process flow arrangement engine, and a human-computer interaction and editing output module. All of the above modules are connected and deployed on an application server via an internal bus and data interface.

[0007] In the adaptive process flow orchestration system based on dual verification of graphic semantics and visual features, the CAD drawing parsing and feature recognition module is configured to process the input mechanical CAD drawing files in parallel through a vector parsing channel and a visual inspection channel. The vector parsing channel parses the underlying data structure of the drawing, extracts geometric primitives, text annotations, and layer attributes, and constructs a vector feature set based on geometric topological relationships. The visual inspection channel converts the mechanical CAD drawing file into a raster image, uses a deep learning object detection model to identify manufacturing feature semantics in the image, and constructs a visual feature set. This module establishes a unified drawing world coordinate system, performs spatial mapping and alignment between the vector feature set and the visual feature set, and performs consistency verification on the aligned features.

[0008] This invention also provides a feature recognition risk quantification method using the aforementioned system. This method performs a compatibility comparison between the geometric semantics extracted by the vector analysis channel and the visual semantics recognized by the visual detection channel. Based on the consistency of the semantic comparison results and the confidence score of the visual detection, the system calculates a feature recognition risk factor. This feature recognition risk factor, as a dimensionless quantification index measuring the reliability of front-end feature extraction, is written into the feature data structure to generate a reconstructed feature vector containing geometric parameters, tolerance information, and the recognition risk factor. Thus, the system transforms the uncertainty in the unstructured drawing recognition process into structured data that can be processed by downstream algorithms.

[0009] Furthermore, this invention provides an adaptive process orchestration method based on risk cost, executed by an intelligent process orchestration engine. During the processing method selection phase, this engine dynamically adjusts the matching strategy based on risk factors identified through feature recognition. The system reads the risk factor values ​​from the reconstructed feature vector and uses these values ​​to correct the feature type weight coefficients and size weight coefficients in the processing method matching function. For high-risk features, the system automatically reduces the dependence weight on precise geometric dimension matching, prioritizing processing methods with higher fault tolerance, thereby achieving adaptive adjustment of process decisions to the uncertainty of input data.

[0010] During the process sequencing and optimization phase, the intelligent process orchestration engine constructs a global optimization model that incorporates virtual risk costs. The system combines feature identification of risk factors and error sensitivity parameters of processing methods to calculate nondeterministic virtual risk costs, which are then added to the basic processing costs of each process. By solving for the process sequence that minimizes the overall objective function, the system automatically avoids combinations of high-risk features and high-error-sensitive processing methods at the planning level, balancing theoretical processing costs with actual execution risks.

[0011] Furthermore, this invention also relates to a dynamic generation mechanism for process constraints. The system identifies risk factors based on their numerical values, classifies risks into different levels, and dynamically generates corresponding additional process constraints. For features whose identified risks exceed a preset threshold, the system automatically inserts mandatory online inspection procedures into the process route, adjusts the dimensional tolerances and allowance parameters of semi-finishing procedures to implement allowance compensation, or generates process separation constraints to prohibit parallel processing. The human-computer interaction and editing output module ultimately converts the process flow data containing the above adaptive adjustment strategies into visualized process charts and standard process documents, and highlights high-risk process nodes.

[0012] This invention provides a process flow arrangement system based on mechanical CAD drawing recognition. It has the following beneficial effects: 1. This invention performs parallel processing of mechanical drawings by configuring dual channels of vector analysis and visual inspection, establishes a unified coordinate system to achieve alignment and consistency verification between geometric features and visual semantic space, calculates feature identification risk factors based on semantic comparison status, and transforms unstructured drawing information into structured feature vectors containing risk dimensions. This solves the problem of low recognition accuracy of traditional single analysis mode when drawing annotation is not standardized, and provides a quantitative data reliability basis for subsequent process decisions.

[0013] 2. This invention introduces a dynamic weight adjustment mechanism based on feature recognition risk factors in the process matching stage. The weight ratio of feature type and size parameter in the processing method matching function is adjusted according to the risk value. When there is uncertainty in the front-end feature recognition, the dependence on precise geometric parameters is automatically reduced, and processing methods with high fault tolerance and strong versatility are prioritized. This realizes the adaptive compensation of the process decision logic for the quality of drawing input, and reduces the probability of process method matching errors caused by recognition deviation.

[0014] 3. This invention constructs a global optimization model for virtual risk costs and a dynamic constraint generation mechanism. It combines the error sensitivity of processing methods to transform identified risks into quantifiable cost items. It also automatically generates mandatory online detection, process separation, and margin compensation constraints for high-risk characteristics. At the process route planning level, it balances theoretical processing costs with actual execution risks, ensuring that the final generated process plan is feasible and safe even when the input data is not ideal. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the architecture and software module functions of the adaptive process flow arrangement system based on dual verification of graphic semantics and visual features of the present invention. Figure 2 This is a schematic diagram illustrating the geometric analysis based on vector data, feature detection based on visual semantics, and feature space mapping and alignment processes of the present invention. Figure 3 This is a schematic diagram illustrating the process of feature consistency determination, risk factor identification calculation, and enhanced feature vector construction in this invention. Figure 4 This is a schematic diagram illustrating the dynamic weighting process matching, dynamic objective function construction based on uncertainty cost, and dynamic constraint supplementation process of the present invention. Figure 5 This is a schematic diagram of the data input and transmission shaft component processing flow of a specific embodiment of the present invention; Figure 6 This is a schematic diagram of the final process route output interface of the present invention.

[0016] Among them, 100 is the adaptive process flow arrangement system; 101 is the CAD drawing parsing and feature recognition module; 102 is the manufacturing information structured storage module; 103 is the process knowledge base module; 104 is the intelligent process flow arrangement engine; 105 is the human-computer interaction and editing output module; 106 is the application server; 1011 is the vector data parsing unit; 1012 is the visual feature detection unit; 1013 is the dual-channel consistency verification unit; 1041 is the adaptive process matching unit; and 1042 is the dynamic programming solution unit. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See attached document Figure 1This invention provides an adaptive process flow arrangement system 100 based on dual verification of graphic semantics and visual features. The adaptive process flow arrangement system 100 includes a CAD drawing parsing and feature recognition module 101, a manufacturing information structured storage module 102, a process knowledge base module 103, an intelligent process flow arrangement engine 104, and a human-computer interaction and editing output module 105. The CAD drawing parsing and feature recognition module 101, the manufacturing information structured storage module 102, the process knowledge base module 103, the intelligent process flow arrangement engine 104, and the human-computer interaction and editing output module 105 are connected via an internal bus and data interface and deployed on an application server 106. The application server 106 provides data computing and processing capabilities and data storage space.

[0019] The CAD drawing parsing and feature recognition module 101 is used to import and parse the mechanical CAD drawing files of the target part. Internally, the CAD drawing parsing and feature recognition module 101 includes a vector parsing channel and a visual inspection channel. The vector parsing channel parses the underlying data structure of the mechanical CAD drawing file, extracts geometric primitives, text annotations, and layer attributes, and constructs a vector feature set. The visual inspection channel converts the mechanical CAD drawing file into a raster image, uses a target detection algorithm to identify manufacturing features, and constructs a visual feature set. The CAD drawing parsing and feature recognition module 101 is configured to perform spatial alignment and consistency checks between the vector feature set and the visual feature set, calculate feature recognition risk factors, and output reconstructed feature vectors.

[0020] The manufacturing information structured storage module 102 is connected to the CAD drawing parsing and feature recognition module 101. The manufacturing information structured storage module 102 receives and stores the reconstructed feature vectors output by the CAD drawing parsing and feature recognition module 101, forming a part manufacturing information model. The part manufacturing information model includes a feature list, tolerance information, surface quality requirements, and material information. The feature list includes feature type, geometric parameters, and location information.

[0021] The process knowledge base module 103 stores process flow arrangement rules and data. It includes machining method and feature mapping rules, a machining capability database, process decision logic, and a standard process parameter library. The machining method and feature mapping rules define the correspondence between feature types and machining methods. The machining capability database records machine tool performance parameters. The process decision logic includes process sequencing principles. The standard process parameter library contains cutting speed and feed rate parameters for different combinations of materials, tools, and machining methods.

[0022] The intelligent process orchestration engine 104 is connected to the manufacturing information structured storage module 102 and the process knowledge base module 103. The intelligent process orchestration engine 104 is used to call upon the part manufacturing information model and the data from the process knowledge base module 103 to execute automated process orchestration. The intelligent process orchestration engine 104 is configured to dynamically adjust the process matching strategy and the process sequencing optimization model based on feature-based risk factor identification. During the processing method selection stage, the intelligent process orchestration engine 104 constructs a feature vector for each manufacturing feature: ; In the formula: For feature vectors; For manufacturing feature types; These are geometric dimension parameters; This is tolerance information; Surface roughness.

[0023] The intelligent process orchestration engine 104 provides each candidate processing method with Constructing capability vectors. The intelligent process orchestration engine 104 quantifies and selects the optimal processing method by calculating a matching degree function: ; In the formula: Features With candidate processing methods Match score; These are the feature type weight coefficients; It is an indicator function; Candidate processing methods Capable of handling feature type sets; This is a size weighting coefficient; This is the size fitness function. The intelligent process flow orchestration engine 104 is configured to dynamically adjust the weight coefficients based on the feature recognition risk factors output by the CAD drawing parsing and feature recognition module 101. and Numerical value.

[0024] During the process sequencing stage, the intelligent process orchestration engine 104 models the process sequencing as an optimization problem and uses an objective function to calculate the minimum total cost and time. ; In the formula: The total objective function value; For process Processing time; Process Index Clamping time; and For constraint term indexing. The intelligent process orchestration engine 104 is configured to calculate the objective function. During the process, risk factors are identified based on characteristics, and virtual risk costs are added to processes involving high-risk characteristics.

[0025] The human-computer interaction and editing output module 105 is connected to the intelligent process flow arrangement engine 104. The human-computer interaction and editing output module 105 receives the initial process flow data generated by the intelligent process flow arrangement engine 104 and presents it visually. The human-computer interaction and editing output module 105 is equipped with an editing interface to respond to user operations such as reviewing, modifying, adjusting, and optimizing the initial process flow. The human-computer interaction and editing output module 105 is used to output the finalized process flow as a standard format process file.

[0026] This invention provides an adaptive process flow arrangement system software architecture, which includes a CAD drawing parsing and feature recognition module 101, a manufacturing information structured storage module 102, a process knowledge base module 103, an intelligent process flow arrangement engine 104, and a human-computer interaction and editing output module 105. The modules interact with each other through defined data structures and application programming interfaces.

[0027] The CAD drawing parsing and feature recognition module 101 internally includes a vector data parsing unit 1011, a visual feature detection unit 1012, and a dual-channel consistency verification unit 1013. The vector data parsing unit 1011 is configured to read the binary data stream of the mechanical CAD drawing file. The vector data parsing unit 1011 parses the entity segment data within the file, extracting geometric primitives such as points, lines, circles, arcs, and spline curves. The vector data parsing unit 1011 extracts associated text annotation objects, dimension annotation objects, and layer attribute information. The vector data parsing unit 1011 constructs a vector feature set based on the topological relationships of the geometric primitives.

[0028] The visual feature detection unit 1012 is configured to perform a graphics rendering rasterization operation, converting mechanical CAD drawing files into high-resolution bitmap images. The visual feature detection unit 1012 loads a pre-trained deep learning object detection model. The visual feature detection unit 1012 performs convolution operations, feature map extraction, and non-maximum suppression operations on the bitmap image. The visual feature detection unit 1012 outputs a set of visual features containing feature category labels, confidence scores, and bounding box coordinate data.

[0029] The dual-channel consistency verification unit 1013 is configured to establish a unified drawing world coordinate system. The dual-channel consistency verification unit 1013 performs spatial overlap mapping between the geometric coordinate data in the vector feature set and the bounding box coordinate data in the visual feature set. The dual-channel consistency verification unit 1013 compares the vector parsing semantics with the visual detection semantics within the mapped region. The dual-channel consistency verification unit 1013 calculates the feature recognition risk factor based on the semantic comparison result. The dual-channel consistency verification unit 1013 writes the feature recognition risk factor as an independent dimension into the feature data structure and outputs a reconstructed feature vector containing the feature recognition risk factor.

[0030] The manufacturing information structured storage module 102 is configured to receive reconstructed feature vectors. The manufacturing information structured storage module 102 maps the reconstructed feature vectors to object-oriented entity class data structures. The manufacturing information structured storage module 102 establishes the association relationships between feature entities, tolerance entities, and technical requirement entities, generating a part manufacturing information model.

[0031] The intelligent process orchestration engine 104 internally includes an adaptive process matching unit 1041 and a dynamic programming solution unit 1042. The adaptive process matching unit 1041 is configured to read reconstructed feature vectors from the part manufacturing information model. It retrieves processing methods and feature mapping rules from the process knowledge base module 103. The adaptive process matching unit 1041 reads the feature identification risk factor values ​​from the reconstructed feature vectors. Based on the feature identification risk factor values, the adaptive process matching unit 1041 dynamically sets the feature type weight coefficient and size weight coefficient in the matching degree function. The adaptive process matching unit 1041 calculates the matching degree score between features and candidate processing methods, generating a set of candidate processing methods.

[0032] The dynamic programming solver unit 1042 is configured to receive a set of candidate processing methods. The dynamic programming solver unit 1042 constructs a mathematical model for process sequencing optimization. Based on the identified risk factors and processing method error sensitivity parameters, the dynamic programming solver unit 1042 calculates the virtual risk cost value. The dynamic programming solver unit 1042 adds the virtual risk cost value to the process cost term in the objective function. The dynamic programming solver unit 1042 executes a global optimization algorithm to find the process sequence and resource allocation scheme that minimizes the total value of the objective function, generating initial process flow data.

[0033] The human-computer interaction and editing output module 105 is configured to parse the initial process flow data. The module renders the process flow diagram on the user interface. It detects process nodes in the initial process flow data that are associated with high-risk identification factors. It highlights and renders these process nodes. The module receives editing commands from the user via an input device to update the process flow data. Finally, it calls a file generation interface to serialize the final process flow data into a standard process document format.

[0034] See attached document Figure 2 This invention provides a geometric analysis method based on vector data, which is executed by a vector data analysis unit 1011. The vector data analysis unit 1011 receives a mechanical CAD drawing file as an input data stream. The vector data analysis unit 1011 reads the header data of the mechanical CAD drawing file, parses the file version number and character encoding format. The vector data analysis unit 1011 locates the entity data segment and traverses all graphic object data blocks stored in the entity data segment.

[0035] The vector data parsing unit 1011 performs type identification and classification extraction on the graphic object data blocks. For graphic objects identified as straight lines, the vector data parsing unit 1011 extracts the starting point coordinate data and ending point coordinate data. For graphic objects identified as circles, the vector data parsing unit 1011 extracts the center coordinate data and radius data. For graphic objects identified as arcs, the vector data parsing unit 1011 extracts the center coordinate data, radius data, starting angle data, and ending angle data. For graphic objects identified as polylines, the vector data parsing unit 1011 extracts the vertex coordinate sequence data and convexity parameter data.

[0036] The vector data parsing unit 1011 performs layer attribute filtering. The vector data parsing unit 1011 reads the layer index attribute of each graphic object. According to a preset layer naming rule table, the vector data parsing unit 1011 divides the graphic objects into a set of contour geometric objects, a set of dimension objects, and a set of text annotation objects. The set of contour geometric objects contains geometric primitives describing the physical shape of the part. The set of dimension objects contains linear dimensioning, diameter dimensioning, and angle dimensioning primitives. The set of text annotation objects contains technical requirements and tolerance numerical text.

[0037] The vector data parsing unit 1011 performs a topology reconstruction operation on the set of contour geometric objects. The vector data parsing unit 1011 calculates the Euclidean distance between the coordinates of the endpoints of each geometric primitive. When the Euclidean distance between two endpoints is less than a preset overlap threshold, the vector data parsing unit 1011 determines that the two endpoints are connected. The vector data parsing unit 1011 traces the geometric primitive links based on the connection relationship, identifies and constructs closed geometric loops. The vector data parsing unit 1011 marks the closed geometric loops as candidate processing feature contours.

[0038] The vector data parsing unit 1011 performs the association mapping operation between geometric features and non-geometric information. The vector data parsing unit 1011 calculates the bounding box center coordinates of each text object in the text annotation object set. The vector data parsing unit 1011 calculates the geometric center coordinates of the candidate processing feature contour. The vector data parsing unit 1011 calculates the distance vector between the bounding box center coordinates of the text object and the geometric center coordinates of the candidate processing feature contour. The vector data parsing unit 1011 associates the text object with the candidate processing feature contour with the smallest distance vector magnitude.

[0039] The vector data parsing unit 1011 constructs a vector feature set based on the association mapping results. Each element in the vector feature set contains the geometric shape definition data of the feature, the associated dimensional numerical data, and the extracted tolerance text data. The vector data parsing unit 1011 stores the vector feature set in a memory buffer for subsequent use by the dual-channel consistency verification unit 1013.

[0040] Meanwhile, the visual semantic-based feature detection method is executed by the visual feature detection unit 1012. The visual feature detection unit 1012 receives mechanical CAD drawing files as input data. The visual feature detection unit 1012 calls the graphics rendering engine interface to set the rasterization sampling resolution parameters. The visual feature detection unit 1012 maps the vector data stream in the mechanical CAD drawing file into a pixel matrix, generating a high-resolution raster bitmap image. The rasterization sampling resolution parameters are set to values ​​that ensure the smallest annotation characters and surface texture features in the drawing are clearly distinguishable in the raster bitmap image.

[0041] The visual feature detection unit 1012 performs image preprocessing operations on the raster bitmap image. The image preprocessing operations include grayscale conversion, histogram equalization, and noise suppression filtering. The visual feature detection unit 1012 inputs the preprocessed raster bitmap image into a preset deep learning object detection model. The deep learning object detection model includes a convolutional neural network feature extraction backbone network, a feature pyramid fusion network, and a detection head network.

[0042] The visual feature detection unit 1012 utilizes a convolutional neural network feature extraction backbone network to perform convolution and pooling operations on the raster bitmap image, generating a multi-scale feature map set. The visual feature detection unit 1012 then uses a feature pyramid fusion network to perform upsampling and feature concatenation operations on the multi-scale feature map set, fusing deep semantic features with shallow detail features. Finally, the visual feature detection unit 1012 uses a detection head network to perform anchor box regression prediction and class probability classification prediction on the fused feature map.

[0043] The visual feature detection unit 1012 is configured to identify manufacturing features with specific visual semantics in a raster bitmap image. These manufacturing features include threaded texture regions, surface roughness symbols, geometric tolerance frames, datum symbols, and knurled texture regions. The detection head network output contains a data list of multiple candidate detection boxes. Each candidate detection box includes a predicted category label, a confidence score, and bounding box pixel coordinates. The bounding box pixel coordinates consist of the top-left x-coordinate, top-left y-coordinate, bottom-right x-coordinate, and bottom-right y-coordinate.

[0044] The visual feature detection unit 1012 performs non-maximum suppression on the candidate detection box data list. The visual feature detection unit 1012 calculates the intersection-union ratio (CUI) between overlapping candidate detection boxes. When the CUI value is greater than a preset overlap threshold, the visual feature detection unit 1012 retains the candidate detection box with the highest confidence value and discards the remaining overlapping candidate detection boxes. The visual feature detection unit 1012 filters out candidate detection boxes with confidence values ​​lower than a preset effective threshold.

[0045] The visual feature detection unit 1012 encapsulates the detection results after non-maximum suppression and threshold filtering operations into a visual feature set. Each element in the visual feature set contains a semantic category identifier, a visual recognition confidence score, and bounding box data in the raster bitmap image coordinate system. The visual feature detection unit 1012 transmits the visual feature set to the dual-channel consistency verification unit 1013.

[0046] Subsequently, feature space mapping and alignment operations are performed. The dual-channel consistency verification unit 1013 receives vector feature sets from the vector data parsing unit 1011 and visual feature sets from the visual feature detection unit 1012. The dual-channel consistency verification unit 1013 reads the drawing boundary coordinate data from the vector feature sets, including the coordinates of the lower left and upper right corners of the drawing area. The dual-channel consistency verification unit 1013 also reads the raster bitmap image resolution data corresponding to the visual feature sets, including the image width (in pixels) and image height (in pixels).

[0047] The dual-channel consistency verification unit 1013 establishes a linear mapping relationship between the vector world coordinate system and the visual pixel coordinate system. The dual-channel consistency verification unit 1013 calculates the horizontal and vertical scaling factors of the vector coordinate data relative to the pixel coordinate data of the raster bitmap image. The dual-channel consistency verification unit 1013 calculates the translational offset of the vector coordinate origin relative to the origin of the raster bitmap image. The dual-channel consistency verification unit 1013 constructs a coordinate transformation matrix using the horizontal scaling factor, the vertical scaling factor, and the translational offset.

[0048] The dual-channel consistency verification unit 1013 performs an inverse projection transformation on the bounding box pixel coordinate data of each element in the visual feature set using a coordinate transformation matrix. The dual-channel consistency verification unit 1013 converts the bounding box coordinates of all features in the visual feature set into vector world coordinates that are coherent with the vector feature set. The dual-channel consistency verification unit 1013 generates a visual feature mapping set in a unified coordinate system.

[0049] The dual-channel consistency verification unit 1013 performs spatial grid partitioning on the vector feature set and constructs a spatial index tree structure. The dual-channel consistency verification unit 1013 traverses each visual feature element in the visual feature mapping set. The dual-channel consistency verification unit 1013 uses the spatial index tree to retrieve candidate vector features in the vector feature set that overlap in spatial location with the given visual feature element.

[0050] The dual-channel consistency verification unit 1013 calculates the intersection-union ratio (IUR) of the bounding boxes between elements in the visual feature mapping set and candidate vector features. The dual-channel consistency verification unit 1013 sets a spatial association determination threshold. When the calculated IUR value is greater than the spatial association determination threshold, the dual-channel consistency verification unit 1013 determines that the visual feature and the vector feature describe the same physical object. The dual-channel consistency verification unit 1013 establishes a feature alignment association index, linking the semantic label and confidence data of the visual feature to the corresponding vector feature data structure.

[0051] The dual-channel consistency verification unit 1013 handles multiple mapping scenarios. When a vector feature spatially overlaps with multiple visual features, the dual-channel consistency verification unit 1013 selects the visual feature with the highest intersection-union ratio (IU) to establish an association. When a vector feature does not find any overlapping visual features, the dual-channel consistency verification unit 1013 marks the associated visual semantic field of that vector feature as null and sets the visual channel confidence score to zero. The dual-channel consistency verification unit 1013 outputs a list of spatially aligned paired features for subsequent risk quantification calculations.

[0052] See attached document Figure 3This invention provides a feature consistency determination method, which is executed by a dual-channel consistency verification unit 1013. The dual-channel consistency verification unit 1013 reads a paired feature list generated through spatial alignment operations. Each item in the paired feature list contains geometric feature attribute data from the vector resolution channel and visual feature attribute data from the visual detection channel. The geometric feature attribute data includes primitive type, closed contour shape, and associated text annotation content. The visual feature attribute data includes the predicted category label of the detection box and the visual confidence score.

[0053] The dual-channel consistency verification unit 1013 performs a semantic normalization mapping operation. The dual-channel consistency verification unit 1013 calls a preset manufacturing semantic mapping table. The manufacturing semantic mapping table defines the correspondence between two-dimensional geometric primitive types and machining feature types. Based on the geometric primitive type and closed contour shape, the dual-channel consistency verification unit 1013 converts the vector feature attribute data into a first semantic label. The dual-channel consistency verification unit 1013 directly reads the predicted category label from the visual feature attribute data as the second semantic label.

[0054] The dual-channel consistency verification unit 1013 performs a semantic compatibility comparison operation. The dual-channel consistency verification unit 1013 constructs a semantic compatibility matrix. The row index of the semantic compatibility matrix corresponds to the first semantic label, the column index corresponds to the second semantic label, and the matrix element values ​​are Boolean compatibility identifiers. When the first semantic label is a circular closed contour and the second semantic label is a threaded hole texture, the corresponding matrix element value is true. When the first semantic label is a rectangular closed contour and the second semantic label is a countersunk hole symbol, the corresponding matrix element value is false.

[0055] The dual-channel consistency verification unit 1013 determines the feature consistency status based on the semantic compatibility comparison results and the visual confidence score. The dual-channel consistency verification unit 1013 defines three feature consistency statuses: complete consistency, potential conflict, and missing feature.

[0056] When the corresponding element values ​​of the first semantic label and the second semantic label in the semantic compatibility matrix are true, and the visual confidence value is greater than the preset high confidence threshold, the dual-channel consistency verification unit 1013 determines that the feature is in a completely consistent state. In this state, the system recognizes that the vector parsing result and the visual detection result corroborate each other, and the feature definition is accurate.

[0057] When the corresponding element values ​​of the first and second semantic labels in the semantic compatibility matrix are false, and the visual confidence value is greater than the preset effective detection threshold, the dual-channel consistency verification unit 1013 determines that the feature is in a potential conflict state. In this state, the system determines that there is a logical contradiction between the geometric drawing information and the visual annotation symbol information of the drawing, indicating that there is an annotation error or version inconsistency problem in the drawing.

[0058] When the paired feature list contains only vector feature attribute data, while the corresponding visual feature attribute data is empty or the visual confidence value is lower than the preset effective detection threshold, the dual-channel consistency verification unit 1013 determines that the feature is in a feature missing state. In this state, the system determines that the feature lacks visual semantic support and is at risk of being omitted or drawn in a non-standard manner. The dual-channel consistency verification unit 1013 writes the determined consistency status identifier into the feature data structure as the input basis for subsequent calculation of feature recognition risk factors.

[0059] This invention provides a method for calculating risk factors, which is executed by a dual-channel consistency verification unit 1013. Based on the feature consistency status and visual confidence score output by the feature consistency judgment logic, the dual-channel consistency verification unit 1013 performs risk quantification processing on each feature to be processed. It is a dimensionless scalar whose range of values ​​is a closed interval. Identify risk factors The higher the value, the more unreliable the feature is in subsequent process decisions.

[0060] The dual-channel consistency verification unit 1013 executes the first calculation logic for features in a completely consistent state. The dual-channel consistency verification unit 1013 reads the visual confidence value corresponding to that feature. Since the vector semantics and visual semantics corroborate each other at this point, risk factors are identified. The main limitation is the confidence probability of the visual detection model. The dual-channel consistency verification unit 1013 will identify risk factors. The value is set to be negatively correlated with the visual confidence score. Risk factors are identified when the visual confidence score approaches 1. Approaching 0.

[0061] The dual-channel consistency verification unit 1013 executes a second calculation logic for features in a potentially conflicting state. Since the geometry defined by vector analysis and the texture symbols recognized by visual inspection have semantic contradictions, this indicates a high degree of uncertainty in the drawing. The dual-channel consistency verification unit 1013 will identify risk factors. A preset conflict penalty constant is set. This conflict penalty constant is set to a high-risk value close to 1. The dual-channel consistency verification unit 1013 fine-tunes this value according to the severity of the conflict. The severity is determined based on the conflict level defined in the semantic compatibility matrix.

[0062] The dual-channel consistency verification unit 1013 executes a third calculation logic for features in a feature-missing state. Since only vector geometric information exists at this time without visual semantic support, the dual-channel consistency verification unit 1013 will identify risk factors. A preset blind zone risk constant is set. This blind zone risk constant is set to a medium-to-high risk value between 0.5 and 0.8. The dual-channel consistency verification unit 1013 performs a weighted adjustment of this constant based on the geometric size of the feature. For features with geometric sizes smaller than a preset small feature threshold, the dual-channel consistency verification unit 1013 increases their corresponding identification risk factor. Numerical value.

[0063] The dual-channel consistency verification unit 1013 completes the identification of risk factors. After calculation, a feature vector reconstruction operation is performed. The dual-channel consistency verification unit 1013 combines the original geometric attribute data and tolerance attribute data with the calculated risk identification factors. The features are fused to construct a high-dimensional feature vector that includes the risk dimension. The dual-channel consistency verification unit 1013 outputs the reconstructed feature vector. Its mathematical expression is as follows: ; In the formula: The reconstructed feature vector; To create the feature type, this type is determined by the geometric primitive type extracted from the vector analysis channel; These are geometric dimensional parameters, including diameter, depth, and length values; The tolerance information is characterized by dimensional tolerances and geometric tolerances. The required surface roughness value is a characteristic feature. The calculated risk factors are used to identify the risk factors.

[0064] The dual-channel consistency check unit 1013 will generate the feature vector The data is serialized and transmitted to the manufacturing information structured storage module 102 for persistent storage, and simultaneously transmitted to the intelligent process flow orchestration engine 104 as input variables for subsequent adaptive process matching. This is achieved by identifying risk factors. Explicitly included in the feature vector In this process, the system ensures that subsequent process decision-making algorithms can perceive the uncertainties in front-end feature recognition.

[0065] Furthermore, enhanced feature vector construction is performed. This invention provides an enhanced feature vector construction method, which is executed by a dual-channel consistency verification unit 1013. The dual-channel consistency verification unit 1013 is configured to receive feature attribute data after risk quantification calculation and encapsulate it into a standardized mathematical vector format.

[0066] Dual-channel consistency verification unit 1013 for feature type data The digital encoding operation is performed. The dual-channel conformance verification unit 1013 calls the preset feature classification encoding table. The feature classification encoding table contains the mapping relationship between manufacturing feature categories and unique integer identifiers defined by the international standard STEP format. The dual-channel conformance verification unit 1013 converts the manufacturing semantic type in text form into the corresponding integer identifier. When the manufacturing semantic type is a general thread blind hole, the dual-channel conformance verification unit 1013 assigns it the corresponding specific integer code.

[0067] Dual-channel consistency verification unit 1013 checks geometric parameter data The system performs unit normalization and parameter completion operations. The dual-channel consistency verification unit 1013 checks the unit of measurement attributes of the geometric data. It converts all imperial unit values ​​to metric millimeter unit values. The dual-channel consistency verification unit 1013 constructs a geometric parameter vector. For features of revolution, the dual-channel consistency verification unit 1013 extracts the diameter and depth values ​​as vector elements. For features of non-revolution, the dual-channel consistency verification unit 1013 extracts the length, width, and height values ​​as vector elements.

[0068] Dual-channel consistency verification unit 1013 checks tolerance information data Standardized assignment operations are performed. The dual-channel consistency verification unit 1013 checks whether the feature has explicitly marked tolerance values. When an explicit tolerance is detected, the dual-channel consistency verification unit 1013 directly extracts the upper and lower deviation values. When an unmarked tolerance is detected, the dual-channel consistency verification unit 1013 queries the pre-stored standard tolerance grade table based on the feature's nominal geometric dimensions. The dual-channel consistency verification unit 1013 obtains the limit deviation values ​​for the corresponding accuracy grade and fills them into the tolerance information data field. The dual-channel consistency verification unit 1013 calculates the tolerance band width value and uses it as a quantitative indicator to measure machining accuracy.

[0069] The dual-channel consistency verification unit 1013 will calculate the identified risk factors. The feature vector data structure is embedded as an independent dimension. The dual-channel consistency verification unit 1013 constructs a five-dimensional data tuple containing feature type encoding, normalized geometric parameters, standardized tolerance values, surface roughness values, and risk identification factors.

[0070] The dual-channel consistency check unit 1013 serializes the five-dimensional data tuple into an enhanced feature vector. The dual-channel consistency check unit 1013 is an enhanced feature vector. Assign a globally unique feature index ID. The dual-channel consistency check unit 1013 will enhance the feature vector... Data is transmitted via an internal data bus to the manufacturing information structured storage module 102, serving as the underlying data node for constructing the part manufacturing information model. The dual-channel consistency verification unit 1013 simultaneously transmits the enhanced feature vector... The data is transmitted to the intelligent process flow orchestration engine 104 for subsequent algorithm calls. By constructing enhanced feature vectors, the system transforms unstructured drawing information and implicit risk identification information into structured numerical matrices that can be directly processed by computer algorithms.

[0071] See attached document Figure 4 This invention provides an adaptive process orchestration method based on risk cost, which is executed by an adaptive process matching unit 1041. The adaptive process matching unit 1041 receives an enhanced feature vector from a dual-channel consistency verification unit 1013. Adaptive process matching unit 1041 analyzes enhanced feature vectors. Extract the manufacturing semantic type encoding. Geometric parameter data and identification of risk factors .

[0072] The adaptive process matching unit 1041 establishes a communication connection with the process knowledge base module 103. The adaptive process matching unit 1041 encodes according to the manufacturing semantic type. A query command is sent to the process knowledge base module 103. The process knowledge base module 103 returns a set of candidate processing methods. Each candidate processing method in the set... It includes processing capability attributes, a list of applicable feature types, and standard process parameters.

[0073] The adaptive process matching unit 1041 is configured to execute a risk factor-based dynamic weight adjustment strategy. The adaptive process matching unit 1041 reads a preset static weight configuration table. The static weight configuration table defines the feature type matching weight baseline value and the size adaptability weight baseline value. The adaptive process matching unit 1041 utilizes identified risk factors... The static weight baseline value is adjusted, and the dynamic weight coefficient is calculated.

[0074] The adaptive process matching unit 1041 uses a linear attenuation model to calculate the dynamic weighting coefficients. ; In the formula: The calculated dynamic weighting coefficients; The baseline weight value read from the static weight configuration table; This is a preset risk sensitivity coefficient, and its value range is a real number greater than 0; The calculated risk factors are used to identify the risk factors.

[0075] The adaptive process matching unit 1041 sets risk sensitivity coefficients according to the geometric attribute categories of manufacturing features. The numerical value. For feature type matching weights that are sensitive to geometric accuracy, the adaptive process matching unit 1041 sets a higher risk sensitivity coefficient. When identifying risk factors As the weight increases, the value of the feature type matching weight decreases significantly. This logic is used to reduce the system's dependence on specific processing methods when feature recognition is uncertain.

[0076] The adaptive process matching unit 1041 utilizes the calculated dynamic weighting coefficients. Substituting the matching degree function defined in the aforementioned embodiments In the middle. The adaptive process matching unit 1041 will dynamically weight the coefficients. Assigned to the feature type weight coefficients in the matching function and size weighting coefficient .

[0077] The adaptive process matching unit 1041 traverses each candidate in the candidate processing method set. The adaptive process matching unit 1041 calculates the enhanced feature vector. With each candidate processing method The weighted matching score is calculated. The adaptive process matching unit 1041 stores the calculation results in the matching score list.

[0078] The adaptive process matching unit 1041 performs a candidate set filtering operation based on a risk threshold. The adaptive process matching unit 1041 sets a minimum availability threshold. The adaptive process matching unit 1041 removes processing methods from the matching score list whose scores are lower than the minimum availability threshold. The adaptive process matching unit 1041 sorts the remaining processing methods in descending order of matching score to generate a preferred sequence of available process methods.

[0079] The adaptive process matching unit 1041 transmits the optimal sequence of available process methods to the dynamic programming solution unit 1042. By introducing a dynamic weight calculation mechanism, the adaptive process matching unit 1041 automatically reduces the weight dependence on precise geometric matching when the feature recognition risk is high, thereby prioritizing the selection of processing methods with stronger versatility and higher fault tolerance, and realizing adaptive compensation for the uncertainty of front-end recognition in process decision-making.

[0080] Subsequently, a dynamic objective function based on uncertainty costs is constructed. The dynamic programming solver 1042 receives the optimal sequence of available process methods output by the adaptive process matching unit 1041. The dynamic programming solver 1042 also receives data including information on identified risk factors. The enhanced feature vector sequence.

[0081] The dynamic programming solver unit 1042 establishes a global optimization mathematical model for process route planning. The dynamic programming solver unit 1042 defines the global optimization objective function as minimizing the generalized processing cost. The generalized processing cost consists of two parts: deterministic basic processing cost and non-deterministic virtual risk cost. The dynamic programming solver unit 1042 traverses each feature node in the enhanced feature vector sequence and calculates the generalized processing cost value of that feature when using a specific processing method.

[0082] The dynamic programming solver 1042 calculates the deterministic basic processing cost. The dynamic programming solver 1042 reads the standard time quota data and equipment unit-time depreciation rate data for candidate processing methods from the process knowledge base module 103. The dynamic programming solver 1042 multiplies the standard time quota data by the equipment unit-time depreciation rate data to obtain the deterministic basic processing cost for that process.

[0083] The dynamic programming solver 1042 calculates the nondeterministic virtual risk cost. The dynamic programming solver 1042 reads the error sensitivity coefficient of the candidate machining method. The error sensitivity coefficient is a dimensionless parameter pre-stored in the process knowledge base module 103, used to characterize the machining method's tolerance to errors in the definition of input features. For grinding machining methods requiring high-precision positioning references, the error sensitivity coefficient is set to a value greater than 1; for rough milling methods, the error sensitivity coefficient is set to a value less than 1.

[0084] Dynamic programming solution unit 1042 combines risk factor identification in enhanced eigenvectors Based on the error sensitivity coefficient of the processing method, a generalized cost calculation formula for a single process is constructed: ; In the formula: This refers to the generalized processing cost of a single process. This refers to the deterministic basic processing cost calculated based on the time quota; The risk factors are identified by features, and their values ​​range from [value range missing]. ; This is the error sensitivity coefficient of the processing method; The preset risk penalty factor is used to adjust the weight ratio of virtual risk costs in the total cost.

[0085] Dynamic programming solver 1042 constructs a dynamic objective function for overall process optimization based on the generalized cost calculation formula for a single process. : ; In the formula: The total generalized cost target value for the process route; This represents the total number of processes included in the technological route. For the first The broad processing cost of each step; For from the first The process flow is transferred to the next step. The logistics and switching costs generated by each process.

[0086] The dynamic programming solver unit 1042 executes the dynamic programming algorithm. The dynamic programming solver unit 1042 models the process routing problem as a multi-stage decision-making process. The dynamic programming solver unit 1042 uses process nodes as stage variables and processing method selection as state variables. The dynamic programming solver unit 1042 uses recursive formulas to calculate the cumulative minimum generalized cost of each state node.

[0087] The dynamic programming solver 1042 implements a risk avoidance strategy during the solution process. This involves identifying risk factors related to features. When the value is high, according to the generalized cost calculation formula for a single process, this feature adopts a high error sensitivity coefficient. The processing method generates a huge increase in virtual risk cost, thereby increasing the objective function. The total value. During the optimization process, the dynamic programming solver 1042 will automatically tend to select the error sensitivity coefficient. Lower processing methods, or choosing processing methods that include online measurement and calibration steps, can reduce the total generalized cost.

[0088] The output of the dynamic programming solver 1042 makes the objective function... The optimal process sequence and corresponding resource allocation scheme are minimized numerically. The dynamic programming solution unit 1042 generates the optimal process sequence as initial process flow data and transmits it to the human-computer interaction and editing output module 105. This is achieved by introducing a risk factor-based approach. And error sensitivity coefficient The system quantifies the uncertainty of front-end feature recognition at the mathematical model level, ensuring that the generated process route achieves the optimal balance between theoretical cost and execution risk.

[0089] Simultaneously, the constraints are dynamically supplemented. After constructing the dynamic objective function including virtual risk costs, the dynamic programming solver 1042 further performs an adaptive constraint generation operation. The dynamic programming solver 1042 reads the basic constraint set stored in the process knowledge base module 103. The basic constraint set includes machine tool travel range constraints, tool reachability constraints, and clamping stability constraints.

[0090] The dynamic programming solution unit 1042 traverses the enhanced eigenvector sequence and identifies risk factors. The numerical values ​​are used to dynamically expand the basic constraint set. The dynamic programming solution unit 1042 sets hierarchical risk response thresholds, including mandatory inspection threshold, margin compensation threshold, and process separation threshold.

[0091] The dynamic programming solver 1042 executes the logic for generating mandatory constraint checks. The dynamic programming solver 1042 compares the identification risk factors for each feature. Compared to the preset mandatory test threshold. When identifying risk factors. When the value exceeds the mandatory verification threshold, the dynamic programming solver 1042 generates a mandatory verification constraint. This constraint stipulates that an online measurement operation must immediately follow the cutting operation involving this feature. The dynamic programming solver 1042 restricts this measurement operation to call the machine tool probe or external measuring equipment to obtain the actual geometric parameters after machining.

[0092] The dynamic programming solver unit 1042 executes the residual compensation constraint generation logic. The dynamic programming solver unit 1042 compares the identification risk factors for each feature. Compared to the preset margin compensation threshold. When risk factors are identified. When the tolerance exceeds the margin compensation threshold, the dynamic programming solver 1042 modifies the dimensional tolerance constraints of the semi-finishing process. The dynamic programming solver 1042 calculates the safety margin increment, which is then correlated with the identified risk factors. The relationship is directly proportional. The dynamic programming solver 1042 sets the lower limit of the allowance size after semi-finishing to the sum of the standard allowance value and the safety margin increment value. This operation ensures that even if there is a deviation in the feature recognition size, the subsequent finishing process still has enough material to remove, preventing the part from being overcut and scrapped due to the front-end recognition size being too large.

[0093] The dynamic programming solver unit 1042 executes the constraint generation logic for process separation. The dynamic programming solver unit 1042 compares the identification risk factors for each feature. Separate from the preset process threshold. When risk factors are identified. When the process separation threshold is exceeded, the dynamic programming solver 1042 generates a mutual exclusion constraint. This mutual exclusion constraint prohibits the high-risk feature from being combined with other features in the same composite process for parallel machining. The dynamic programming solver 1042 forces the machining process of this feature to be an independent process step, and sets upper limits for the spindle speed and feed rate during the execution of this process step to reduce the potential risk of cutting force damage to uncertain geometry.

[0094] The dynamic programming solver 1042 merges the generated mandatory inspection constraints, margin compensation constraints, and process separation constraints into the constraint equation set of the global optimization mathematical model. Under the premise of satisfying all basic constraints and dynamic supplementary constraints, the dynamic programming solver 1042 solves the dynamic objective function constructed in the above embodiment. The dynamic programming solver 1042 outputs a final process route that satisfies the constraints and minimizes the generalized machining cost. This route is then converted into CNC machining code or process card file via the human-computer interaction and editing output module 105.

[0095] See attached document Figure 5 This invention provides a specific implementation example background and input content, which relates to the processing of a drawing of an automotive transmission driveshaft part. A dual-channel manufacturing feature recognition and process scheduling system 100 receives the automotive transmission driveshaft part drawing file. The automotive transmission driveshaft part drawing file is stored in DXF data exchange format.

[0096] The drawing file for the automotive transmission driveshaft includes the two-dimensional geometric contour data of the driveshaft. This data consists of multiple straight-line and circular arc entities. The drawing file also includes dimension entities, which indicate the diameter and length values ​​at key journals of the driveshaft. Finally, the drawing file contains non-geometric attribute annotations, including surface roughness symbols and geometric tolerance frames.

[0097] The vector data parsing unit 1011 reads the drawing file of the automotive gearbox drive shaft part. The vector data parsing unit 1011 extracts the geometric primitive information of key features. The key feature is identified as a precision journal. The vector data parsing unit 1011 analyzes and obtains the nominal diameter of the precision journal as 50.00 mm. The vector data parsing unit 1011 analyzes and obtains the nominal length of the precision journal as 120.00 mm. The vector data parsing unit 1011 extracts the associated tolerance zone designation text k6. The vector data parsing unit 1011 consults the standard tolerance table based on the tolerance zone designation text, obtaining an upper deviation value of +0.018 mm and a lower deviation value of +0.002 mm.

[0098] The visual feature detection unit 1012 renders the automotive gearbox driveshaft part drawing file into a raster bitmap image with a resolution of 4096 x 4096 pixels. The visual feature detection unit 1012 performs a target detection operation on the raster bitmap image. The visual feature detection unit 1012 detects a surface roughness annotation symbol in the image region corresponding to the precision journal. The visual feature detection unit 1012 identifies the numerical text contained in the surface roughness annotation symbol as 0.8. The visual feature detection unit 1012 outputs a visual confidence value for the surface roughness annotation symbol.

[0099] In this implementation example, the visual feature detection unit 1012 outputs a visual confidence score of 0.88 for the surface roughness marking symbol. The visual feature detection unit 1012 also detects a coaxiality tolerance frame. The visual confidence score output by the visual feature detection unit 1012 for the coaxiality tolerance frame is 0.65. The visual feature detection unit 1012 generates a set of visual features containing the above visual recognition results.

[0100] The dual-channel consistency verification unit 1013 receives the geometric dimension data and tolerance data output by the vector data parsing unit 1011. The dual-channel consistency verification unit 1013 also receives the visual feature set output by the visual feature detection unit 1012. The system then enters the feature alignment and risk quantification calculation stage, preparing to instantiate the enhanced feature vector using the extracted data.

[0101] Furthermore, a processing demonstration is provided, which is collaboratively executed by the various components of the dual-channel manufacturing feature recognition and process orchestration system 100. The dual-channel consistency verification unit 1013 first performs a spatial association operation. The dual-channel consistency verification unit 1013 determines that the precision journal geometry region extracted by the vector analysis channel overlaps with the surface roughness annotation symbols and coaxiality tolerance frames identified by the visual inspection channel in the pixel coordinate system. The dual-channel consistency verification unit 1013 establishes a unique feature index association.

[0102] Dual-channel consistency verification unit 1013 performs risk factor identification. The instantiation calculation is performed. For the surface roughness attribute, the visual confidence score is 0.88. The dual-channel consistency verification unit 1013 calculates the corresponding first risk component as 0.12. For the coaxiality tolerance attribute, the visual confidence score is 0.65. The dual-channel consistency verification unit 1013 calculates the corresponding second risk component as 0.35. The dual-channel consistency verification unit 1013 uses the maximum value selection logic to comprehensively identify the risk factors of this precision journal feature. The value is determined to be 0.35. The dual-channel consistency check unit 1013 constructs an enhanced feature vector. The specific numerical sequence is as follows: ; In the formula: Integer encoding for the cylindrical features; These are normalized geometric dimension parameters; These are the values ​​for upper deviation, lower deviation, and coaxiality tolerance. This represents the surface roughness value; To comprehensively identify risk factors.

[0103] The adaptive process matching unit 1041 receives the enhanced feature vector. The adaptive process matching unit 1041 queries the process knowledge base module 103 to obtain CNC turning and high-precision cylindrical grinding as candidate machining methods. The adaptive process matching unit 1041 reads the error sensitivity coefficient of the high-precision cylindrical grinding method. The preset value of this coefficient is 1.5. The adaptive process matching unit 1041 calculates the dynamic weight coefficient by combining the comprehensive identification risk factor of 0.35. Due to the high risk factor, the adaptive process matching unit 1041 automatically reduces the weight value of the dimensional accuracy matching dimension to prevent the algorithm from prematurely locking the grinding process parameters that are extremely sensitive to the input parameters.

[0104] The dynamic programming solver 1042 receives the optimal sequence of available process methods. The dynamic programming solver 1042 calculates the generalized machining cost. When calculating the virtual risk cost of the grinding process, the dynamic programming solver 1042 substitutes the comprehensive identified risk factor of 0.35 into the single-process generalized cost calculation formula. Due to the existence of the identified risk factor, the calculated cost of the grinding process increases, reflecting the potential scrap risk.

[0105] The dynamic programming solver 1042 dynamically supplements the logic based on the comprehensive identification of risk factor 0.35, which triggers the constraint conditions. Since 0.35 is greater than the preset mandatory inspection threshold of 0.30, the dynamic programming solver 1042 inserts an online probe inspection process between the semi-finishing turning process and the finishing grinding process. This process is used to measure the coaxiality error of the journal before grinding.

[0106] The dynamic programming solver 1042 simultaneously triggers the allowance compensation constraint logic. Since the comprehensive identified risk factor of 0.35 is greater than the preset allowance compensation threshold of 0.25, the dynamic programming solver 1042 adjusts the process parameters of the semi-finish turning operation. The dynamic programming solver 1042 increases the single-sided allowance after semi-finish turning from the standard 0.15 mm to 0.25 mm. This operation aims to compensate for the risk of grinding black skin or runout exceeding tolerances caused by the uncertainty in coaxiality identification.

[0107] The human-computer interaction and editing output module 105 ultimately outputs an adaptively adjusted process route: rough turning—semi-finish turning (with a machining allowance of 0.25mm)—online inspection (verifying coaxiality)—finish grinding (processing based on measured data). Compared to the standard process route, this route adds an inspection step and adjusts the machining allowance, demonstrating the system's adaptive fault-tolerant planning capability in the face of low confidence in visual recognition (coaxiality confidence of 0.65) in the input drawings.

[0108] See attached document Figure 6 This invention provides a final process route output, which is generated and output by a human-computer interaction and editing output module 105. The human-computer interaction and editing output module 105 receives optimized process sequence data transmitted by a dynamic programming solver unit 1042. The human-computer interaction and editing output module 105 calls a pre-set document generation engine to convert binary process objects in memory into structured documents conforming to the CAPP (Computer-Aided Process Design) standard.

[0109] The human-machine interaction and editing output module 105 generates the first operation instruction, marked as operation 10: rough turning of the outer diameter. The human-machine interaction and editing output module 105 writes the cutting parameters into this instruction, setting the spindle speed to 800 rpm and the feed rate to 0.25 mm per rpm.

[0110] The human-machine interaction and editing output module 105 generates a second process instruction, marked as process 20: semi-finish turning of the outer diameter. The human-machine interaction and editing output module 105 explicitly includes the dimensional control parameters calculated after risk compensation in this instruction. The human-machine interaction and editing output module 105 sets the target machining diameter to 50.50 mm. The human-machine interaction and editing output module 105 automatically fills in a single-sided allowance of 0.25 mm in the instruction remarks column; this value is directly derived from the risk factor identification method described in the aforementioned embodiment. The execution of the margin compensation logic.

[0111] The human-machine interaction and editing output module 105 generates the third process instruction, marked as process 30: online inspection. The human-machine interaction and editing output module 105 defines this as a mandatory execution step. The human-machine interaction and editing output module 105 specifies the inspection item as the coaxiality of the reference journal. The human-machine interaction and editing output module 105 sets the inspection equipment call instruction to enable the in-machine infrared probe. The human-machine interaction and editing output module 105 sets the pass / fail judgment logic: only when the measured coaxiality error is less than 0.02 mm is the process allowed to proceed to the next process; otherwise, an abnormal alarm is triggered and the machine tool is locked.

[0112] The human-machine interaction and editing output module 105 generates the fourth process instruction, marked as process 40: fine grinding of the outer diameter. The human-machine interaction and editing output module 105 sets this process to use an adaptive machining mode. The human-machine interaction and editing output module 105 configures the control system to read the measured geometric data generated in process 30 and dynamically adjust the feed origin coordinates of the grinding wheel accordingly. The human-machine interaction and editing output module 105 sets the final dimensional acceptance standard to 50.00 mm and the tolerance zone to k6.

[0113] The human-computer interaction and editing output module 105 constructs an output data package containing information from all the above-mentioned processes. The human-computer interaction and editing output module 105 uses JSON (JavaScript Object Notation) format to serialize and encapsulate data packets. Data packets The mathematical structure is defined as follows: ; In the formula: This is the header information, which includes the part name, gearbox driveshaft, and drawing number. It is an ordered collection of process objects, containing detailed parameters and control instructions for processes 10 to 40; This is a risk tracing log, recording the identification of risk factors. The source of the constraint changes and the resulting change records.

[0114] The human-computer interaction and editing output module 105 serializes the data packets. The data is written to local storage, generating a process data file with the .json extension. The human-computer interaction and editing output module 105 simultaneously calls the PDF rendering engine, based on the data packet... The content is presented in a visual process card document. The process card document highlights the allowance compensation value for process 20 and the mandatory inspection requirements for process 30 in red, intuitively reminding operators to pay attention to the special process requirements introduced by the uncertainty of front-end feature recognition.

[0115] The human-computer interaction and editing output module 105 transmits the process data file to the CNC machine tool controller or the workshop manufacturing execution system via a network interface. The system thus completes the entire closed-loop process from unstructured drawing input to adaptive process route output, achieving automatic correction and fault-tolerant planning of the process scheme even when feature recognition presents risks.

Claims

1. A process flow arrangement system based on mechanical CAD drawing recognition, characterized in that, It includes a CAD drawing parsing and feature recognition module, a manufacturing information structured storage module, a process knowledge base module, an intelligent process flow arrangement engine, and a human-computer interaction and editing output module; The CAD drawing parsing and feature recognition module is used to perform vector data parsing and visual feature detection on mechanical CAD drawing files, perform spatial alignment and consistency verification between vector features and visual features, calculate feature recognition risk factors, and output a reconstructed feature vector containing the feature recognition risk factors. The manufacturing information structured storage module is connected to the CAD drawing parsing and feature recognition module, and is used to receive the reconstructed feature vector and construct the part manufacturing information model; The process knowledge base module is used to store process flow arrangement rule data; The intelligent process flow orchestration engine connects the manufacturing information structured storage module and the process knowledge base module, and is used to call the reconstructed feature vector, dynamically adjust the process matching strategy and the process sorting optimization model according to the feature identification risk factors, and generate initial process flow data. The human-computer interaction and editing output module is connected to the intelligent process flow arrangement engine, and is used to perform visualization and editing operations on the initial process flow data, and output the finalized process flow.

2. The process flow arrangement system based on mechanical CAD drawing recognition according to claim 1, characterized in that, The CAD drawing parsing and feature recognition module includes a vector data parsing unit and a visual feature detection unit. The vector data parsing unit is used to parse the underlying data structure of mechanical CAD drawing files, extract geometric primitives, text annotations and layer attributes to construct a vector feature set; The visual feature detection unit is used to convert mechanical CAD drawing files into raster images and to use target detection algorithms to identify manufacturing features and construct a visual feature set.

3. The process flow arrangement system based on mechanical CAD drawing recognition according to claim 2, characterized in that, The vector data parsing unit is used to construct a closed geometric loop based on the topological relationship of geometric primitives, calculate the distance vector between the center coordinates of the bounding box of the text annotation object and the geometric center coordinates of the closed geometric loop, and associate the text annotation object with the closed geometric loop based on the distance vector. The visual feature detection unit is used to identify manufacturing features including thread processing texture, surface roughness symbol, geometric tolerance frame and datum symbol using a deep learning model, and outputs the visual feature set including feature category label, confidence value and bounding box coordinate data.

4. The process flow arrangement system based on mechanical CAD drawing recognition according to claim 2, characterized in that, The CAD drawing parsing and feature recognition module contains a dual-channel consistency verification unit. The dual-channel consistency verification unit is used to establish a unified drawing world coordinate system and to perform an inverse projection transformation operation on the visual feature set using a coordinate transformation matrix; The dual-channel consistency verification unit is used to retrieve candidate vector features that have spatial overlap with the elements of the visual feature set within the vector feature set using a spatial index tree, and to establish an alignment association index between the vector features and the visual features when the geometric bounding box intersection ratio is greater than a preset spatial association judgment threshold.

5. A process flow arrangement system based on mechanical CAD drawing recognition according to claim 4, characterized in that, The dual-channel consistency verification unit is used to calculate the feature identification risk factor based on the semantic comparison results and confidence scores. When vector semantics and visual semantics corroborate each other, the dual-channel consistency verification unit sets the feature recognition risk factor to a value that is negatively correlated with the confidence level. When there is a logical contradiction between vector geometric information and visual annotation information, the dual-channel consistency verification unit sets the feature recognition risk factor as a preset conflict penalty constant. In the absence of visual semantic support, the dual-channel consistency verification unit sets the feature recognition risk factor as a preset blind zone risk constant.

6. A process flow arrangement system based on mechanical CAD drawing recognition according to claim 1, characterized in that, The manufacturing information structured storage module is used to store part manufacturing information models that include feature lists, tolerance information, surface quality requirements, and material information. The reconstructed feature vector data structure includes manufacturing feature types determined by vector data parsing, normalized geometric dimension parameters, standardized tolerance information, surface roughness values, and the feature identification risk factor written as an independent dimension.

7. A process flow arrangement system based on mechanical CAD drawing recognition according to claim 1, characterized in that, The process knowledge base module is used to store process flow arrangement rule data, which includes processing methods and feature mapping rules, processing capacity database, process decision logic and standard process parameter library; The processing method and feature mapping rules define the correspondence between feature types and processing methods. The processing capability database records the performance parameters of machine tools and equipment. The process decision logic includes the process sequencing principle.

8. A process flow arrangement system based on mechanical CAD drawing recognition according to claim 1, characterized in that, The intelligent process orchestration engine includes an adaptive process matching unit. The adaptive process matching unit is used to identify risk factors based on the features, correct the preset feature type matching weight benchmark value and the preset size adaptability weight benchmark value, and calculate the dynamic weight coefficient. The adaptive process matching unit is used to calculate the matching score between features and candidate processing methods using dynamic weight coefficients. When the value of the feature identification risk factor increases, the feature type matching weight value is reduced to generate a preferred sequence of available process methods.

9. A process flow arrangement system based on mechanical CAD drawing recognition according to claim 1, characterized in that, The intelligent process flow orchestration engine contains a dynamic programming solution unit. The dynamic programming solution unit is used to identify risk factors and processing method error sensitivity coefficients based on the features, calculate nondeterministic virtual risk costs, add nondeterministic virtual risk costs to the process cost items included in the generalized processing cost objective function, and solve for the process sequence that minimizes the total value of the generalized processing cost objective function. The dynamic programming solver is used to dynamically generate one of the following constraints based on the identified risk factors: mandatory verification constraints, margin compensation constraints, and mutual exclusion constraints.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the functions of the process flow arrangement system based on mechanical CAD drawing recognition as described in any one of claims 1 to 9.

Images