MBD model processing feature recognition and extraction method and system based on neural network

Abstract

The invention belongs to the technical field of MBD application, and discloses an MBD model processing feature recognition and extraction method and system based on a neural network, and the method comprises the steps: firstly constructing an MBD ontology model and a process model of an optical component, converting a three-dimensional model into a two-dimensional feature image, and generating a data set; training a processing feature classifier by using a neural network, and carrying out feature recognition on the image and video file of the three-dimensional model; and finally, in combination with PMI information and a knowledge base, calculating a matching degree and screening a processing method to form feature recognition extraction and optimization. According to the method, the relevance of geometric and process characteristics is autonomously extracted from dimension reduction data by using a neural network. The calculation burden of traditional three-dimensional voxel processing is overcome, and the recognition precision of complex curved surfaces (such as aspheric surfaces and microstructures) is remarkably improved. Multi-modal feature joint reasoning is realized by combining PMI information, a design-manufacturing closed loop is finally formed, and human experience dependence in process iteration is effectively reduced.

Description

Technical Field

[0001] The present invention belongs to the field of MBD application technology, and in particular relates to a method and system for identifying and extracting MBD model processing features based on a neural network. Background Art

[0002] Traditional machining feature recognition relies on manual experience or rule-based geometric analysis. This makes it difficult to guarantee efficiency and accuracy when dealing with optical components with free-form surfaces and microstructures. This is especially true when dealing with the dynamic matching of unstructured features with process parameters, where the lack of semantic information from the 3D model often leads to delayed process decision-making. While existing MBD technologies integrate 3D models with manufacturing information, they lack automated feature extraction mechanisms for optical machining characteristics. The static association between process knowledge bases and design models makes them difficult to adapt to the demands of high-variety, low-volume production. Furthermore, deep learning algorithms based on 3D voxels suffer from excessive computational resource consumption and redundant feature representation. Summary of the Invention

[0003] In view of this, the present invention provides a method and system for identifying and extracting MBD model processing features based on a neural network to solve the above problems.

[0004] To solve the above technical problems, the present invention provides a method for identifying and extracting MBD model processing features based on a neural network, comprising:

[0005] S1. Construct the MBD ontology model of the optical component;

[0006] S2. Constructing an MBD process model based on the process information and step information of the optical components;

[0007] S3, converting the three-dimensional model of the optical element into a corresponding two-dimensional processing feature image to form a preliminary feature data set; cleaning and annotating the preliminary feature data set to generate a processing feature data set as an input parameter of the neural network;

[0008] S4. Input the processing feature data set into the neural network for training, and obtain the processing feature classifier through multiple iterations and optimization of the network structure and hyperparameters;

[0009] S5. Preprocessing the three-dimensional model to obtain image files and video files, using a processing feature classifier to perform processing feature recognition on the image files and video files, obtaining processing features contained in the optical element and outputting recognition results of the processing features;

[0010] S6. Extract the PMI information of the optical component, calculate the matching degree between the MBD ontology model features and the knowledge base based on the preset process flow and knowledge base, and screen the processing method based on the matching degree.

[0011] As an optional method, in the above S1, the features of the MBD ontology model include geometric features, material features, surface features, surface features and usage features; wherein,

[0012] Geometric features include planes, free-form surfaces, spheres, aspheric surfaces and dimensional tolerances; dimensional tolerances also include perpendicularity, thickness and overall dimensions;

[0013] Material characterization includes optical glasses and crystals;

[0014] Surface features include surface error, surface roughness and frequency; surface error also includes PV value and RMS value, surface roughness includes Ra value, Rz value, Rt value, Rq value, Rsk value, Rku value and P value, and frequency characteristics include low frequency, medium frequency and high frequency;

[0015] Surface features include pinholes, pits, and scratches;

[0016] Causing characteristics are used to express damage thresholds.

[0017] As an optional method, process information includes process number, process name, process content, process inspection, whether it is a critical process, whether it is a special process and whether it is a final product inspection; work step information includes work step sequence number, work step number, work step name, work step content and processing requirements, whether it is a critical work step and precautions; MBD process model consists of annotation information, process attribute information and three-dimensional models corresponding to parts at different production stages.

[0018] As an optional method, in the above S3, cleaning and labeling include:

[0019] Images that do not meet the preset acquisition standards are eliminated, and LabelImg is used to annotate each 2D processing feature image and save it as a text document. The text document includes the location information and category label of each processing feature.

[0020] As an optional method, in the above S4, the YOLOv5 algorithm is used for feature recognition, and its recognition process includes:

[0021] Define the image path and clear the table data, initialize the YOLOv5 model, and determine the number of images;

[0022] If it meets the preset requirements, read the image for target detection:

[0023] If an object is detected, draw the bounding box and label, convert the Chinese name of the label, add the result to the table, and display the image in the corresponding label;

[0024] If no target is detected, the image is displayed directly on the corresponding Label;

[0025] If the number of images is greater than the preset value, the YOLOv5 object is released and the process ends.

[0026] As an optional method, 3D model annotation is performed based on the S4 recognition results, including:

[0027] Establish 3D annotation and MBD information annotation for optical components, interpret the annotated information to obtain dimension information, processing information and annotation information; merge the corresponding information to generate PMI information and processing features, and then enter them into the information management system for annotation management.

[0028] As an optional method, in the above S6, based on the UG platform and using the UG / Open API interface provided by it, the PMI information in the MBD model is automatically extracted, and the process information of the parts is classified into design information, basic attribute information and processing technology information to participate in the construction of the MBD process model; wherein,

[0029] Design information is process information directly associated with the geometric features of the 3D model, including dimensional information, tolerance information, surface accuracy information, roughness information, surface error information, and damage threshold information. The above information can be directly marked on the 3D model. Design information is composed of the design information of the design model, the design information of the blank model, and the design information of the process model, and is the basis for all model building.

[0030] Basic attribute information is a general description of the optical component, including name, material, quantity and number, which can be stored in the form of an attribute information table and associated;

[0031] Processing technology information is the process information generated by parts during the production and manufacturing stage, including tool attribute information of machine tools, cutting tools, and fixtures used in the processing project, as well as processing parameter information such as spindle speed, feed rate, and back-cutting depth. Processing technology information is directly related to the process flow of parts, guiding the processing and manufacturing process of parts, and associating processing technology information with corresponding process models in the form of information tables and information texts.

[0032] As an optional method, in the above S6, the distance information between the processing feature and the part material is obtained by the nearest neighbor algorithm, the distance information between the main dimension and the characteristic dimension during the processing is obtained, the root mean square value, peak-to-valley value and surface roughness of the surface accuracy are obtained, and the upper and lower tolerances of the main dimension are obtained;

[0033] Set up target layer, criterion layer and solution layer. The target layer is used to determine the processing method; the criterion layer is used to reflect the key information that affects the selection of the processing method, including geometric features, material indicators, surface indicators, surface features and service features; the solution layer includes the processing methods in the knowledge base;

[0034] A judgment matrix of characteristic attributes is constructed, and the characteristic attributes of the information are scaled according to the definition of the judgment matrix scale to obtain the final judgment matrix. After the final judgment matrix is normalized, the weight of each key information feature is obtained; after introducing the random consistency index, the consistency ratio is calculated, and the comprehensive distance measurement value is calculated by weighted summation to evaluate the similarity between the information to be matched and the process knowledge in the knowledge base, and based on the similarity, the processing method is extracted.

[0035] As an optional method, when calculating similarity, first calculate the comprehensive distance measurement values of all candidate process knowledge and sort them in ascending order; then the processing method with the smallest comprehensive distance measurement value is displayed on the system interface as the optimal matching result; if the comprehensive distance measurement values of multiple process knowledge are the same, the process knowledge with the highest similarity calculated first is given priority.

[0036] On the other hand, the present invention also provides a neural network-based MBD model processing feature recognition and extraction system, comprising:

[0037] Model building module, used to build MBD body model and MBD process model of optical components;

[0038] A dimension conversion module is used to convert the three-dimensional model of the optical element into the corresponding two-dimensional processing feature image to form a preliminary feature data set;

[0039] A preprocessing module is used to clean and annotate the preliminary feature data set to generate a processed feature data set as an input parameter of the neural network; and to preprocess the three-dimensional model to obtain image files and video files;

[0040] The neural network module is used to process feature data sets and input them into the neural network for training. After multiple iterations and optimizations, the network structure and hyperparameters are adjusted to obtain the processing feature classifier.

[0041] A feature recognition module is used to use a processing feature classifier to perform processing feature recognition on image files and video files, obtain processing features contained in optical elements and output recognition results of processing features;

[0042] The matching calculation module is used to extract the PMI information of optical components, calculate the matching degree between the MBD ontology model features and the knowledge base based on the preset process flow and knowledge base, and screen the processing methods based on the matching degree.

[0043] The beneficial effects of the present invention are:

[0044] The present invention maps three-dimensional geometric features into two-dimensional high-density information carriers by constructing an MBD ontology model and a process model. Combined with a dynamic recognition and matching mechanism driven by a neural network, it forms a logical system from feature extraction to process decision-making. This effectively solves systemic problems encountered by traditional methods, such as information loss during data conversion, insufficient intelligent level of process reasoning, and low efficiency of multimodal manufacturing data collaboration, and provides a solution for the integration of design and process in the field of optical manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic flow chart of a method for identifying and extracting machining features of an MBD model provided by an embodiment of the present invention;

[0046] Figure 2 A schematic diagram illustrating the association between the process structure tree and the MBD process model provided in an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of MBD ontology model information provided by an embodiment of the present invention;

[0048] Figure 4 A schematic diagram of the classification of optical element process information provided by an embodiment of the present invention;

[0049] Figure 5 A schematic diagram of the identification process of an optical element processing feature identifier provided by an embodiment of the present invention;

[0050] Figure 6 Schematic diagram of the hierarchical structure of optical element processing information provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with specific implementation methods.

[0052] See also Figures 1-6 This embodiment provides a method for identifying and extracting MBD model processing features based on a neural network, comprising:

[0053] S1. Construct the MBD ontology model of the optical component;

[0054] S2. Constructing an MBD process model based on the process information and step information of the optical components;

[0055] S3, converting the three-dimensional model of the optical element into a corresponding two-dimensional processing feature image to form a preliminary feature data set; cleaning and annotating the preliminary feature data set to generate a processing feature data set as an input parameter of the neural network;

[0056] S4. Input the processing feature data set into the neural network for training, and obtain the processing feature classifier through multiple iterations and optimization of the network structure and hyperparameters;

[0057] S5. Preprocessing the three-dimensional model to obtain image files and video files, using a processing feature classifier to perform processing feature recognition on the image files and video files, obtaining processing features contained in the optical element and outputting recognition results of the processing features;

[0058] S6. Extract the PMI information of the optical component, calculate the matching degree between the MBD ontology model features and the knowledge base based on the preset process flow and knowledge base, and screen the processing method based on the matching degree.

[0059] In the above S1, this embodiment first performs MBD body modeling on the optical element, and assigns it geometric features, material features, surface features, surface features and service features.

[0060] Among them, geometric features include planes, free-form surfaces, spheres, aspheric surfaces and dimensional tolerances, and dimensional tolerances include perpendicularity, thickness and overall dimensions; material features include two major categories: optical glass (fused quartz) and crystals; surface features include surface error, surface roughness and frequency, and surface error includes PV value and RMS value, surface roughness includes Ra value, Rz value, Rt value, Rq value, Rsk value, Rku value and P value, frequency characteristics include low frequency, medium frequency, and high frequency; surface features include scraping defects, pitting defects and scratch defects; service characteristics express its damage threshold.

[0061] In the above S2, in an optional scenario, the process information is set to include the process number, process name, process content, process inspection, whether it is a critical process, whether it is a special process, and whether it is a final product inspection; the work step information includes the work step sequence number, work step number, work step name, work step content and processing requirements, whether it is a critical work step, and precautions and remarks. For MBD process modeling, this embodiment takes into account that the part process model based on MBD can effectively solve the problem of poor correlation between process information and three-dimensional models in the design and processing of optical elements. Directly applying the MBD process model to guide the subsequent production and processing of parts can greatly shorten the production cycle and improve production efficiency. This embodiment defines the part process model Mtm based on MBD as follows:

[0062] Mtm={Mg,Am,Ap}

[0063] Where Mg represents the 3D model of the part's MBD process model, including a 3D model that reflects the part's shape and function, as well as a 3D model that reflects the material removed from the part during the various stages of the manufacturing process. Am represents annotation information associated with the part's 3D model, such as basic dimensions, tolerances, surface accuracy, and surface roughness. Ap represents the part's process attribute information, which includes basic process information such as the part's name, material, and quantity, as well as processing information related to the part's manufacturing process, such as cutting tools, machine tools, fixtures, processing methods, and processing parameters.

[0064] Since the manufacturing process of a part is a process of obtaining a product part by continuously removing materials from the blank according to the designed processing route, the MBD process model of the part can well show the changes in this process from a geometric perspective. Therefore, the three-dimensional model Mg of the MBD process model of the part can be defined as:

[0065] Mg={Md,Mr,Mp1,Mp2,……,Mpn}

[0066] In the formula, Md represents the design model that reflects the three-dimensional shape and function of the product part. Mr represents the blank model that reflects the three-dimensional shape and function of the machined part. Mpi represents the process model corresponding to the remaining portion of the part after material is removed through the i-th machining process. represents the number of machining processes included in the part manufacturing process. According to the definition of the MBD-based part process model, the part MBD process model consists of annotation information, process attribute information, and the corresponding three-dimensional models of the part at different production stages.

[0067] This example then uses the UG / OpenAPI interface provided by the UG platform to automatically extract PMI information from the MBD model. Corresponding information extraction modules are developed for different types of PMI information (including dimensional tolerances, geometric tolerances, surface shape requirements, and service requirements). All extracted PMI information is stored in TXT format. This can include: removing images that do not meet preset acquisition standards; using LabelImg to annotate each 2D machining feature image and saving it as a TXT file containing the location information and category label for each machining feature.

[0068] In order to better manage the process information of parts, it is necessary to classify the process information of parts. Based on the different expressions of process information, this paper divides the process information stored in the MBD process model into: design information, basic attribute information and processing technology information.

[0069] Design information refers to process information directly associated with the geometric features of a 3D model, such as dimensions, tolerances, surface accuracy, roughness, surface shape error, and damage threshold. This information can be directly annotated on the 3D model. Design information is composed of design information for the design model, the rough model, and the process model, and serves as the basis for all model building.

[0070] Basic attribute information is a general description of the part, such as the part's name, material, quantity, number, etc. It has no direct relationship with the geometric features of the model, so it generally does not need to be directly marked on the three-dimensional model. This embodiment associates the basic attribute information with the design model in the form of an attribute information table.

[0071] Machining process information is generated during the manufacturing phase of a part. It includes tool attributes such as the machine tool, cutting tool, and fixture used in the machining process, as well as machining parameters such as spindle speed, feed rate, and back-cut depth. Machining process information is directly linked to the part's manufacturing process and guides the manufacturing process. Therefore, this article associates this information with the corresponding process model in the form of information tables and text files, making it easier for process personnel to directly retrieve and view it.

[0072] For S3 and S4, this embodiment adopts a processing feature recognition method based on the YOLOv5 algorithm, taking advantage of the algorithm's efficient feature extraction and real-time advantages, and using the MBD three-dimensional model of optical components to accurately identify processing features. By achieving information exchange between CAD and CAPP systems, data barriers are broken down to provide technical support for intelligent manufacturing.

[0073] First, a processing feature dataset was constructed. Based on the principle of processing feature classification, a three-dimensional model of the optical component was constructed using UG three-dimensional design software. The three-dimensional model was converted into its corresponding two-dimensional processing feature image using the screenshot method, ensuring that the image could reflect the geometric form of the processing feature. In order to improve the diversity and adaptability of the dataset, a multi-angle and multi-level method was adopted when taking screenshots. These operations increased the richness of the dataset. After these processed image sets, a preliminary dataset containing diverse processing features was formed. Then, in order to make the dataset meet the training requirements of the deep learning model, LabelImg software was used to annotate each processing feature image. The annotation results of each image will be automatically saved as a TXT format file. These files contain the specific location information and category labels of the processing features in the image. Finally, through this series of steps, the construction and preprocessing of the processing feature dataset were completed.

[0074] Secondly, a processing feature identifier based on the YOLOv5 algorithm is constructed. In order to achieve efficient and accurate processing feature recognition, it is necessary to construct a processing feature recognition classifier based on the YOLOv5 algorithm. The above-mentioned pre-processed processing feature data set is input into the constructed YOLOv5 model as the data source for network training. Subsequently, through multiple rounds of training and parameter optimization, the network structure and hyperparameters are continuously adjusted to improve the recognition accuracy and training efficiency of the model. During the training process, the algorithm extracts the key feature points of the processing features from a large number of samples through a learning mechanism, and can gradually adjust the weight values in the network, thereby achieving accurate recognition of different processing features. After sufficient training and verification, a YOLOv5 deep neural network model with optimal performance and the highest recognition accuracy is obtained. Finally, the trained YOLOv5 model is encapsulated and deployed in the CAPP system developed in this embodiment.

[0075] The recognition process includes: defining the image path and clearing the table data, initializing the YOLOv5 model, and determining the number of images; if it meets the preset requirements, reading the image for target detection: if the target is detected, drawing the bounding box and label, converting the Chinese name of the label and adding the result to the table, and displaying the image in the corresponding label; if no target is detected, displaying the image directly in the corresponding label; if the number of images is greater than the preset value, releasing the YOLOv5 object and ending.

[0076] Finally, machining feature recognition is performed on the 3D model of the optical component. Within the CAD system, UG secondary development technology is used to preprocess the 3D model of the optical component and convert it into image and video files. This ensures that various machining features can be effectively extracted and identified during subsequent processing. The generated image and video files are then transferred to the CAPP system. Within the CAPP system, machining feature recognition is performed on the input image and video files using a pre-built machining feature classifier. Based on a pre-trained deep learning model, the machining feature classifier automatically identifies various machining features contained in the optical component, such as planes, spherical surfaces, and aspherical surfaces. After recognition is complete, the system outputs the machining feature recognition results in two formats: an image format, where the identified machining features are visually displayed within the optical component image as bounding boxes, making them easy for designers and engineers to quickly understand and review; and a text format, where the system lists the identified machining features in text format, facilitating subsequent process design and data archiving.

[0077] After completing processing feature recognition and PMI information extraction, since the two types of information come from different processing flows, the processing features and their corresponding PMI information are separated and cannot be directly associated. To achieve efficient application of information, this embodiment systematically integrates and processes the originally scattered and disorganized data into a detailed information set based on processing features. This integrated information set is called the Processed Feature Information Set (PFIS). It provides effective data support for intelligent process design by accurately mapping processing features and PMI information. It can be described by the following formula:

[0078] PFIS={T, D1, D2, D3, PV, T u , T l}

[0079] Where T is the type of processing feature, D1 is the first dimension information of the processing feature (such as the length of the optical element), D2 is the second dimension information (such as the width of the optical element), D3 is the third dimension information (such as the height of the optical element), PV is the surface error, T u is the upper tolerance, T l The upper and lower tolerances are the tolerance information of the main dimensions of the processing features.

[0080] The core task of the machining feature information set is to associate each machining feature with its corresponding PMI information. This organized information becomes more organized and clearly reflects information such as the dimensional requirements, tolerance standards, and surface roughness of each machining feature. This not only helps improve data readability but also allows this information to be directly applied to subsequent process design. In intelligent process design, in particular, using these structured data sets, designers can determine the machining process and select appropriate machining equipment based on specific machining feature types, thereby providing data support for the entire production process.

[0081] In addition, this embodiment is also provided with a three-dimensional model annotation system, including: establishing three-dimensional annotations and MBD information annotations for optical components, interpreting the annotated information to obtain size information, processing information and annotation information; merging the corresponding information to generate PMI information and processing features and then entering them into the information management system for annotation management.

[0082] The ontology of an optical element can describe geometric information and non-geometric information in a unified syntax. However, the ontology modeling process is highly subjective, making its results lack authority and professionalism. To this end, this embodiment introduces the STEP / STEP-NC standard, which provides an authoritative and professional high-level information model for describing physical objects in the processing field. Optionally, two sub-processes, geometric information and non-geometric information, can be set, and finally the processing steps are used as indexes to merge and obtain the implementation data ontology. The construction of the implementation model focuses on the implementation data, so this embodiment selects the processing feature as the smallest unit to describe the ontology. After constructing the processing feature ontology, the implementation geometric data can be imported to generate a single optical element ontology instance. Subsequently, by adding rules, all instances of this type of optical element can be generated at one time.

[0083] By using the feature tree traversal method, all PMI information in the MBD ontology model is scanned. Based on the process execution results and the nearest neighbor algorithm, the optical component implementation model is optimized by setting weights to complete the optimization of the optical component process implementation model, including:

[0084] After years of research and practice, the inventors took into account that the core of the nearest neighbor algorithm lies in the calculation of similarity, and the similarity matching between knowledge depends on the measurement of this similarity. When performing similarity calculations, commonly used distance measurement methods include Euclidean distance, Manhattan distance, cosine similarity, and Minkowski distance. Different numerical types and data characteristics determine which distance measurement method to choose. According to the specific application scenario and data characteristics, selecting a suitable measurement method is the key to ensuring matching accuracy. In an optional scenario, the information required for matching the processing method of this embodiment includes processing features, feature size 1, feature size 2, feature size 3, surface roughness, upper tolerance, lower tolerance, and part material. In order to facilitate the description of the matching process, this embodiment represents the information to be matched as X (x1, x2, x3, x4, x5, x6, x7, x8), where x1 to x8 respectively represent the various pieces of information matched by the above-mentioned processing method. The relevant information already existing in the knowledge base is represented as Y(y1, y2, y3, y4, y5, y6, y7, y8). The distance measurement calculation method between different information is as follows:

[0085] Machining features and part materials are both categorical variables, and there is no direct or indirect correlation between them. Different machining features and materials can lead to significant differences in the selection of machining methods. For example, holes are typically machined using drilling, while slots require milling. Therefore, to measure the differences between these categorical variables, the Hamming distance is used as a distance metric. The specific calculation method is shown below:

[0086]

[0087] Among them, d i is the distance metric, x i is the processing feature or part material information of X, and y is the processing feature or part material information of Y. If the two information are the same, the value is 0, otherwise it is 1.

[0088] In this embodiment, three dimensions of the processing feature are extracted. When determining the processing method, it is usually only necessary to consider the dimension that has the greatest impact on the processing process, that is, the main dimension. However, in some cases, other dimensions of the processing feature may also have a significant impact on the processing method. Therefore, in such cases, the impact of other dimensions on the processing method must be comprehensively considered. Based on this, this embodiment selects weighted Euclidean distance as the distance measurement method for feature dimensions to more accurately reflect the relative importance of each dimension. The specific calculation formula is shown below:

[0089]

[0090] Among them, d2 is the distance measurement value of the characteristic size, x2, x3, and x4 are the characteristic size 1, characteristic size 2, and characteristic size 3 of X respectively, y2, y3, and y4 are the characteristic size 1, characteristic size 2, and characteristic size 3 of Y respectively, and w1, w2, and w3 are the weights of the characteristic size 1, characteristic size 2, and characteristic size 3 respectively. In this embodiment, the characteristic size 1 is the main size of the processing feature, so w1=0.8, w2=0.1, and w3=0.1 are taken, and the sum of the three is 1.

[0091] In order to accurately capture nonlinear features such as PV value, RMS value, and surface roughness Rq value, this embodiment uses Manhattan distance as the distance measurement method, which is calculated as shown in the following formula:

[0092] d5=|x5-y5|

[0093] Wherein, d5 is the distance measurement value of the surface roughness, x5 is the surface roughness information of X, and y5 is the surface roughness information of Y.

[0094] During the process design, the choice of processing method mainly depends on the dimensional accuracy requirements of the parts. However, since dimensional accuracy itself is difficult to directly quantify using traditional distance measurement formulas, this embodiment uses the upper and lower tolerances of the main dimensions as quantitative indicators of dimensional accuracy, and uses weighted Manhattan distance to evaluate the differences between different tolerance requirements. The weighted Manhattan distance can effectively reflect the width of the tolerance zone and its impact on the selection of processing methods. Its calculation formula is shown below:

[0095] d6=w ９ |x6-y6|+w x |x7-y7|

[0096] Among them, d6 is the distance measurement value of the tolerance, x6 and y6 are the upper tolerance values of X and Y respectively, x7 and y7 are the lower tolerance values of X and Y respectively, and w ９ and w x are their respective weight coefficients. In this embodiment, w ９ =0.5, w x =0.5, the sum of the two is 1.

[0097] After calculating the distance metric values for each piece of information, its weight must be determined. Because the choice of processing method is influenced by multiple dimensions of information, including processing characteristics, feature dimensions, surface roughness, and tolerance requirements, and these factors have complex interactions, this embodiment uses the analytic hierarchy process (AHP) to determine the weight values for each piece of information.

[0098] The purpose of building a hierarchical structure is to clearly divide the hierarchical relationship of each decision-making factor, thereby forming a clear and systematic hierarchical indicator system. In this embodiment, the entire hierarchical structure is designed and divided into three main levels: the target level, the criterion level, and the solution level.

[0099] The goal layer, the highest level in the hierarchy, represents the ultimate goal of the entire decision-making process: selecting the appropriate machining method. The determination of the machining method directly impacts production efficiency, product quality, and cost control, making it a crucial decision in process design. Therefore, the goals of this layer are clear and guiding, forming the core of the entire decision-making process. The criteria layer then primarily includes key information influencing machining method selection: geometric characteristics, material specifications, surface characteristics, surface features, and service characteristics. These criteria represent the various process conditions that must be considered when selecting a machining method, providing multi-dimensional support and basis for decision-making at the goal layer. Finally, the solution layer primarily encompasses various machining methods within the knowledge base. These methods represent machining knowledge accumulated based on experience, specifications, and historical data, encompassing all technical solutions related to the goal and criterion layers. The solution layer provides specific implementation plans for the decisions made at the upper layers, selecting the most suitable machining method within the requirements and constraints defined by the criterion layer.

[0100] The judgment matrix U of the feature attribute is as shown below:

[0101]

[0102] Among them, u ij for u i To u j The relative importance scale of the information is then used, and then the characteristic attributes of the information are scaled according to the definition of the judgment matrix scale, as shown in Table 1.

[0103] Table 1 Judgment matrix scale and its definition

[0104] <![CDATA[Scale (u ij )]]> Definition (Importance) 1 <![CDATA[u ｉ and u j Equally important]]> 3 <![CDATA[u ｉ and u j Slightly more important]]> 5 <![CDATA[u i and u j Obviously more important]]> 7 <![CDATA[u i and u j Strongly important]]> 9 <![CDATA[u ｉ and u j Extremely important<!-- 8 --> ]]> 2、4、6、8 Importance is between 1, 3, 5, 7, 9 <![CDATA[Reciprocal of scale (u j )]]> <![CDATA[u ｊ =in i -1 ]]>

[0105] This embodiment comprehensively considers the influence of geometric features, material indicators, surface indicators, surface features, and service characteristics on the processing method, and obtains the importance judgment between the information, as shown in Table 2.

[0106] Table 2 Judgment of importance between information

[0107] Geometric features Material indicators Surface characteristics causative features Face shape indicators Geometric features 1 3 3 5 7 Face shape indicators 1 / 3 1 1 3 5 Surface characteristics 1 / 3 1 1 3 5 causative features 1 / 5 1 / 3 1 / 3 1 3 Material indicators 1 / 7 1 / 5 1 / 5 1 / 3 1

[0108] The judgment matrix U is finally obtained from the above table, as shown in the formula:

[0109]

[0110] Normalize each column of the judgment matrix U using the following formula:

[0111]

[0112] in, is the normalized value of the element in row i and column j, is the sum of all elements in the jth column. The normalized data is shown in Table 3, and the results retain three decimal places.

[0113] Table 3 Normalized values of matrix U

[0114] Geometric features Material indicators Surface characteristics causative features Face shape indicators Geometric features 0.498 0.542 0.542 0.405 0.333 Material indicators 0.166 0.181 0.181 0.243 0.238 Surface characteristics 0.166 0.181 0.181 0.243 0.238 causative features 0.099 0.060 0.060 0.081 0.143 Face shape indicators 0.071 0.036 0.036 0.027 0.048

[0115] Next, take the average of the normalized values in each row to get the weight of each piece of information. The arithmetic mean is calculated using the following formula:

[0116]

[0117] Among them, w i is the average value of the i-th row, that is, the weight of the item, and m is the number of columns in the matrix. Based on the above data processing, the weight coefficient of each information obtained in this embodiment is: w i =(0.464, 0.202, 0.202, 0.089, 0.043), which are the weight coefficients of processing features, dimensional tolerances, surface roughness, feature dimensions, and materials, respectively. The result is retained to three decimal places.

[0118] Since it is difficult to achieve completely consistent measurements when comparing information attributes pairwise, there will usually be certain errors. Therefore, it is necessary to perform consistency testing on the judgment matrix to ensure its accuracy. The consistency index Cl is introduced as shown in the following formula:

[0119]

[0120] Among them, λ max is the maximum eigenvalue of the matrix U, which is calculated to be 5.1284. n is the matrix order. All calculations result in a Cl value of 0.0321.

[0121] Then, the consistency ratio CR is calculated, and the average random consistency index Rl needs to be considered. The Rl value is found through the order of the matrix and Table 4.

[0122] Table 4 RI values of judgment matrix

[0123] Order 1 2 3 4 5 6 7 8 9 RI 0 0 0.58 0.90 1.12 1.24 1.32 1.41 1.45

[0124] In this embodiment, the order of the matrix U is 5, so the Rl value obtained by looking up the table is 1.12. The CR value is calculated as shown in the following formula:

[0125]

[0126] From the above formula, we can get the CR value of 0.029, that is, CR<0.10, so the matrix U has satisfactory consistency.

[0127] After completing the distance metric calculation for each piece of information and determining its weight coefficient, the comprehensive distance metric value is calculated by weighted summation to evaluate the similarity between the information to be matched and the process knowledge in the knowledge base. The calculation of the comprehensive distance metric value is shown in the following formula:

[0128] d2=0.464d1+0.202d2+0.202d5+0.089d6+0.043d8

[0129] Where d2 is the comprehensive distance metric between the information to be matched and a piece of process knowledge in the knowledge base. The smaller this value, the higher the similarity and the better the match. During the processing method matching process, the system calculates the comprehensive distance metric between the information to be matched and each piece of process knowledge in the knowledge base according to the above formula and selects the processing method with the highest similarity.

[0130] To ensure the uniqueness of the matching results, this embodiment adopts the following strategy: First, the comprehensive distance metric values of all candidate process knowledge items are calculated and sorted in ascending order. The processing method with the smallest comprehensive distance metric value (i.e., the highest similarity) is then displayed as the optimal match on the system interface. In extreme cases, if multiple pieces of process knowledge have the same comprehensive distance metric value, the system will prioritize the process knowledge item with the highest similarity calculated first to avoid ambiguity in the matching results.

[0131] On the other hand, this embodiment also provides a neural network-based MBD model processing feature recognition and extraction system, including:

[0132] Model building module, used to build MBD body model and MBD process model of optical components;

[0133] A dimension conversion module is used to convert the three-dimensional model of the optical element into the corresponding two-dimensional processing feature image to form a preliminary feature data set;

[0134] A preprocessing module is used to clean and annotate the preliminary feature data set to generate a processed feature data set as an input parameter of the neural network; and to preprocess the three-dimensional model to obtain image files and video files;

[0135] The neural network module is used to process feature data sets and input them into the neural network for training. After multiple iterations and optimizations, the network structure and hyperparameters are adjusted to obtain the processing feature classifier.

[0136] A feature recognition module is used to use a processing feature classifier to perform processing feature recognition on image files and video files, obtain processing features contained in optical elements and output recognition results of processing features;

[0137] The matching calculation module is used to extract the PMI information of optical components, calculate the matching degree between the MBD ontology model features and the knowledge base based on the preset process flow and knowledge base, and screen the processing methods based on the matching degree.

[0138] Through the above scheme, this embodiment projects the three-dimensional model of the optical element into a two-dimensional feature image, and uses a neural network to autonomously extract the correlation between geometric and process features from the dimensionality-reduced data. It not only overcomes the computational burden of traditional three-dimensional voxel processing, but also significantly improves the recognition accuracy of complex surfaces (such as aspheric surfaces and microstructures). Through end-to-end learning, the neural network accurately captures implicit features such as curvature continuity and smoothness, and combines PMI information to achieve multi-modal feature joint reasoning, ultimately forming a design-manufacturing closed loop, effectively reducing the reliance on human experience in process iteration.

[0139] The above are merely preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be construed as limiting the present invention, and the scope of protection of the present invention should be determined by the scope defined in the claims. Persons skilled in the art will appreciate that improvements and modifications may be made without departing from the spirit and scope of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying and extracting MBD model processing features based on neural network, characterized in that: include: S1. Construct the MBD ontology model of the optical component; S2. Constructing an MBD process model based on the process information and step information of the optical components; S3, converting the three-dimensional model of the optical element into a corresponding two-dimensional processing feature image to form a preliminary feature data set; cleaning and annotating the preliminary feature data set to generate a processing feature data set as an input parameter of the neural network; S4, inputting the processing feature data set into a neural network for training, and obtaining a processing feature classifier through multiple iterations and optimization adjustment of the network structure and hyperparameters; S5. Preprocessing the three-dimensional model to obtain an image file and a video file, performing processing feature recognition on the image file and the video file using a processing feature classifier, obtaining processing features contained in the optical element, and outputting recognition results of the processing features; S6. Extract the PMI information of the optical component, calculate the matching degree between the MBD ontology model features and the knowledge base based on the preset process flow and knowledge base, and screen the processing method based on the matching degree.

2. The method for identifying and extracting MBD model processing features based on a neural network according to claim 1, characterized in that: In the above S1, the features of the MBD body model include geometric features, material features, surface features, surface features and usage features; wherein, The geometric features include planes, free-form surfaces, spherical surfaces, aspherical surfaces and dimensional tolerances; the dimensional tolerances also include perpendicularity, thickness and overall dimensions; Said material characteristics include optical glass and crystals; The surface features include surface error, surface roughness and frequency; the surface error also includes PV value and RMS value, the surface roughness includes Ra value, Rz value, Rt value, Rq value, Rsk value, Rku value and P value, and the frequency characteristics include low frequency, medium frequency and high frequency; The surface features include scratch defects, pit defects and scratch defects; The servitude characteristic is used to express the damage threshold.

3. The method for identifying and extracting MBD model processing features based on a neural network according to claim 1, characterized in that: The process information includes the process number, process name, process content, process inspection, whether it is a critical process, whether it is a special process and whether it is sent for standard inspection; the work step information includes the work step sequence number, work step number, work step name, work step content and processing requirements, whether it is a critical work step and precautions and remarks; the MBD process model is composed of annotation information, process attribute information and three-dimensional models corresponding to parts at different production stages.

4. The method for identifying and extracting MBD model processing features based on a neural network according to claim 1, characterized in that: In the above S3, the cleaning and labeling include: Images that do not meet the preset acquisition standards are eliminated, and LabelImg is used to annotate each two-dimensional processing feature image and save it as a text document. The text document includes the location information and category label of each processing feature.

5. The method for identifying and extracting MBD model processing features based on a neural network according to claim 1, characterized in that: In the above S4, the YOLOv5 algorithm is used for feature recognition, and its recognition process includes: Define the image path and clear the table data, initialize the YOLOv5 model, and determine the number of images; If it meets the preset requirements, read the image for target detection: If an object is detected, draw the bounding box and label, convert the Chinese name of the label, add the result to the table, and display the image in the corresponding label; If no target is detected, the image is displayed directly on the corresponding Label; If the number of images is greater than the preset value, the YOLOv5 object is released and the process ends.

6. The method for identifying and extracting MBD model processing features based on a neural network according to claim 5, characterized in that: 3D model annotation based on S4 recognition results, including: Establish 3D annotation and MBD information annotation for optical components, interpret the annotated information to obtain dimension information, processing information and annotation information; merge the corresponding information to generate PMI information and processing features, and then enter them into the information management system for annotation management.

7. The method for identifying and extracting MBD model processing features based on a neural network according to claim 3, characterized in that: In the above S6, based on the UG platform and using the UG / OpenAPI interface provided by it, the automatic extraction of PMI information in the MBD model is realized, and the process information of the parts is classified into design information, basic attribute information and processing technology information to participate in the construction of the MBD process model; wherein, The design information is process information directly associated with the geometric features of the 3D model, including dimensional information, tolerance information, surface accuracy information, roughness information, surface shape error information, and damage threshold information. The above information can be directly marked on the 3D model. The design information is composed of the design information of the design model, the design information of the blank model, and the design information of the process model, and is the basis for all model building. The basic attribute information is a general description of the optical element, including name, material, quantity and number, which can be stored in the form of an attribute information table and participate in the association; The processing technology information is the process information generated by the parts during the production and manufacturing stage, including the tool attribute information of the machine tools, cutting tools, and fixtures used in the processing project, as well as the processing parameter information of the spindle speed, feed rate, and back cutting amount; the processing technology information is directly related to the process flow of the parts, guiding the processing and manufacturing process of the parts, and associating the processing technology information with the corresponding process model in the form of information tables and information texts.

8. The method for identifying and extracting MBD model processing features based on a neural network according to claim 1, characterized in that: In the above S6, the distance information between the processing feature and the part material is obtained by the nearest neighbor algorithm, the distance information between the main dimension and the characteristic dimension during the processing is obtained, the root mean square value, peak-to-valley value and surface roughness of the surface accuracy are obtained, and the upper and lower tolerances of the main dimension are obtained; Setting up a target layer, a criterion layer and a solution layer, wherein the target layer is used to determine the processing method; the criterion layer is used to reflect the key information affecting the selection of the processing method, wherein the key information includes geometric features, material indicators, surface indicators, surface features and service features; The solution layer includes processing methods in the knowledge base; A judgment matrix of characteristic attributes is constructed, and the characteristic attributes of the information are scaled according to the definition of the judgment matrix scale to obtain the final judgment matrix. After normalizing the final judgment matrix, the weight of each key information feature is obtained; after introducing the random consistency index, the consistency ratio is calculated, and the comprehensive distance measurement value is calculated by weighted summation to evaluate the similarity between the information to be matched and the process knowledge in the knowledge base, and based on the similarity, the processing method is extracted.

9. The method for identifying and extracting MBD model processing features based on a neural network according to claim 8, characterized in that: When calculating similarity, first calculate the comprehensive distance measurement value of all candidate process knowledge and sort them in ascending order; then the processing method with the smallest comprehensive distance measurement value is regarded as the optimal matching result and displayed on the system interface; If the comprehensive distance measurement values of multiple pieces of process knowledge are the same, the process knowledge with the highest similarity calculated first will be selected first.

10. A neural network-based MBD model processing feature recognition and extraction system, characterized in that: include: Model building module, used to build MBD body model and MBD process model of optical components; A dimension conversion module is used to convert the three-dimensional model of the optical element into the corresponding two-dimensional processing feature image to form a preliminary feature data set; A preprocessing module, configured to clean and label the preliminary feature data set to generate a processed feature data set as an input parameter of a neural network; and preprocessing the three-dimensional model to obtain image files and video files; A neural network module is used to input the processing feature data set into the neural network for training, and obtain a processing feature classifier through multiple iterations and optimization adjustments of the network structure and hyperparameters; A feature recognition module is used to use a processing feature classifier to perform processing feature recognition on image files and video files, obtain processing features contained in optical elements and output recognition results of processing features; The matching calculation module is used to extract the PMI information of the optical component, calculate the matching degree between the MBD ontology model features and the knowledge base based on the preset process flow and knowledge base, and screen the processing method based on the matching degree.

Images