A novel and smart interactive feature recognition system for rotational parts using a STEP file

Ketan and Yaqoub [10] proposed an AFR system for recognising the symmetrical features of rotational parts. In this system, a STEP file is used to save a 2D upper profile of the part from a CAD package. The system also includes the implementation of the Sweeping Primitive Rule (SPR) algorithm, which was constructed based on three techniques: syntactic pattern recognition, sweeping operation, and logic rule based. An example was presented to validate the approach using mechanical desktop 0.9 in order to draw the part profile. Whilst a 2D part representation can be created using lines and arcs, in this work the part profile is restricted to containing only line segments. Also, the system is limited to recognising thirteen predefined symmetrical rotational features.

Meeran and Pratt [11] applied both syntactic pattern recognition and intelligent logic techniques to recognise the features of a 2D prismatic part design using DXF file. The entities in the DXF file are classified and resorted into three groups pertaining to the front, side, and top views. These three groups are then exported to the recogniser, which is provided with a library of predefined features based on syntactic pattern recognition, for recognising and categorising simple isolated features. Loops of entities that correspond to predefined features are removed from the database, whilst the remaining entities are examined using intelligent logic to check whether they correspond to one of three possibilities: intersecting, protrusion, or depression features. According to the authors, the sequence of recognition processes is not in any sense optimal. Furthermore, the adopted approach is capable of recognising very specific and simple feature types along with more general protrusion and depression ones.

Yip-Hoi et al. [12] utilised a 2D area decomposition method for recognising turned features from axisymmetric parts. It starts with identifying the machining axis and determining the Maximum Turn-able State (MTS) of the part, which is the state of the part in which moving more material will gouging surfaces of the final part. The maximum turn-able volume (MTV) is the material that can be removed by turning, and it is found by subtracting the MTS from the stock workpiece. The system consists a set of predefined features with a unique combination of elements for each feature. A pattern recognition algorithm is used to search for matching between the maximum turn-able volumes (MTVS) decomposition and each predefined feature pattern. After finishing the feature recognition, three steps of precedence generation are included: determining the feature’s ordering, mapping volume-to-operations, and mapping operation-to-resource. This is to be used as a part of a CAPP system. However, the implementation of this method was limited for the 2D patterns that contains only straight-lines segments. Based on the authors, it is possible to extend it to include arcs or other curved segments.

Sivakumar and Dhanalakshmi [13] developed an AFR logic rule-based methodology to extract manufacturing features for cylindrical parts. The system comprises four major units, with the first including a part model created using Unigraphics and saved as a STEP file. A set of nine different cylindrical features was defined in the second unit. Next, a feature analyser was developed using the C++ language to extract a part’s features and save it in a text file. Finally, the extracted data with other information, such as type of material and cutting conditions, were used to generate an NC code. During the experimental work, the authors claimed a range of deviations in dimensions between 0.082 and 4.125% from the reference data. However, eight of the nine predefined features were identified based on one surface and hence, other features that require more than one surface were not included, for example, a square groove that comprises three adjacent surfaces.

In a multi-tasking machine, both turning and milling machining can be performed. ZHU et al. [14] applied the concept of attributed adjacency graph (AAG) to recognise cylindrical and prismatic features, which makes it valuable in multi-tasking machines. According to the authors, the system is capable of generating process planning from a CAD model, which is achieved by saving it as a STEP file and sending it to the feature recogniser. In total, the system is able to recognise nine turning features and eight milling ones. Then, the information of the manufacturing features is returned to the user, in order to select the cutting tools manually or automatically. The final step includes the generation of process planning, which comprises the sequence of operations and the NC tool path data. The generated process planning is optimised based on the machining cost evaluation.

Kang et al. [15] presented an approach to link CAD and CAPP systems via STEP files and a hint-based technique. The system includes four steps, starting with the generation of a STEP 203 file from UniGraphics CAD software. Next, the part’s geometrical information of the STEP 203 is transformed into Parasolid entities. This is to check for the correctness of the STEP 203 import and to assign other information to the following step, such as surface roughness, and dimensional or geometric tolerances. In the third step, the hint-based concept is implemented to convert the geometric features into manufacturing features by applying the Integrated Incremental Feature Finder (IF2). Finally, the resultant manufacturing information is translated into a physical STEP AP224 file, which contains the relevant process planning data to manufacture the designed part. Despite the system being able to remove the main barrier between CAD and CAPP, recognition of complex shapes remains a bottleneck, which requires a mature feature recognition system.

Sunil and Pande [4] developed an intelligent Artificial Neural Network (ANN) system to recognise prismatic features from B-Rep CAD models. In this approach, a unique 12-node vector scheme was created in order to classify machining features as families based on their geometrical and topological information. In this system, the input is a SAT file exported from Solidworks software, which is sent to a pre-processing module to construct a part face adjacency graph and to create a feature representation vector (FRV). All the FRVs are presented to the trained ANN unit for feature classification. Finally, all the classified features from the ANN unit, with their relevant parameters, are post-processed to create a feature-based model. Hence, the feature-based model file can be linked to a CAPP system for CNC machining.

A hybrid AFR method was developed by Rameshbabu and Shunmugam [16] to recognise the intersecting manufacturing features of a prismatic part using a STEP AP203 file. The AFR algorithm is a combination of face adjacency graph and volume subtraction techniques. This is to simplify a feature’s graph by reducing the number of faces in a feature’s volume and hence, the complexity of feature representation is minimised. The proposed system is also able to generate a part of CAPP in terms of determining a setup sequence to machine a component’s features. Other aspects of CAPP, such as machine tool selection, operation selection, operation sequencing and machining sequencing, have not been considered, since it is assumed that only one machine is required to manufacture the product. Li et al. [17] also presented a hybrid AFR approach for the purpose of recognising intersecting features from a design part. In this system, three AFR techniques are used, which are hint based, AAG, and ANN. In the design part, the face loops are defined as generic feature hints, which can be extracted and used as clues for interacting features based on an enhanced attributed adjacency graph (EAAG) algorithm. Next, the relationships between the face loops are determined in order to create F-Loop Graphs (FLGs). The use of FLGs simplifies the graph representation of a part in the next step, which includes a final feature recognition using an ANN unit. Although the system can recognise complex intersecting features, it is restricted to planar faces and quadric surfaces. Tseng and Joshi [18] presented a machining volume generation method for recognising rotational and prismatic features of mill-turn parts. In order to generate feature volumes, the boundary faces of a part are swept along a direction determined based on the type of machining operation. The proposed method includes four main steps. Firstly, the part’s faces are classified according to their geometric shapes and topological relationships as cylindrical, planar, or other types of faces. Then, several rotational machining zones, which represent cylindrical portions of the part that can be produced in turning processes, are determined. Rotational and prismatic machining volumes are generated using rotational sweeping operations for the former and volume decomposition, maximal volume sweeping, and reconstruction for the later. Finally, two approaches are used to recognise different types of features; these are 2D profile patterns for rotational features and 3D face adjacency relationship for prismatic features. Although many types of rotational and prismatic features can be recognised following this method, it has several limitations. For example, a complex geometry can be recognised as rotational or prismatic features, which is unreliable in such cases and requires extra work. Liu et al. [19] utilised the concept of AAG to analyse the geometrical and topological relations of parts’ machined surfaces. Following certain geometric reasoning rules, the AAG net is created by adding, deleting, and improving some linking relations between the graph nodes. Then, the resultant AAG is decomposed into a number of subgraphs in order to construct local rotational primitives. The 2D section of the turning volumes between the stock workpiece and the MTS model are meshed using the local-slicing approach. Also, a 2D shape descriptor for turning feature types was developed. Both the meshes and the descriptor helped creating a layer tree-based element sorting approach, which was established for identifying turning features.

However, beside the mentioned methods, there are many feature recognition techniques that serve different manufacturing processes in addition to those of machining. For example, Shi et al. [20] presented a feature recognition approach using heat kernel signature (HKS), for manufacturability analysis in Additive Manufacturing (AM). This includes an investigation regarding unique characteristics of AM processes, such as unsupported feature, minimum feature size, maximum vertical aspect ratio, minimum spacing, and minimum self-supporting angle. The proposed method has the ability to identify geometric features and manufacturing constrains of different shapes. Such technique can be justified to be compatible with the recognition of parts’ features manufactured using machining processes.

Based on the literature review of AFR techniques, it is concluded that a particular emphasis has been placed on solving some issues, such as intersecting features. Also, features that contain toroidal surfaces have to be recognised, in the same way that features containing planar surfaces are currently identified. Whilst a few systems have been able to deal with these matters, other barriers are still unsolved. For example, any AFR system requires a library comprising a set of predefined features; however, there is none that can be expected to include all the possibilities of features, which could exist in a part’s design. Furthermore, a new feature recognition system is needed, which can automatically recognise new types of features and add them to its database [21].

The STEP AP 203 file is a language-based text file, containing strings and entities describing a product’s representation [25], being divided into two main sections: “HEADER” and “DATA”. The HEADER section provides general product information, such as file name, creation date, company name, and programmer name. The DATA section contains lines of entities, which represent the main geometrical and topological product data. Each entity instance in the DATA section starts with “#” followed by a unique integer, entity name, and related data. The related data can be numbers, strings, Boolean, or even a reference to another entity instance in the file.

The geometrical and topological product data is organised into a hierarchical structure, as shown in Fig. 2. The top level of description in this hierarchy is the shell, which is a topological item that bounds a region in 3D space by joining a number of faces along edges. For example, the cylinder in Fig. 3a is considered a “CLOSED_SHELL”, with the STEP AP 203 file for it containing 253 entity lines in its DATA section. Whilst the STEP file has a hierarchical structure and shows the entity lines in an ordered form, the data arrangement does not appear in a hierarchical sequence in the physical text file. Hence, the “CLOSED_SHELL” node, which represents the top of the hierarchy, does not necessarily exist in the first line of entities. In fact, it could be in any order between the beginning and the end of the DATA section. This statement holds true for all the other levels of data.

The faces represent the second level of description in a STEP AP 203 file. A face is a topological entity that describes a piece of a surface bounded by loops, which shares an edge with exactly one other face to form a CLOSED_SHELL. Figure 3b shows the four faces of the cylinder and some other details, including the edges and vertices of the first face #47. Each face can be defined using two pointers: bound and surface. The bound always takes the first pointer, whereas the second pointer denotes the surface of the face. The ADVANCED_FACE entity ends with a Boolean flag, which indicates whether the loop direction is oriented in accordance with or opposed to the surface normal [26].

As mentioned above, the data in a STEP AP 203 file, which is represented by entity lines, does not appear in a specific logical order. Consequently, the file is almost unreadable since it requires moving from line to line in a non-sequential way. This is unreadable not only for humans, but also for feature recognition systems. It is therefore necessary to develop a parser so as to re-arrange the data in a comprehensible way. The parser that was developed in this work imports a STEP AP 203 file and scans all of its lines to find how many closed shells there are in the design as well as how many faces each contains. Then, each face is displayed with all of its geometrical and topological information before moving to the next face. When a face is declared, the first bound, which might be either face bound or face outer bound, will appear in the following line. Next, the parser shows the edge-loop with indicators of the edges that form the face. All the edges are given with their related information until the lowest level of Cartesian points. If the face contains another bound, the parser displays it directly in the same way as the previous one, otherwise the surface type of the face is declared with all of its details before moving on to the next face. Figure 4a and b show part of the original and the resultant STEP file AP 203 after applying the parser, respectively. The parser is developed using the C# programming language, and it aims to facilitate the feature recognition task. However, more manipulating, editing, and filtering are needed before moving on to the feature recognition module. This will be achieved using developed algorithms that are explained in the next section in detail.

Different algorithms have been proposed for this system to manipulate the results from the parser before moving on to the feature recognition module. These have the aim of facilitating the task of the feature recogniser, by delivering the information of the part’s design in a more organised way.

By using different algorithms to manipulate the results from the STEP file parser, the data describing a part are organised as external and internal faces nodes. Each of these nodes is intended to be used in the recognition phase of predefined features. The proposed AFR system adopts a rule-based approach, whereby it recognises a feature by scanning the faces in each node and matching them with the predefined features based on certain rules that are characteristic to that feature. The details about scanning faces and recognising features are as follows. The SI-AFR system’s database contains varying types of features in terms of the number of faces that form one, and

depending on the geometrical and topological description, a feature might include from one to five faces. Thus, the predefined feature set is categorised into five groups, with those in each group having the same number of faces. Regardless of the number of faces that form the feature, each of these predefined ones has a unique description, which is saved in the database, with two extra faces denoted as F_gbandF_ga representing guides before and after a feature, respectively. For example, a square groove feature with two round corners in the base is shown in Fig. 14 and saved in the database, as in Table 1.

The SI-AFR system works as loops. Backing to the sorted external and internal face nodes, the first loop starts by holding a bundle of the first seven faces from a node and matches them with each predefined feature that includes five related faces. If a match is found:

On the other hand, if no matching is confirmed, the first loop is continued, and the bundle is updated by keeping the first six faces and subtracting the last one. Then, the system compares the amended bundle with all the predefined features that include four related faces. This process continues, with the same decisions being taken, until all the faces in a node are covered. However, any loop is stopped, if the faces of a bundle reaches three items F_gb + F_n + F_ga and no matching is found. In this case, the face F_n is declared as undefined, and the second part of the system, which is the interactive feature recognition, is activated. Figure 15 shows a map of the predefined feature recognition loops.

One of the contributions with this methodology is the reduction of the stages required to recognise some multi-face features. This can be explained by taking different types of square groove features. The previous AFR systems, which were developed for rotational parts, can recognise blind square grooves, as in Fig. 16. This blind square groove consists of three faces, these being: a cylindrical surface F_{n + 1} in the middle and two bigger surfaces F_n and F_{n + 2} on the sides, which might be cylindrical, conical, or toroidal. However, most of these AFR systems have not been developed to recognise a square groove with extra components and hence, the square groove with extra rounded corners in the base of Fig. 14 is declared as five separated faces F_n + F_{n + 1} + F_{n + 2} + Fn + 3 + F_{n + 4}. This is an unreliable solution and consequently, such systems require extra processing in order to recognise this type of feature. The proposed SI-AFR system has been developed to expect such features and thus, avoids the need for extra processing.

A part has been designed using SOLIDWORKS and saved as a STEP AP 203. This model has only predefined features, which form the closed-shell comprising an external shape and three internal ones. Figure 19 shows the top section of the part with its details, where the red numbers and arrows represent the external shape, and the blue, purple, and orange are the internal shapes. After importing the STEP file into the system, the developed parser reads it, and shows the details to the user as a text via the main window. In this example, the closed-shell has 64 faces, which will be manipulated in the next step, before moving on to the feature recognition. The manipulating processes include: check the convexity or concavity of toroidal faces, finding MaxZ and MinZ as well as MaxX and MinX for each. Then, merge similar cylindrical faces, similar toroidal faces, adjacent toroidal faces and similar conical faces. Also, sorting faces based on MaxZ and splitting ones that related to holes. The final step in this example is the feature recognition process, when the system compares the part’s faces with the set of predefined features using the proposed methodology provided in Section 4.1. Moreover, it provides information to the user about the external and three internal shapes. Whilst both the external and internal shapes include the feature’s name and information about the faces that form it, the specific details for them are different. That is, regarding the external shape, the system gives the following for each face: X and Z start point values (X_sp, Z_sp), X and Z end point values (X_ep, Z_ep) as well as the type of movement from the start to end points (linear or circler). Also, if the face is toroidal, additional data are provided: X centre value of the curve (X_cc), curve radius (CR), and the direction of the curve (CW or CCW). Furthermore, the maximum depth (D) and width (W ) of each feature are calculated and printed in the interface window. Whereas the internal shapes’ information are detailed as follows: X centre value (X_c), Y centre value (Y_c), Z centre start (Z_cs), radius of Z_cs (RZ_cs), Z centre end (Z_ce), radius of Z_ce (RZ_ce), and the type of movement from the start to end points (linear or circler). Also, the X centre value of the curve (X_cc), curve radius (CR), and the direction of the curve (CW or CCW) are provided in the case of toroidal faces. Figure 20 and Tables 3, and 4 show the main window after applying the three steps of parsing, manipulating, and recognition; the details of the recognised external features; and those of the three internal shapes, respectively.

All the details that the system provides can be used in a smart CAPP system, which is the objective of future work. This does not include just the point values, the types of movement from point to point and the direction, for the system also demonstrates in an efficient way the location and specification of the internal shapes. In this example, the through, front, and back internal shapes are provided, as in Fig. 21. Such information can be used whilst generating CAPP, because it helps to select the cutting tool path and whether this part requires being flipped or not. For instance, the CAPP system would select the required tools, processes, and sequence of operations to manufacture both the front and back holes. However, the CAPP system is provided with sufficient information so that when the front hole is manufactured, the system gives instruction to flip the part before machining the back hole.

This example explains how the system identifies a new feature and adds it to the database. For this purpose, the previous design has been used with an alteration, whereby the square groove feature (4) is replaced with an undefined feature, as shown in Fig. 22. To avoid repetition, the steps of the parser, modification, and predefined feature recognition will not be described again here. During the feature recognition process, the new feature is declared in a pop-up window, as in Fig. 23, which has only one meaning: “the smart interactive feature recognition has been activated”. By clicking on the “OK” button, another pop-up window is shown, which includes two text fields. The first is automatically filled with an explanation about the new feature and provides information to the user, such as the number of faces that form the feature as well as the type and specification of each of these faces. In addition, it provides details about the faces that come before and after the feature. Such information allows for understanding regarding the geometrical and topological nature of the new feature, thus helping the user to suggest a name in the second text field, as shown in Fig. 24. In this example, the user has suggested a feature name “V groove concave base”, which is the closest to the feature description. Whilst it is possible to avoid this simple interaction part and give the new feature a name automatically, it makes more sense to leave this option for the user, because giving a random name to the new feature might not be satisfying to the user since this name does not describe the nature of the new feature.

By clicking on the “Save Feature” button, the feature is saved in the database with its name and description. Hence, the total number now of the predefined features in the database is 55, instead of the initial number of 54.

The proposed SI-AFR system provides high efficiency regarding the required time to accomplish the feature recognition task. For example, the total time that the system takes to recognise the external and internal features of the model design in Fig. 22 is 5 seconds. This includes one second for each of the parsing process, achieving all the manipulating algorithms, and recognising all the predefined features, as well as two seconds to diagnose, process, and add the new feature to the system’s database.