Innovative Production Machines and Systems Conference

A method for automatic generation of feature recognition rules that is applicable in different application domains is proposed. This is a new method for creating knowledge bases of AFR systems that elevates the knowledge acquisition issues associated with the development of rules-based AFR systems. The two most important characteristics of this method are:
1) The application of inductive learning techniques for identification of hidden patterns in sets of feature examples.
2) The utilisation of two representation schemes that code feature information at different levels of abstraction to complement and extend the learning capabilities of the method.
This research also suggests that the utilisation of inductive learning techniques for AFR has several advantages:
1) It provides a formal mechanism for rule definition and also assures the consistency of the generated rule sets.
2) The development of AFR systems for different application domains requires only representative training sets to be formed for each of them. This is a major advantage of the proposed approach due to the domain-dependent nature of features.
3) Due to the generalisation capabilities of inductive learning techniques, AFR systems could recognise features that are not present in the training sets.
4) The knowledge base of such systems could be extended easily to cover new or user-defined features.

The proposed method for automatic formation of feature recognition rules was implemented in the learning module of a prototype AFR system. This system also includes a feature recognition module that employs the rule bases generated applying the proposed method. Both modules of the prototype system were developed using the Java™ programming language. They use as input 3D CAD models in STEP format created using the STEP Application Protocol 203 supported by most commercially available CAD packages. Each module includes several parsers that extract automatically B-Rep data from CAD models in STEP format. The feature recognition module was applied on the test part shown in this slide, which was used for validation purposes by other researchers. The prototype system recognised nine of the eleven features present in this part. Two features could not be recognised because such feature classes were not defined in the taxonomy adopted in this study.

The RULES-5 algorithm is applied consecutively on both training sets to extract rules that depict feature patterns at both levels of abstraction. In particular, RULES-5 created 15 rules from the first training set that encapsulates information at the lowest level of abstraction. The application of the same algorithm on the second training set resulted in 9 rules. This set of rules is shown in the table included in this slide.

This illustrative example presents a possible implementation of the proposed method for automatic formation of feature recognition rules. In this example, the taxonomy that is adopted categorises features belonging to the machining domain. One of the B-Rep models used as an example of a blind hole is shown in this slide together with its code/characteristic vector generated at the highest level of abstraction.

To address knowledge acquisition issues associated with the development of AFR systems, a new method for generating automatically feature recognition rules is proposed. In particular, these rules are formed by applying an inductive learning algorithm on training data consisting of feature examples. The creation of training data is comprised of three steps. First, a taxonomy that represents the feature classes for a given application domain is defined. Second, a set of B-Rep models representing examples of features is designed for each class of a given taxonomy. Third, the B-Rep models of features are converted into data files that are suitable for inductive learning. In particular, two levels of abstraction are considered for coding topological and geometrical information from the B-Rep feature models. The algorithm utilised in this research is RULES-5. During the induction process, RULES-5 repeatedly takes an example not covered by the previously created rules and forms a new rule for it until all examples in the training data are covered by the generated rule set.

For the last twenty-five years, many AFR techniques have been proposed for recognising both simple and interacting features. Such techniques implement approaches based on rules or neural networks for example. However, the focus of the research efforts was on developing techniques for recognising features in the context of a particular manufacturing application, and especially for a range of machining processes. A true CAD/CAM integration requires the development of AFR systems that could be easily applied in different domains. For example, such a system could be used to assess the manufacturability of a given product or part with regard to various production processes.
The objective of this research is to address knowledge acquisition issues associated with the development of AFR systems that could be employed in different application domains.

The design of products and their consecutive manufacturing requires information at different levels of abstraction. One of the data representation schemes that is widely used to interface Computer-Aided Design (CAD) and Computer-Aided Manufacturing (CAM) processes is the Boundary Representation (B-Rep) scheme. In particular, the geometrical data stored using the B-Rep scheme cannot be utilised directly for process design because this data lacks high-level geometrical entities that are meaningful from a manufacturing point of view. To bridge this information gap between CAD and CAM, Automatic Feature Recognition (AFR) techniques are applied to identify geometrical entities, features in the CAD model, which are semantically significant in the context of specific downstream manufacturing activities.

This presentation will start by introducing the problem that automatic feature recognition techniques are trying to solve. Then, the specific objective of this research will be stated and the proposed method for automatic formation of feature recognition rules will be described. An example will be presented in order to illustrate the suggested method. Then, the results obtained from this implementation will be tested on a benchmarking CAD model. Finally, the presentation will summarise the main conclusions from this research.

This paper discusses the application of inductive learning techniques for creation of rule sets that could be utilised for Automatic Feature Recognition (AFR) in 3D solid models. AFR techniques are an important tool for achieving a true integration of Computer-Aided Design (CAD) and Computer-Aided Manufacturing (CAM) processes. In particular, AFR systems allow the identification in CAD models of high-level geometrical entities, features that are semantically significant for manufacturing operations. In this paper, a method is proposed to meet the specific requirements imposed by the utilisation of inductive learning for acquisition of feature recognition rules. The method presented in this study is implemented within a prototype feature recognition system and its capabilities are verified on a benchmarking part.