Session: 02-08-02: Innovative Product and Process Design II

Paper Number: 70255

Start Time: Thursday, 05:20 PM

70255 - Comparison of Clustering Techniques for Feature-Based Toolpath Generation in Dieless Manufacturing 

The aim of the current research is to compare the data clustering techniques for geometrical feature extraction. The CAD models of the geometries are sliced to generate the data sets for clustering. The K-means, Spectral, DBSCAN and Single Linkage Hierarchical clustering techniques are compared here for clustering arbitrary shaped contours and the formation of groups based on the means of the contours. The inputs required for clustering, e.g., the numbers of clusters need to be defined a priori; time taken for the contour clustering, ability to identify the arbitrary shaped contours, and the ability to identify the features that can be used for feature-based toolpath generation in dieless manufacturing applications are the factors that are considered in this work for the comparison.

The work includes the development of the feature extraction algorithms, consisting of contour clustering algorithm and group formation algorithm. The contour clustering algorithm separates multiple contours available on each slice, whereas the group formation algorithm forms the groups of the contours to build different features. Three different approaches to modify the feature extraction algorithms are designed and discussed in this paper. In the feature extraction algorithms, the DBSCAN clustering is replaced with K-means, Spectral, and Single Linkage Hierarchical clustering techniques, respectively. The outputs obtained by using DBSCAN clustering in the contour clustering algorithm and group formation algorithm are used as input to other clustering techniques.

In approaches one and two, only the contour clustering algorithm is modified. In approach one, the final number of features obtained from the DBSCAN clustering is used as input to the other clustering techniques; whereas, in approach two, the number of clusters obtained for each slice is used. It is found that, in approach one, all the clustering techniques other than DBSCAN fails to separate the available contours on the slices; while in approach two, along with the DBSCAN, Spectral, and Single Linkage Hierarchical are also able to separate the available contours on the slices. However, the DBSCAN performs the contour clustering faster than the Spectral and Single Linkage Hierarchical clustering techniques.

In the third approach, the second approach is combined with the modification in the group formation algorithm. In the group formation algorithm, the final number of features obtained from DBSCAN clustering is used as input to the other clustering techniques. From the comparative analysis, it is found that DBSCAN clustering and Single Linkage Hierarchical Clustering techniques can form the groups that can be used for feature-based toolpath generation in dieless manufacturing, whereas the other two fails to perform the same. The algorithms using DBSCAN clustering perform the group formation task faster than the Single Linkage Hierarchical clustering.

From the obtained results, it can be concluded that the DBSCAN clustering is the best suitable technique for contour clustering and group formation algorithms. Besides, if the number features are known in advance, the Single Linkage Hierarchical clustering can be used in place of DBSCAN clustering in the group formation algorithm.

Presenting Author: Puneet Tandon IIITDM Jabalpur

Authors:

Aniket Nagargoje deLOGIC Lab IIITDM Jabalpur
Pavan Kumar Kankar System Dynamics Lab, IIT Indore
Prashant Kumar Jain deLOGIC Lab IIITDM Jabalpur
Puneet Tandon deLOGIC Lab IIITDM Jabalpur