Hybrid approach towards the assessment of a drill condition using deep learning and the Support Vector Machine

摘要

This paper describes an application of a novel method which relies on applying the fusion of two sets of features obtained from deep learning and hand crafted features. Moreover, in terms of deep learning, transfer learning in deep learning has been applied. A hybrid method has been compared to classical deep learning approach, and let us state that the new approach can obtain more accuracy than regular methods. Transfer learning in deep learning should be used in case of lack of data to train the use of deep learning algorithms. We have used this method for the classification of drill wear state on the basis of drilled hole images. The specialists divided the whole data set into three classes: red, yellow, and green, which correspond to 3 stages of drill wear. The red class corresponds to a drill which is worn out and should be replaced; the yellow class should send a warning message to an operator for a manual check of the state of a drill; and the last one, the green one, is connected with the state of a drill which should be further used in production (it is still sharp enough). The important advantage of this approach is that it collects features from two sources: one set is automatically extracted by means of deep learning and the other is manually extracted by a researcher. The next advantage of this approach is training classification models only on the basis of a small portion of data, which in case of popular deep learning methods is too small to achieve reasonable accuracy. Hence, in order to get around this issue connected with a small portion of training data, transfer learning in deep learning has been applied. The achieved results confirm the fact that this approach can be applied in this situation.