Comparison of some Techniques for Learning-Based Pattern Classification in Technical Diagnosis

This paper reports a comparison study of 5 selected techniques for learning-based pattern classification. The techniques are the Linear Discriminant, the k-Nearest Neighbour Rule, the Classification Tree, and the 2 types of feed-forward artificial neural networks called the Backpropagation network and the Cascade-Correlation network. Each technique is briefly introduced and a test case involving a centrifugal pump driven by an electric motor is then presented. An extensive set of vibration and process measurements was collected for the machinery both in healthy and faulty conditions, and this was then used for training and testing the different classifiers.