Artificial Neural Network Architecture Design Using Similarity Measure With Applications in Engineering Monitoring and Diagnosis

Abstract

Artificial Neural Network (ANN) has been widely used for engineering monitoring and diagnosis. However, there are still several important problems unsolved and one of them is the architecture design of the ANN (namely, choosing the number of nodes in the hidden layer). In this technical brief, a new method of ANN architecture is introduced based on the idea that an ANN represents a mapping of training samples. Hence, the best ANN should represent the mapping that is most similar to the training samples. The method is tested using three practical engineering monitoring and diagnosis examples, including tool condition monitoring in turning, cutting condition monitoring in tapping, and metallographic condition monitoring in welding. It is demonstrated that the proposed method can improve the monitoring and diagnosis by approximately 3 percent.