An Improved Artificial Neural Network Method for Monitoring and Diagnosis of Engineering Processes With Applications

Abstract

An artificial neural network (ANN) is a common method used for the monitoring and diagnosis of engineering processes. The use of ANN consists of two phases: training and reasoning. ANN requires supervised training, which involves designing its structure (select the number of layers and number of nodes at each layer) and setting up its target outputs. In general, the structure design is done by trial and error, and the target outputs are set as "1-0" (1 means a specified process condition occurs and 0 means otherwise). This article demonstrates that the performance of ANN can be improved by adopting proper structure and by using the mean of the training samples as the target outputs. The proposed method is based on the idea that the best ANN is the one being most similar to the training samples or, equivalently, the one minimizing the entropy of the training samples. A procedure of using ANN for monitoring and diagnosing engineering processes is developed accordingly. Using three practical examples (tool condition monitoring in turning, machining condition monitoring in tapping, and metallographic condition monitoring in welding), it is shown that the proposed procedure improves the performance of ANN by at least 10%. In addition, the training time is also reduced.