Equipment fault diagnosis—A neural network approach

Abstract

This paper describes a project involving the development of a neural network (connectionist) expert system for mechanical equipment fault diagnosis in an integrated steel industry. The objective of this work is to automate the diagnostic process by an artificial intelligence approach. The approach is characterized as a neural network expert system. It is different from traditional rule-based expert systems. A neural network consists of a collection of interconnected nodes. Each node has an activation value and each interconnecting arc has a numerical weight. The network can be trained by giving training examples. Learning here is the generation of optimal weights. The resulting weight matrix serves as the knowledge base of the system. A hybrid approach has been employed for inferencing, comprising forward chaining and backward chaining. These two processes continue as long as required by the satisfiability criteria. The system has a powerful explanation facility. An example for engine fault diagnosis is presented. Problems encountered during the development are discussed.