Machine learning and its applications in reliability analysis systems

Abstract

In this thesis, we are interested in exploring some aspects of Machine Learning (ML) and its application in the Reliability Analysis systems (RAs). We begin by investigating some ML paradigms and their- techniques, go on to discuss the possible applications of ML in improving RAs performance, and lastly give guidelines of the architecture of learning RAs. Our survey of ML covers both levels of Neural Network learning and Symbolic learning. In symbolic process learning, five types of learning and their applications are discussed: rote learning, learning from instruction, learning from analogy, learning from examples, and learning from observation and discovery. The Reliability Analysis systems (RAs) presented in this thesis are mainly designed for maintaining plant safety supported by two functions: risk analysis function, i.e., failure mode effect analysis (FMEA) ; and diagnosis function, i.e., real-time fault location (RTFL). Three approaches have been discussed in creating the RAs. According to the result of our survey, we suggest currently the best design of RAs is to embed model-based RAs, i.e., MORA (as software) in a neural network based computer system (as hardware). However, there are still some improvement which can be made through the applications of Machine Learning. By implanting the 'learning element', the MORA will become learning MORA (La MORA) system, a learning Reliability Analysis system with the power of automatic knowledge acquisition and inconsistency checking, and more. To conclude our thesis, we propose an architecture of La MORA.