Generation of diagnostic trees by means of simplified process models and machine learning

Abstract

Fault diagnosis by means of diagnostic trees is of considerable interest for industrial applications. The drawbacks of this approach are mostly related to the knowledge elicitation through laborious enumeration of the tree structure and ad hoc threshold selection for symptoms definition. These problems can be alleviated if a more profound knowledge of the process is brought into play. The main idea of the paper consists of modeling the nominal and faulty states of the plant by means of interval-like component models derived from first-principles laws, e.g. the conservation law. Such a model serves to simulate the entire system under different fault conditions, in order to obtain the representative patterns of measurable process quantities, i.e. training examples. To match these patterns by diagnostic rules, multistrategy machine learning is applied. As a result, binary decision trees that relate symptoms to faults are obtained, along with the thresholds defining the symptoms. This technique is applied to a laboratory test process operating in the steady state, and is shown to be suitable for handling incipient single faults. The proposed learning approach is compared with two related machine learning methods. It is found that it achieves similar classification accuracy with better transparency of the resulting diagnostic system.