A connectionist expert system approach to fault diagnosis in the presence of noise and redundancy

The author differentiates between physical redundancy involving duplicate measurements of the same quantity and analytical redundancy involving the behavior of a collection of sensors measuring different quantities. If there are a finite number of possible faults, if each fault has a known set of ideal instrument readings (in the absence of noise), and if a model of the noise is available, then analytical redundancy relationships exist. The task of constructing expert systems for problems involving noise and redundancy is then considered. The author reviews an automated method for constructing diagnostic expert systems (MACIE). This approach is based on machine learning techniques for connectionist network models and is well suited for noisy problems. The main advantage of the MACIE system is that it only requires training examples of desired behavior to generate the final expert system. Moreover, this approach takes advantage implicitly of both types of redundancy, without the need for explicit probabilistic analysis