Neuro-fuzzy approaches to fault diagnosis and identification

Systems are becoming increasingly complex, with widespread use made of autonomous controllers, as in the case of decentralised control for complex chemical or flexible manufacturing systems. Furthermore, autonomous control system design focuses on the identification of analytical (as opposed to hardware) `redundancy´ in the form of a control theoretic plant model. The major drawbacks of any such analytical framework are: the computational overheads associated with supporting real time operation and, more significantly, the inability of the detection system performing any fault diagnosis operation to distinguish between modelling errors and the failure modes. This leads to a significant emphasis being placed on identifying a suitably robust representation of the plant. To be more specific, we identify three sources of deviation from the nominal analytical plant model: 1. fault conditions-we interpret these as having a deterministic form; 2. modelling errors-these we classify as having a possibilistic (fuzzy) interpretation; errors as a result of modelling approximations such as over simplification or idealising assumptions; 3. system and measurement noise-this represents a probabilistic (random) source of error