The basic requirement in designing an operational support system for faults diagnosis is in interpretation of transients, knowledge acquisition and knowledge representation. This study uses a:wavelet transform to extract features of the transients which are then interpreted in terms reflecting the main features of an ART2 neural network. The detailed results are stored in a knowledge base embedded in trained fuzzy feedforward neural networks. The system can deal with numerical dynamic signals and carry out heuristic reasoning, and the knowledge base is easily maintained and expanded, and so provides an evolutionary adaptive capability. The approach is illustrated by reference to an industrial fluid catalytic cracking process.