A modular neural net approach for fault detection and diagnosis

This paper describes an approach based on the neural networks technique for fault detection and diagnosis in systems with discrete events and temporal constraints (like manufacturing systems). In the structure of a supervisor the authors identify the functions of detecting, diagnosis, decision and recovery. They focus their work on the fault detection and diagnosis. Each of these two functions is implemented with a multilayer neural network, using the backpropagation algorithm to learn normal and off-normal situations. As an application example, a factory cell with conveyors and a machining station is studied. The temporal constraints and the significant events are used to build the training set. A recognition set, containing unlearned situations, is then used to test the performances of this approach.<>