Abstract A novel scheme is proposed in which a neural network is trained with qualitative data, based on information obtained from a Failure Mode and Effects Analysis. To perform the diagnosis, plant sensor data are normalised through comparison with appropriate reference values corresponding to the operating point of the circuit. Reference values which correspond to “correct” operation can be established through simulation or from measurement. The normalised data are quantized in order to classify signals as normal, high, low or zero signals with blurred boundaries. The quantized signal is used as an input to the trained network which, in turn, generates a set of candidate faults. The paper illustrates the approach using two simple hydraulic circuits and demonstrates the ability of the trained net to accommodate imprecise sensor data.
Edge et al. (Sun,) studied this question.