A method comprises obtaining signals from a plurality of sensors Si, S2, Sn and feeding these into an encoder 12, which translates signals into vectors VE characterising the state of the operational parameters of the monitored system. The encoded vector VE is fed into a translation engine 13 which translates the encoded vector VE into feature space to form a feature vector VF. The feature vector VF is subsequently fed into a residual vector generator 16 which compares the feature vector VF with a predicted vector Vp generated by a prediction engine 14 and thereby output a residual vector VR which characterises any differences between the feature vector VE and the predicted vector Vp. In addition, the feature vector VF may also be fed directly into the prediction engine 14 to update subsequent predictions. The formed residual vector VR may then be input into a computation unit 18 for analysis, such as to determine whether the differences identified between the predicted and feature vectors Vp, VF indicate that there is a fault in the monitored system. Vectors may be encoded as binary numbers where the number of digits corresponds to the number of classes into which measured values may be classified. Markov chains may be used to model the probability of transitions between states.
|Publication status||Published - 22 May 2019|