Feature Extraction Using S-Transform and 2DNMF for Diesel Engine Faults Classification

This paper investigates the supervised classification of a distribution fault of an internal combustion Diesel engine using vibration measurement. For 3 inlet valve clearance values, the standard S-transform is used to produce a time-frequency representations of the vibration signals. The large size of time frequency images is then reduced to a set of lower sizes using two-dimensional non-negative matrix factorization . A multilayers perceptron neural network is then trained and applied to classify the test data. The optimal size of feature set is computed, for the best classification and the lowest elapsed CPU time at the training and testing classification phases. It has been found that the performance of the multilayers perceptron neural network classifier is, generally, enhanced and the CPU time is minimized for a reduced feature set size.