A feature selection approach based on principal component analysis and synaptic weights for mechanical faults diagnosis

The use of vibration data analysis in faults identification often comes up against the question of definition and choice of pertinent features to feed a diagnosis system. An experimental study has been performed on 162 different operation conditions induced on a mechanical test rig by combination of 4 multilevel fault types. Vibration signals are processed in time and frequency domain to build the feature space for the neural network. Classification rates are satisfying in both the two domains. This paper presents a feature influence analysis in time domain to determine relevant features corresponding to induced faults. This analysis is based on a joint consideration of synaptic weights of the neural network and loading factors. Diagnostic decisions from features judged as relevant gives 95% of success rate while it was only 89% with all the features space. The methodology presented in this paper can be used experimentally to define relevant features from physical vibration measures. The aim is to lead to heuristics about feature selection by faults type.