The Concept of the Time-Frequency-energy 3D Feature Library and Its Application on Structural Health Monitoring of the EMU Train

In this study, the concept of a time-frequency-energy 3D feature library has been introduced through advanced signal processing technology and pattern recognition methods. Sixty fault vibration signals generated by the axle box bearing of the EMU train are used for analysis. They are firstly pre-processed using discrete wavelet transform for both signal de-noising and singularity detection. Then feature extraction of the signature waveforms standardized by first-order differences and normalization has been performed from three perspectives: time, frequency and energy. Principal component analysis algorithm is then applied for dimensionality reduction in order to achieve integrated features which could better reflect vibration fluctuations in much smaller data sets. And finally the recognition model is established using selforganizing maps network with first principals obtained from the three domains as the input vector. It is found that using this artificial intelligence approach, excellent correlation between unidentified vibration fluctuations and point positions at a timefrequency- energy 3D feature space is achieved. And 100% accuracy rate of intelligent classification for the six vibration categories is realized with both reductions in data processing load and improvement in detection efficiency. The present work contributes to a new idea for the establishment of an intuitionistic feature library to realize real-time discrimination of all kinds of vibration signals generated by the EMU train. It may have a great prospect in the field of rail transit fault diagnosis