A Hybrid Temporal Feature for Gear Fault Diagnosis Using the Long Short Term Memory

The vibration of the rotating machinery for condition monitoring in gear fault detection is a popular area of study. Reliable improvements to the rotating machinery can be obtained by enhancing the machine condition monitoring. The automatic detection of a gear fault at an early stage is required to guarantee a reliable and robust rotating machinery system. In this paper, a novel method of gear fault diagnosis is proposed based on extracting a computational cheap hybrid hand-crafted feature set including the Gamma Tone Cepstral Coefficient (GTCC) and the Mel-Frequency Cepstral Coefficient (MFCC), extracted temporally from the vibration signal. The vibration signal faults have a temporal nature, so the Long Short-Term Memory (LSTM) classifier is adopted because it is suitable for time series signals. To evaluate the proposed model, a ten-fold cross validation approach is applied to two different datasets. The results obtained show that the adopted features and the LSTM classifiers are effective for gear fault detection. Additionally, the performance of the fusion of 14 coefficients for both the GTCC and MFCC exceed the state-of-the-art performance for gear fault detection and for those which use learned features using a pre-trained model.