Gear fault detection using noise analysis and machine learning algorithm with YAMNet pretrained network

Abstract

Gears are an essential machine element in every machinery. Faults in gears must be detected promptly to reduce downtime and avoid the sudden failure of the machine. Intelligent fault diagnosis techniques are gaining impetus due to Industry 4.0. This work proposes to use the YAMNet network for fault detection of gear pairs using transfer learning. YAMNet is a pre-trained network trained on audio data for sound event detection. The knowledge of the pretrained network is transferred to the fault diagnosis model. Research in this area suggests that noise and vibration data in the time-frequency domain is used to train pre-trained networks for fault detection. However, the research on applying pre-trained audio networks is minimal. An experimental setup is fabricated with two identical gear pair sub-assemblies, one in healthy condition and another with one gear tooth failure condition. Noise data is collected and preprocessed to extract Mel Spectrogram and is given at the input size requirement of the YAMNet pretrained network. The final connected layer is replaced, and hyperparameters are fine-tuned to improve accuracy. The final prediction accuracy is 95% for the random test dataset.