Deep Autoencoder Model for Hypoid Gear Fault Diagnosis With Dynamic Torque and Rotation Speed

In this study the applicability of autoencoder-based deep neural networks to solve fault classification problems using vibration signals. This work will aim to construct CNN and LSTM type autoencoder-based neural networks, which would be suitable for solving fault classification tasks, and compare them with traditional neural networks. This study explores the applicability of autoencoder-based deep neural networks to solve fault classification problems using vibration signals. The study examines 7 different types of gear failures of hypoid gear. CNN-2D, CNN-2D with autoencoder, LSTM, and LSTM with autoencoder were developed and compared. CNN-2D model reached 0.72 of overall accuracy while CNN-2D with autoencoder reached only 0.65. LSTM models (with and without autoencoder) showed worse overall accuracy that CNN-2D. The general conclusion of the research is that in order to achieve better results of autoencoders use for the classification of vibration signal data, it is necessary to perform more tests: to test different data augmentations, architectures of the models.