Investigation of deep neural networks for hypoid gear signal classification to identify anomalies

Abstract

A breakthrough of deep learning methods as automated feature extraction techniques for fault further evaluation and classification has blossomed in recent years. Multiple novel approaches of pattern recognition for fault diagnostic algorithms were proposed recently for vibration signal processing. In this paper, deep learning algorithms such as one- and two-dimensional convolutional neural networks (CNN-1D and CNN-2D), long short-term memory (LSTM) and Transformer were developed for hypoid gear faults multi-class and binary classification. The best model for seven gear conditions classification was the CNN-2D with 81.1% accuracy, while fault detection in binary classification achieved 100% accuracy. Also, LSTM and Transformer neural network showed extremely high accuracy result for binary classification. Gear condition without any faults was the most easily classifiable and showed the best statistics of model fit for all the models except for CNN-1D. The investigation revealed that experiments with lower torque achieved better classification accuracy, while different rotation speed had no significant effect. This study showed that superior accuracy results for hypoid gear fault classification were reached by using deep neural networks models and with vibration signal as input information.