Data-driven bearing fault detection using hybrid autoencoder-LSTM deep learning approach

Abstract

Artificial intelligence (AI) and its sub-domains of machine learning and deep learning have kindled the interests of both industry practitioners and academicians. Its contribution to the manufacturing industry in making intelligent predictions about a machinery's health and its working has seen a huge surge in the research carried in recent years. Nowadays, AI in manufacturing is popularly applied for the efficient fault detection of machinery using data analytics. Traditional fault predictive classification and further diagnosis have pitfalls such as low prediction accuracy, poor feature extraction and susceptibility to noise. To overcome these disadvantages, this paper proposes the deep-learning-based hybrid autoencoders (AE) - long-short-term memory (LSTM) framework for fault detection. The dimensionality reduction with automatic latent feature extraction by autoencoders and temporal feature consideration by LSTM help to achieve high fault diagnosis accuracy. The empirical results show that fault detection of roll bearings based on the proposed hybrid AE-LSTM deep learning technique achieved superior results in comparison to the traditional K-means clustering technique.