Attention-based Explainable Feature Extraction Model for Anomaly Detection in Gearbox

Abstract

Gearbox fault detection is an important task in rotating machinery that has been carried out for the sustainable operation of the industry. Since gearbox is a device for transmitting power, it is composed of designed gears and bearings to minimize heat generation and power loss caused by friction. If faults are not found timely manner, time is wasted due to replacement of defective parts and costs are lost due to repairs, and thereafter, there is also an opportunity cost loss due to downtime. Vibration signals are acquired with piezoelectric sensor and in the form of an amplitude in the time domain or frequency domain to diagnosis the fault. Owing to its non-destructive nature, vibration-based condition monitoring has been widely used to diagnosis mechanical failures. The researchers studied how the distribution of faulty signals differs from the distribution of the signals for a gearbox in healthy condition. However, condition indicator monitoring requires a signal preprocessed and experts in the domain knowledge. On the one hand, machine learning (ML) techniques are also useful at dealing with vibration signals in order to diagnosis gearbox faults. Artificial neural networks, support vector machines, manifold learning and logic regression are examples of typical machine learning approaches. In recent years, deep learning (DL)-based methods are utilized with the advantages of automation of feature extraction and automation of feature selection have been used. DL-based approaches include convolutional neural networks (CNNs), stacked sparse autoencoders (SSAEs), and deep belief networks (DBNs). However, one limitation of such DL approaches is that a large amount of data is needed to extract a sufficiently effective feature from the data. The phenomenon of data imbalance is widespread in real industry since the number of fault data that causes serious problems in machinery operation is extremely small. Our Study focused on developing the DL models dealing with the limited data sample. Our purpose is to detect anomalies with training only normal dataset. Motivated by the previous studies and address related issue such as data imbalance problem in real industry field, we propose a application study of DL-based anomaly detection (AD) in the gear and bearing system.