An investigation of conventional machine learning approaches vs deep learning approaches in manufacturing context : bearing fault detection

Abstract

In the realisation that ball bearing fault is the number one fault that most commonly occur in industrial applications and the potential hazard that it can bring, this paper aims to tackle the problem of bearing fault detection.

With the recent development and boom of Deep-learning approaches in the machine learning space, there is an increasing focus on Smart manufacturing. In traditional machine learning approaches, feature engineering is the most crucial process and bearing fault detection has been heavily reliant on subject matter experts for curating suitable features for prediction. Deep-learning is an alternative method that does not require the feature engineering process. Deep-learning approaches are able to automatically learn features by modelling them as nested layers of abstraction of knowledge from the data itself. Convolutional Neural Network (CNN) is one of such Deep-learning approaches.

In this paper, conventional machine learning approaches are compared to CNN in terms of their performances. For conventional machine learning approaches, the result of Fast Fourier Transformed (FFT) is being used as features for classification. For CNN, we will explore the claim of Deep-learning approaches in its ability to automatically learn features. Raw data will be fed to CNN after converting to a 2-dimensional grey image.

The result ascertained that Deep-learning approaches are able to learn features automatically without the need of domain knowledge expertise. Besides, the result shows that Deep-learning approaches can greatly outperform conventional machine learning approaches.