Unsupervised Gear Monitoring using Deep Convolutional Auto-Encoders and K-Means : Application to Gotix Dataset

Abstract

Vibration-based diagnostic approaches using deep learning have attracted attention of the academia and industry. Nevertheless, most of these methods are supervised diagnostic approaches that require a large amount of labeled training data and different working conditions in addition to being time consuming. Therefore, this paper proposes an unsupervised diagnostic model by integrating a deep convolutional Auto-Encoder with a clustering algorithm, and aims to evaluate the potential of unsupervised and blind learning techniques in the context of unlabeled data for gear fault detection. The proposed method does not need statistical feature extraction, and directly uses the normalized frequency-domain signals as inputs. It has been validated with a gear wear fault dataset.