A Seq2Seq transformation strategy for generalizing a pre-trained model in anomaly detection of rolling element bearings

Abstract

The monitoring of the health of Rolling Element Bearings (REBs) in the rolling mill process was recently automated through signal processing and machine learning technologies. However, large data sets are required, and existing methods require precise labels. Moreover, the scarcity of equipment anomalies challenges their practical identification. In addition, machine learning approaches require human resources to manage machine learning models because all REBs in the production line are not identical. Thus, we propose herein a Seq2Seq Transformation Strategy inspired by the accurate performance of such the methods in Natural Language Processing (NLP) and Computer Visioning (CV). Our strategy aims to use an anomaly detection model trained on one signal to detect anomalies in other signals, reducing the need for model management in practical scenarios. Specifically, we considered time series vibration signals collected from the REBs to be analogous to linguistic data, which are sequential and time variant. In essence, our strategy generates synthetic data from a vibration signal represented as the target domain by using a transformer. Subsequently, a model trained by another signal called the source domain is employed to detect the conditions of the target domain. The goal of transformation is to minimize the differences between the distributions of the synthetic and source data. We assumed that the target conditions could be identified from the synthetic data by the source model if the detection model trained on the source domain was precise. Our strategy was experimentally evaluated on six large, unlabeled data sets collected over 1 year from a rolling mill production line in a steel factory. The strategy significantly improved the accuracy of anomaly detection, which had initially been unacceptable when the source model was directly applied to the detection of the target conditions.