Vibration Anomaly Detection Using Multivariate Time Series

Abstract

The paper presents a set of deep learning algorithms for detecting vibration anomalies in bearings using multivariate time series on datasets provided by Case Western Reserve University. The study considers a problem of multiclassification of the condition of the bearings depending on the type of defect, but also on the degree of defect, considering only punctual defects in an incipient phase. Once the data sets are correctly labeled and the algorithms are trained on this data, they can accurately predict the type and the size of defect. The model with the best results in the set is RNN - CNN (Recurrent Neural Network with Convolutions) giving an accuracy greater than 97% in all (load) cases.