Wavelet Scattering Network Based Bearing Fault Detection

Abstract

Rotating machines are broadly used in the manufacturing industry and they usually need to remain operational for an extended period and in harsh environments, which causes degradation and can eventually lead to failures in components.

Bearings are one the most critical components in rotating machines. Bearing failures result in unplanned downtime and thus impacting the production and maintenance costs. To reduce the costs, monitoring the condition of bearings plays, therefore, a vital role in the maintenance program of the industry. It could allow to move from a time-based preventive maintenance program to a condition-based maintenance (CBM) or  predictive maintenance (PdM) strategy. An essential aspect in bearing fault detection are methodologies that compute and/or calculate condition indicators (features) from noisy vibration signals acquired during early fault stages. These strategies ought to be: (1) data-efficient: one is faced, in many industrial settings, with a lack of training data, a lack of ground truth and scarce information on machine design and operational conditions, (2) power-efficient: it is important that the edge device that is monitoring the health of the machine is power-efficient, as edge devices are often battery-powered, and (3) insensitive to industrial disturbances such as signals originating from other mechanical or electrical components or sensor noise. These are challenges often faced by companies. Providing technological solutions to these issues reduces unexpected downtime and additional costs, including avoiding unnecessary and costly replacement of healthy machine components.

In light of the aforementioned technological challenges, we present the wavelet scattering network application to bearing fault detection, this technique is based on wavelet transform, where it computes translation-invariant features that are locally stable to deformation, making it particularly useful in classification and clustering purposes. The method is experimentally validated with data acquired using the Smart Maintenance Living Lab of Flanders Make. It consists of seven identical drive train sub-systems representing a fleet of machines, on which accelerated life time tests of bearings have been performed. Seventy runs overall have been gathered, twenty-one of which are healthy bearing runs, and the other forty-nine are runs to failure with a small initial indent on the bearing inner race.

Furthermore, the wavelet scattering network feature extraction method is benchmarked to : (1) manually-engineered statistical features such as root mean square, kurtosis, crest factor and peak value, (2) a physics-based bearing fault detection method that is a version of the squared envelope spectrum method of vibration signals extended with several noise reductions and signal enhancement techniques, (3) a convolutional neural network based detection method that autonomously learns useful features and does not require a significant level of knowledge of signal processing techniques and domain expertise..

This benchmark is done on three industrially relevant performance metrics namely : Penalized accuracy, Data efficiency and Power efficiency.