Deep Multi-task SVDD: A New Robust Online Detection Method of Bearings Early Fault

With the rapid development of sensing technology in recent years, early fault detection (EFD) of rolling bearings in non-stop online scenarios has received widespread attentions. Specifically, the online scenarios raised the following challenges: 1) The initial data of online stage is generally insufficient, making the extracted features less representative and discriminative; 2) Due to running-in, lubrication and other reasons, some irregular data fluctuations frequently appear in normal state, which may cause false alarms; 3) Alarm strategy should be as simple as possible to meet the requirements of online status monitoring. To address such challenges, this paper proposes a new EFD method, named deep multi-task support vector data description SVDD, to achieve robust online detection of early bearing faults. First, regarding the fault detection of single bearing as a one-class anomaly detection task, the proposed method adopts SVDD to build the model of each task, and adds a hypersphere-based similarity loss between all tasks. By training a deep neural network while optimizing a data-enclosing hypersphere of all tasks, the proposed method can extract common features of multiple bearings normal state data. Moreover, the features of sequentially-arrived online data are calculated using the obtained feature representation, and a hypersphere of the online data, named online hypersphere, can be determined accordingly. An anomalous score is then calculated to determine whether an early fault occurs by measuring the distance of one sample beyond the online hypersphere, while the online hypersphere is also updated iteratively. Experimental results on the IEEE PHM Challenge 2012 bearing data set show that the proposed method is robust to the fluctuation of online data, and can be effectively recognize the status change of online data with a lower false alarm rate.