Unsupervised Deep Multitask Anomaly Detection With Robust Alarm Strategy for Online Evaluation of Bearing Early Fault Occurrence

Abstract

Early fault detection of rolling bearings under online mode focuses on the evaluation of fault occurrence without system halt and is becoming a new research hotpot. In this problem, however, the monitoring data is generally collected in streaming form, lacking prior information about bearing’s running status to build an effective detection model. Irregular noise interference in normal state may further make the detection model biased. Besides, the model is expected to work with less human involvement. Therefore, it is challenging to improve the detection robustness with streaming monitoring data and to avoid false alarms as well. To address this concern, this paper proposes a new unsupervised robust anomaly detection scheme. First, by introducing the idea of rule adaptation into convolutional neural network architecture, this paper builds a deep multi-task one-class anomaly detection model. This model can learn the task-shared feature representation and then transfer the anomaly detection rule from offline normal data to online unlabeled data. An alternative optimization algorithm is then provided to train this model. Furthermore, considering simultaneously the characteristics of bearing fault degradation and faulty data distribution, this paper proposes a new robust alarm strategy to determine the location of early fault occurrence via integrating permutation entropy and hypersphere-formed detection rule. Experimental results on the IEEE PHM Challenge 2012 bearing dataset and the XJTU-SY bearing dataset demonstrate that the proposed scheme can adaptively and accurately evaluate the occurrence of early fault with much lower false alarm rate. More importantly, the proposed scheme does not need to assume the starting part of online streaming data being in normal state, which is of better deployment and practical significance.