Unsupervised Deep Tensor Multitask Anomaly Detection With Rule Adaptation for Online Early Fault Evaluation

This article tries to solve the problem of online early fault detection (EFD) with unlabeled streaming data by addressing the following challenges: 1) no state information is available for online data; 2) online working condition is not determined in advance; and 3) false alarm should be avoided. This article first proposes a deep tensor multitask anomaly detection (Tensor-MAD) model with rule adaptation for the online EFD. Running on multitask learning architecture, Tensor-MAD builds a new tensorized pooling filter to keep the essential information of each task’s operating status from noisy data. With hypersphere-based one-class detection rule representation, Tensor-MAD further constructs a new rule adaptation mechanism to transfer the detection rule from offline labeled data to the online unlabeled data. A training algorithm with an alternating minimization scheme is also provided to update tensor decomposition and rule adaptation. Then, the optimal information filter level and the rule adaptation degree can be determined. Based on the obtained anomalies, this article proposes a nonparametric alarm threshold setting method based on the sequential accumulation of anomaly probability. This threshold can be adaptively chosen once an expected false alarm rate (FAR) is given. A rationality proof is also provided. Experimental results on the IEEE PHM Challenge 2012 bearing dataset demonstrate that the proposed approach can adaptively and accurately evaluate the early fault occurrence from unlabeled streaming data. More importantly, the proposed approach has a much lower FAR and a faster convergence speed, providing an easy-to-deploy and reliable solution for online EFD.

I. Introduction

Running under complex working conditions such as high rotating speed and heavy load, critical elements, such as gearbox, bearings, and other transmission parts, will easily encounter damage and crack that raise the potential risk for the whole machine and even cause major accidents. The requirement of higher reliability and safety for the critical elements is growing significantly [1]. Therefore, it is important to conduct early fault detection (EFD) for the elements in time. Due to the fast development of computer science and sensor technology, various machine learning methods have been applied to intelligent EFD, including shallow models, such as support vector machine (SVM) [2] and Gaussian process [3], and deep models, such as convolutional neural network (CNN) [4] and long short-term memory (LSTM) [5]. At present, building a suitable detection model according to the characteristics of specific applications is still a key challenge of intelligent EFD.