A New Unsupervised Online Early Fault Detection Framework of Rolling Bearings Based on Granular Feature Forecasting

In online scenarios, the monitoring signals are collected in the form of streaming data and would raise some requirements for early fault detection (EFD) of rolling bearings: 1) enhancing the detection accuracy of online data; 2) lowering the computational cost of real-time detection; 3) reducing false alarm rate; 4) deploying easily and working adaptively without manual initialization. To solve this problem, a new unsupervised online EFD framework of rolling bearings is proposed based on granular feature forecasting. First, the proposed framework considers two different online scenarios in extracting granular feature representations of online data. If the offline monitoring data are available, a deep stacked denoising autoencoder (SDAE) network with domain adaptation is introduced to extract common feature representation via decreasing the data distribution differences between offline and online working conditions. If only initial online data are available, a SDAE model is directly used to extract deep features. Second, for the obtained features, a forecasting model with tensor Tucker decomposition and ARIMA is run to predict the degradation trend of all feature sequences quickly and simultaneously. Finally, the deviation degree between the predicted sequence and sequentially-arrived data is calculated for setting alarm threshold. The proposed framework adopts an unsupervised learning mode and has three advantages: 1) flexible applicability to two different online scenarios; 2) automatic detection and easy deployment without manual intervention; 3) high reliability and extremely low false alarm rate. Experimental results on the IEEE PHM Challenge 2012 dataset and XJTU-SY dataset verify the advantages of this proposed framework.