Estimating remaining useful life of rotating machinery using relevance vector machine and deep learning network

Abstract

Listen

Translate article

In today’s manufacturing industry, unexpected equipment failure caused by mechanical degradation usually leads to serious losses in availability, reliability, quality, and cost. To handle the issues of fault diagnosis and remaining useful life (RUL) prediction for rotating machinery, this contribution presents a multi-technique fused method for health prognostics using relevance vector machine (RVM) and deep separable convolutional gated recurrent network (DSCGRN). The proposed data-driven method gets rid of handcrafted feature extraction and considers multi-sensor data and their temporal dependencies at different degradation states. Specifically, the method is implemented in two phases: anomaly detection and RUL prediction. In the first phase, a RVM classifier, which distinguishes the state of mechanical degradation as no-obvious-fault-tendency and fault-prone, provides a fault threshold for the full-life-cycle degradation progressions and enables the method to concentrate more on fault-prone stage. In the second phase, a regression estimator (i.e., DSCGRN) is shaped to learn deep-layer feature representation from the raw multi-sensor data, thus accurate RUL estimation can be carried out, especially for the fault-prone condition. The proposed method is validated through the raw multi-sensor data of bearings accelerated degradation tests. It turns out that the proposed method performs superiority in RUL prediction compared with some existing prognostic approaches.