Remaining Useful Life Prediction Of Precision Bearing Based on Multi-Head Attention Mechanism

Abstract

Abstract The high-precision bearing is the core component of the space actuator, whose remaining useful life (RUL) prediction is an essential issue of constructing condition-based maintenance (CBM) system. To address the current challenges of establishing traditional physical failure models for high-precision bearings and dealing with long time-series vibration data, this paper proposes a precision bearing remaining useful life prediction model based on multi-head attention mechanism. The model utilizes a one-dimensional convolutional neural network (CNN) to compress the input features along the time dimension, then, an encoder-decoder architecture with a multi-head attention mechanism for further feature exploration. Experiments are conducted using vibration signals from the IEEE PHM 2012 Challenge dataset for training and validation. The results indicate that this algorithm exhibits relatively smaller relative errors compared to other algorithms, demonstrating good accuracy and strong stability. This approach shows promise in accurate and reliable precision bearing life prediction.