Multi-Scale and Multi-Branch Transformer Network for Remaining Useful Life Prediction in Ion Mill Etching Process

Abstract

Accurate prediction of the remaining useful life (RUL) of an ion mill is vital for optimizing the overall performance of the ion mill etching (IME) process. However, due to the uneven distribution of important information, and the poorly understood failure mechanisms, fault prognosis in this process presents significant challenges. Deep neural networks have shown promising results for extracting, without domain knowledge, relevant features from condition monitoring data. This study proposes a multi-scale and multi-branch Transformer network based on the vanilla Transformer to predict the RUL of ion mills. To extract features on various scales, multi-scale feature extraction first generates receptive fields of various sizes, which are then integrated to obtain feature representations. The multi-branch Transformer uses the parallel attention mechanism and long short-term memory (LSTM) to capture both the adjacent location information and the crucial information of a given timestamp. Handcrafted features are also incorporated as additional input to enhance the prediction accuracy of the model. The proposed model is evaluated on a dataset from a semiconductor IME process. The experimental results demonstrate that the proposed model outperforms other deep neural network and further highlight the practical feasibility of the proposed method.