전처리 되지 않은 진동 신호로 플랭크 웨어 예측하는 1D-CNN, 2D-CNN과 LSTM로 구성된 인공지능 네트워크

Abstract

Turning processing machines have been widely employed due to their precision and versatility. As the number of cycles increases, the performance of these devices generally degrades owing to tool wear. Therefore, real-time tool condition monitoring (TCM) that utilizes statistical or machine learning methods has gained significant attention in both academia and industry. However, these methods necessitate sufficient data pre-processing, requiring a high degree of academic understanding as well as significant amount of time. Therefore, this research proposes an advanced artificial intelligence network to monitor a wide range of tools by utilizing raw signals without pre-processing. This study first developed a method consisting of 1D and 2D multi filters convolution neural networks (CNNs) and stacked long short term memories (LSTM). To activate the LSTM in a stable manner, the CNN plays a crucial role in dimensionality reduction. Accordingly, two dimensionality reduction approaches were proposed. These were layer normalized 1D&2D-CNN Multi filters. Then, following multi filters, the stacked LSTM was used to extract the sequential features. Next, the performance of the proposed network using the NASA milling dataset was observed and compared between the 1D/2D-CNN without Flank wear information, pre-processing, and previous research network inclusion. Consequently, although the 1D-CNN method did not have them, it achieved a similar level of accuracy as the present method using past Flank wear input.