A data-driven approach to RUL prediction of tools

Abstract

An effective and reliable prediction of the remaining useful life (RUL) of a tool is important to a metal forming process because it can significantly reduce unexpected maintenance, avoid machine shutdowns and increase system stability. This study proposes a new data-driven approach to the RUL prediction for metal forming processes under multiple contact sliding conditions. The data-driven approach took advantage of bidirectional long short-term memory (BLSTM) and convolutional neural networks (CNN). A pre-trained lightweight CNN-based network, WearNet, was re-trained to classify the wear states of workpiece surfaces with a high accuracy, then the classification results were passed into a BLSTM-based regression model as inputs for RUL estimation. The experimental results demonstrated that this approach was able to predict the RUL values with a small error (below 5%) and a low root mean square error (RMSE) (around 1.5), which was more superior and robust than the other state-of-the-art methods.