A novel approach for predicting tool remaining useful life using limited data

Wear, fracture, and other tool faults affect the quality of a machined workpiece and can even damage machine tools. The accurate prediction of remaining useful life (RUL) can prevent a tool from suddenly failing, an ability of significance for ensuring machining quality and providing effective predictive maintenance strategies. Most current approaches for predicting tool RUL are based on historical failure and truncation data. However, for new types of tools or when a similar tool has just launched, such failure and truncation data are limited or even unavailable, making RUL prediction a challenge when using previously proposed methods. To address this problem, a novel method for the prediction of tool RUL using limited data is proposed in this study. A time window is constructed to track the tool condition using sensor data, and its size can be dynamically adjusted according to the wear factor and increase rate. Then, a deep bidirectional long short-term memory (DBiLSTM) neural network in which sequential data are predicted and smoothed by forwards and backwards directions, respectively, is developed to encode temporal information and identify long-term dependencies. On this basis, multi-step ahead rolling predictions are then employed to predict tool RUL. Finally, the effectiveness of the proposed method is verified using the results of milling experiments. These results show that the proposed method is able to predict tool RUL with high accuracy using only limited data. (C) 2020 Elsevier Ltd. All rights reserved.