Tool condition monitoring and remaining useful life prediction based on multi-modal fusion and transfer learning under variable working conditions

Abstract

Tool remaining useful life (RUL) prediction under various machining conditions constitutes crucial technology in the enhancement of machining quality and production efficiency. With the rapid development of intelligent manufacturing, the RUL prediction approach based on deep learning has been extensively employed due to its high efficacy and precision. Nevertheless, within the existing research, the input of single-modal data presents difficulties in comprehensively representing the tool wear feature information, and the generalization capacity of the model under variable working condition scenarios is limited, thereby constraining the practical application efficacy. The objective of this research is to propose a tool RUL prediction method based on multi-modal fusion transfer learning network with channel adaptive stochastic normalization (MFTLNCASN) to solve the existing problems. In the proposed method, the long short-term memory network (LSTM) is employed to extract the time series features of vibration signals and cutting force signals, thereby accomplishing multi-modal fusion within the feature level. A dual-channel prediction model is established by integrating star network (StarNet) and LSTM. Features are extracted from the fused signals and the surface texture images of the workpiece and then fused at the decision level. The channel adaptive stochastic normalization (CASN) method is devised to dynamically adjust the feature channel normalization strategy, to enhance the generalization ability of the model. Simultaneously, the fine-tuning technique is applied to reduce the disparity between the source domain and the target domain, facilitating high-precision RUL prediction under variable working conditions. Experiments were conducted using a face milling cutter. The effectiveness of the proposed method is verified under both constant and variable working conditions. The experimental outcomes demonstrate that MFTLNCASN exhibits superiority over the existing methods with respect to prediction accuracy and robustness. This research provides a new solution within the domain of tool condition monitoring and has significant practical guiding implications for the enhancement of machining quality and efficiency.