Unlabeled-Data-Enhanced Tool Remaining Useful Life Prediction Based on Graph Neural Network

Remaining useful life (RUL) prediction of cutting tools plays an important role in modern manufacturing because it provides the criterion used in decisions to replace worn cutting tools just in time so that machining deficiency and unnecessary costs will be restrained. However, the performance of existing deep learning algorithms is limited due to the smaller quantity and low quality of labeled training datasets, because it is costly and time-consuming to build such datasets. A large amount of unlabeled data in practical machining processes is underutilized. To solve this issue, an unlabeled-data-enhanced tool RUL prediction method is proposed to make full use of the abundant accessible unlabeled data. This paper proposes a novel and effective method for utilizing unlabeled data. This paper defines a custom criterion and loss function to train on unlabeled data, thereby utilizing the valuable information contained in these unlabeled data for tool RUL prediction. The physical rule that tool wear increases with the increasing number of cuts is employed to learn knowledge crucial for tool RUL prediction from unlabeled data. Model parameters trained on unlabeled data contain this knowledge. This paper then transfers the parameters through transfer learning to another model based on labeled data for tool RUL prediction, thus completing unlabeled data enhancement. Since multiple sensors are frequently used to simultaneously collect cutting data, this paper uses a graph neural network (GNN) for multi-sensor data fusion, extracting more useful information from the data to improve unlabeled data enhancement. Through multiple sets of comparative experiments and validation, the proposed method effectively enhances the accuracy and generalization capability of the RUL prediction model for cutting tools by utilizing unlabeled data.