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

Area of Science:

• Manufacturing Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Accurate remaining useful life (RUL) prediction for cutting tools is crucial for efficient manufacturing, preventing downtime and costs.
• Current deep learning methods for RUL prediction are hindered by limited labeled data, which is expensive and time-consuming to acquire.
• Vast amounts of unlabeled machining data are often underutilized in practical applications.

Purpose of the Study:

• To propose a novel method for enhancing tool RUL prediction by effectively utilizing abundant unlabeled data.
• To address the limitations of data scarcity in supervised learning for RUL prediction.
• To improve the accuracy and generalization capabilities of RUL prediction models.

Main Methods:

• Developed a custom criterion and loss function to train models on unlabeled data, incorporating the physical rule of tool wear progression.
• Employed transfer learning to transfer knowledge learned from unlabeled data to a model trained on labeled data.
• Utilized a graph neural network (GNN) for multi-sensor data fusion to extract richer information.

Main Results:

• The proposed method effectively utilizes unlabeled data to enhance tool RUL prediction.
• Transfer learning successfully integrated knowledge from unlabeled data into the RUL prediction model.
• GNN-based multi-sensor fusion improved the effectiveness of unlabeled data enhancement.

Conclusions:

• The unlabeled-data-enhanced approach significantly improves the accuracy and generalization of cutting tool RUL prediction models.
• This method offers a viable solution for leveraging underutilized unlabeled data in industrial settings.
• The integration of GNNs further boosts the performance by enabling sophisticated data fusion.