A Tool Grouping-Based Approach to Remaining Useful Life Prediction for Predictive Maintenance in Semiconductor Manufacturing

This work examines the use of tool grouping strategies in Fleet Learning and Transfer Learning to enhance predictive maintenance strategies for semiconductor manufacturing tools, particularly in cases of data insufficiency. The challenge of domain dissimilarity between tools, which can lead to negative transfer and unreliable predictions, is addressed through the selection of similar tools for training. By employing various tool grouping strategies, the approach aims to reduce noise and improve the accuracy and stability of remaining useful life predictions. A comparative analysis evaluates the impact of these grouping techniques across different data conditions, highlighting that the effectiveness of tool grouping varies depending on the specific data characteristics and tool conditions. These findings suggest that the success of tool grouping is contingent on factors such as data availability, the uniqueness of tool features, and the degree of similarity between the target and source tools.