Task Dependency and Uncertainty-Aware Multi-Task Learning for Machining Process Monitoring

Machining process monitoring (MPM) is crucial for predictive maintenance and quality control within smart manufacturing. Additionally, monitoring multiple elements with limited sensors is essential for real-world applications, making it suitable for a multi-task learning (MTL) framework. MTL is an active research area that improves predictive performance across multiple outputs by extracting shared informative features for several tasks within a single model. However, the performance of MTL models is highly dependent on the weights of multiple loss functions, and determining the optimal weights is challenging due to the varying difficulty of each task. Moreover, conventional MTL models cannot consider the relationships between outputs, leading to inaccuracy and unrealistic predictions due to the loss of useful inter-task information. To address these issues, this study proposes a novel architecture incorporating task uncertainty for optimal weighting and task dependency to better capture inter-task relationships. Experiments are conducted on real-world side end-milling datasets using a computer numerical control (CNC) machine, utilizing vibration data alone to predict cutting forces, tool wear, and surface roughness. The proposed method enables accurate and efficient real-time quality control in smart manufacturing.