In modern manufacturing, the efficiency and accuracy of quality inspection significantly influence production costs and product quality. With the development of sensing technology, signals of process variables can be collected in high resolution, which can be regarded as multichannel profile data. They have abundant information to characterize the multistage manufacturing process and help with quality inspection tasks. Motivated by the urge to take the place of manual inspection, we target at modeling based on multichannel profile data for data-driven quality inspection. However, quality inspection modeling based on multichannel profile data is challenging due to the complexity of data structures and manufacturing processes. Specifically, in the pipe tightening process, accurate diagnos...[ Read more ]

In modern manufacturing, the efficiency and accuracy of quality inspection significantly influence production costs and product quality. With the development of sensing technology, signals of process variables can be collected in high resolution, which can be regarded as multichannel profile data. They have abundant information to characterize the multistage manufacturing process and help with quality inspection tasks. Motivated by the urge to take the place of manual inspection, we target at modeling based on multichannel profile data for data-driven quality inspection. However, quality inspection modeling based on multichannel profile data is challenging due to the complexity of data structures and manufacturing processes. Specifically, in the pipe tightening process, accurate diagnosis of defect types is required, but the available samples for each defect type are limited and imbalanced. Moreover, the profile data is incomplete since the pre-tightening process before the pipe tightening process is unobserved. In the ceramic firing process, the identification of crucial phases and stages is required. Nevertheless, the process data streams from different stages are unsynchronized and high-dimensional, and a hierarchical structure exists between phases and stages. To tackle the challenges, we propose models based on functional data analysis, machine learning, and deep learning. For the pipe tightening process, we propose an innovative classification framework to train on imbalanced datasets based on deep metric learning. A neural network and padding mechanism specially crafted for processing profile data are also proposed to handle incomplete profile data. For the ceramic firing process, we propose a real-time diagnostic system, facilitating the real-time synchronization of unsynchronized and high-dimensional process data streams. We also develop a hierarchical sparse partial functional linear regression model (HSPFLR) and corresponding parameter estimation algorithm, enabling the simultaneous identification of crucial phases and stages. The effectiveness of our frameworks is demonstrated through simulation studies and real-world case studies. The proposed methods can replace traditional manual inspection methods, improving product quality and reducing inspection costs.