Smart system monitoring for high dimensional multistage manufacturing processes

Abstract

This research studies a system-wide approach to monitor product quality in real time to avoid manufacturing defects in high dimensional multistage processes. Traditional control charts have been widely used in various manufacturing industries due to their simplicity. However, in today’s complex manufacturing processes, these charts are not efficient anymore. A complex manufacturing process may include multiple stages with sensors embedded throughout the processes that generate a huge amount of data in high dimensions. Since the numbers of stages and parameters are usually very large, traditional control charts are incapable of handling a multistage high-dimensional problem mainly due to the problem of false alarm rates of simultaneous monitoring. Industry 4.0 and Internet of things (IOT) provides opportunities to achieve better quality products toward a zero defect system. Currently, data is either thrown away or stored in unused databases. In a inefficient approach called “fire-fighting”, when there is a decline in the quality, process engineers go back to archived process data to figure out the problem. However, due to various reasons such as messy and unclean data, outdated data, and common manufacturing data features such as complexity and dimensionality issues, this process may take a long time. In the best cases, researchers provide classification-based process monitoring techniques to use the manufacturing data. However, the state-of-the-art classification-based process monitoring techniques usually provide quality predictions at the end of the manufacturing process and provide no chance to fix the problem. In addition, knowing high dimensional, unbalanced, and newly released manufacturing data, the literature is largely silent on providing a comprehensive study addressing those issues. While addressing above mentioned challenges, the proposed research delivers a stage-wise process monitoring which provides plenty of time for engineers to fix the process before the last point. Then, based on the results from the predictive models, adjustable process parameters can be altered to avoid potential defects. The proposed research relies on predictive models which are built on a series of classifiers. The proposed research is implemented in two different manufacturing application – additive manufacturing (AM) and semiconductor production. The proposed research in the AM industry called the Multi-Layer Classification Process Monitoring (MLCPM) is applied in the Laser Powder Bed Fusion (LPBF) metal printing process. In the semiconductor manufacturing industry, we applied the proposed method on a very imbalanced high dimensional production data called SECOM (SEmiCOnductor Manufacturing) which is publicly available through the UCI repository lab. In this study, we examined various classification models and sampling approaches to find the best results in terms of specific metrics chosen regarding the imbalanced nature of the problem. The applied case studies show the effectiveness of the proposed framework in terms of accurately predict the real-time production quality state. The chance of predicting the quality of the process before the last step provides chances to reduce the waste and save cost and time in the production systems.