Effective and efficient machine tool maintenance plays a significant role in ensuring manufacturing productivity, product quality, operational safety, and profitability. With the advancement of process sensing, the Internet of Things, data analytics and cloud computing, more manufacturing plants are favoring predictive maintenance over traditional preventive maintenance. Predictive maintenance involves monitoring and predicting machine tool condition and performance. It avoids unnecessary maintenance and prevents catastrophic failure of machine tools, thereby saving operational costs and improving production reliability. However, there are barriers to fully implement predictive maintenance such as insufficient accuracy and reliability of machine tool anomaly or fault detection by existing techniques. This award supports fundamental research on designing a next-generation process sensing-machine learning architecture for capturing manufacturing process dynamics that reveals the underlying dependency of product quality on process settings and machine conditions. This research engages industry in assessing the performance and scalability of this novel machine tool health-monitoring technique at actual manufacturing plants, with the outcomes offering a competitive edge to the U.S. manufacturing sector in the global market. The research involves disciplines such as advanced manufacturing, sensor networks, machine learning, and computing. Knowledge gained is applied to developing manufacturing curricula to equip the next generation of engineers with new skills in manufacturing and data sciences.
This project advances the fundamental understanding of complex manufacturing process dynamics for root cause analysis of process sensing variation and detection of machine anomaly occurrences, through discovering the process-observation causal relationships by an innovative machine learning technique. To improve the trustworthiness and computational efficiency of data-driven analysis and decision making in actual manufacturing plants, a next-generation process sensing-machine learning architecture is designed with capabilities in: 1) high-accuracy modeling and high-efficiency computation upon an optimal architecture without extensive manual tuning; 2) physically interpretable discovery of system input-output causal relationships and process dynamics through the integration of model training with manufacturing domain knowledge; 3) allowing for modeling from unbalanced data and incremental learning from evolving machine conditions without repeated model training for robust and scalable anomaly detection. A physically interpretable convolutional neural network with automated architecture search is developed to correlate process parameters, multivariate sensing data (e.g., force, vibration, power), and part quality specifications (e.g., strength, surface quality) that are acquired from turning and milling processes. The discovered relationships are then diagnosed by an incremental sparse classifier for machine fault detection and classification. The outcomes from this project establish a scientific foundation for the systemic realization of machine tool anomaly detection and predictive maintenance that is not achievable with existing techniques.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
