Abstract
Smart manufacturing transforms conventional production systems into data-rich, connected, and intelligent environments capable of sensing operating conditions, interpreting system behavior, making decisions, and adapting to changes in real time. However, the availability of production data and artificial intelligence alone does not make a manufacturing system smart. Manufacturing intelligence requires systematic methods that convert real-time data into interpretable knowledge of the physical system and use this knowledge to guide prediction, optimization, and control.
This dissertation develops data-enabled analytical modeling and intelligent decision-making methods to improve safety in smart manufacturing. Safety is investigated through a progressive, multilevel research path: from the safety and continuity of the overall production system, to the safety of human–robot collaborative processes, and finally to the quality and safety of an individual product. At each level, analytical models are established from manufacturing logic, physical interactions, system structure, and operational data. Properties derived from these models are then used to formulate decision problems and guide the implementation of optimization, machine-learning, and reinforcement-learning methods.
The first part of the dissertation addresses system-level safety, with an emphasis on production continuity, throughput, and production-loss reduction in serial production lines. A serial-line model is first developed for systems experiencing multiple machine downtime events while only a limited number of technicians are available for corrective maintenance. Because the effect of a failure depends on machine location, buffer states, interactions among simultaneous failures, and the evolving condition of the line, maintenance tasks cannot be prioritized effectively by considering each failed machine independently. An online task-allocation method uses real-time system information to assign limited maintenance resources with the objective of minimizing total production loss.
The system-level analysis is subsequently extended to serial production lines containing quality inspection stations and rework loop-back structures. In addition to normal machine downtime events, the model captures quality events in which products fail inspection and are redirected through a rework branch. Analytical system properties are derived to distinguish and evaluate the impacts of equipment downtime, product-quality failures, and rework activities on production throughput. Together, these studies use data-enabled modeling and intelligent decision-making to protect system output and improve operational resilience.
The next part of the dissertation advances from the production-system level to process-level safety in human–robot collaborative manufacturing. This work is motivated by an initial robot-teaching study in which a robot learns a task trajectory from demonstration and adapts the learned motion when the starting or ending position of the task changes. Building on this foundation, a safety-field-based framework is proposed to guide a robot working with a human in a shared workspace. A deep Q-network is trained offline using representative human-motion trajectories, robot states, workspace configurations, and stationary obstacles. During operation, real-time sensing data are used to update the interaction state, and the learned policy controls the robot while maintaining a prescribed safety distance from the human, avoiding workspace obstacles, and continuing the assigned task.
The final part of the dissertation moves from system and process safety to the quality and safety of a single product. A pharmaceutical secondary packaging line with aggregation operations is considered as a quality-critical application. Because aggregation, buffering, and downtime events cause individual products to experience different processing histories, system-level indicators such as throughput and work-in-process cannot fully represent the conditions experienced by a specific product. Product residence time is therefore introduced as an individual-product and process-level quality indicator. An event-based analytical method determines whether a downtime event increases the residence time of a specific product and identifies the maximum duration of the event that can occur without increasing its residence time.
Collectively, this dissertation presents a step-by-step pathway toward safer and smarter manufacturing. It begins with data-enabled modeling of complete production systems to diagnose and minimize production loss, progresses to AI-enabled control of human–robot collaborative processes, and concludes with event-based monitoring of the quality-related experience of an individual pharmaceutical product. Through this multilevel progression, real-time data, analytical modeling, derived system knowledge, optimization, machine learning, and reinforcement learning are integrated to support manufacturing systems that are observable, interpretable, adaptive, intelligent, and safety-conscious.
