Bottleneck Prediction and Resilience Improvement for Manufacturing Systems

Abstract

Improving production system performance is a constant need and a priority for manufacturers. Planning effective operations to maintain high throughput and resilience during production is not a trivial task. Manufacturers are looking for systematic, reliable, and efficient tools to reach their production goals, in the face of increasing production complexity, unplanned disruptions, and uncertain customer demands. This proposed dissertation aims at developing relevant tools to support real-time throughput and resilience improvement decision-making in complex manufacturing systems. First, a methodology is proposed for detecting dynamic throughput bottlenecks for manufacturing systems with non-serial layout conditions. Existing data-driven bottleneck detection tools are well studied for serial lines. In this study, a systematic framework is proposed to extend a well-known bottleneck detection method to complex manufacturing systems, with loop and parallel structures. The spatial distribution of station blockage time and starvation time is selected as the main indicator for identifying bottlenecks, based on their reflection of production system dynamics. Second, the dissertation investigates the predictive modeling for forecasting throughput bottlenecks. The concepts of factory physics by Professor Wallace Hopp are applied to identify dominant features that collaboratively affect the bottleneck locations in a system. A machine learning algorithm (Long Short-Term Memory, LSTM) is adopted to effectively capture time series dynamics and handle high dimensional inputs. Third, the dissertation proposes a framework for resilience dynamics modeling and control. Disruptions near bottleneck locations can have significant impact on system throughput. A resilient manufacturing system is capable of managing and maintaining throughput in an environment with stochastic disruptions. This research explores how to incorporate the bottleneck detection model with the proposed resilience dynamics model for improving real-time production operations (order dispatching) to achieve higher resilience and throughput during production. The proposed methodologies are developed and validated alongside the support of industrial experts. For the bottleneck detection and prediction models, the effectiveness of the approach is evaluated upon daily industrial data in a major automotive OEM with a span over 1 year. The bottleneck detection tools implemented have made tangible impact at the plant. The resilience dynamics modeling and control is formulated with close cooperation with industrial partners to make sure that the solutions are relevant and applicable to their production scenarios. The simulation platform and the results are closely evaluated by industrial researchers and practitioners. Overall, the dissertation proposed methodologies with both scientific and engineering contributions. The tools developed in this dissertation, if deployed correctly, can provide manufacturers greater capabilities in improving its throughput and overall equipment effectiveness (OEE).