A statistical framework of data-driven bottleneck identification in manufacturing systems

Data-driven bottleneck identification has received an increasing interest during the recent years. This approach locates the throughput bottleneck of manufacturing systems based on indicators derived from measured machine performance metrics. However, the variability in manufacturing systems may affect the quality of bottleneck indicators, leading to possible inaccurate detection results. This paper presents a statistical framework (SF) to decrease the data-driven detection inaccuracy caused by system variability. Using several statistical tools as building blocks, the proposed SF is able to analyse the logical conditions under which a machine is detected as the bottleneck, and rejects the proposal of bottleneck when no sufficient statistical evidence is collected. A full factorial design experiment is used to study the parameter effects of the SF, and to calibrate the SF. The proposed SF was numerically verified to be effective in decreasing the wrong bottleneck detection rate in serial production lines.