A Predictive Data-Driven Framework for Multi-Line Manufacturing Throughput Analysis
Abstract
: Understanding and predicting throughput time in multi-line manufacturing environments is a core challenge in industrial simulation and production planning. This paper proposes a simulation-informed analytical framework applied to a real-world event-log dataset comprising 28,026 parts produced across 13 heterogeneous lines over 13 operating days. The framework addresses three objectives: i) characterising per-line throughput distributions, ii) quantifying the impact of equipment downtime on cycle time, and iii) forecasting shift-level production using pre-shift features. Downtime is significantly associated with increased cycle times ( p = 0 . 020), although correlation patterns vary across lines. Change-point detection (PELT) identifies intra-shift disruptions in 11.2% of shifts, typically occurring in the second half, suggesting cumulative degradation effects. A consistent time-of-day effect is observed across most lines. For prediction, global models outperform per-line approaches due to data sparsity. Under a rolling-window protocol, Ridge Regression achieves R 2 = 0 . 570 (MAE = 36 . 3 parts/shift). Feature importance analysis indicates that recent production history dominates predictive performance.