MicroForest: Lightweight Bottleneck Prediction for Manufacturing Processes on Edge Devices

As digital transformation in manufacturing accelerates, process bottleneck prediction has emerged as a central task in industrial automation. To streamline manufacturing processes, where diverse tasks interact in complex ways, it is essential to identify in advance both the location and timing of bottleneck occurrences. However, manufacturing environments often lack high-performance computing resources and must rely on cost-effective, resource-constrained embedded devices, making fast and accurate prediction challenging. We present MicroForest, a lightweight decision tree-based model designed to predict multiple process bottlenecks simultaneously under such resource-constrained environments. MicroForest reassembles the high-information-gain nodes from dozens of large random forests into compact forests. Evaluated on a simulation containing up to 150 production tasks, MicroForest achieves 34%p higher recall scores compared to original random forests while shrinking model size by two orders of magnitude and accelerating inference latency by up to $7.2\times$. Compared with other recent work, MicroForest outperforms them with the highest prediction accuracy (F1 = 0.74) and shows a much gentler increase in latency as process complexity grows.