Real-Time Anomaly Detection in Solar Cell Production Line Sensor Streams Using Hybrid Physics-Informed and Machine-Learning Approaches

Abstract

Real-time anomaly detection across the dense sensor instrumentation of modern solar cell production lines requires reconciling the explainability demands of process engineering with the pattern-recognition capabilities of contemporary machine-learning methods. This work presents a hybrid framework combining physics-informed feature engineering with gradient-boosted decision-tree ensembles and contrasts it with end-to-end deep-learning approaches on identical sensor-stream benchmarks. The hybrid approach yields anomaly-detection performance comparable to deep-learning baselines while preserving feature-attribution interpretability that supports root-cause investigation by process engineers. We discuss deployment patterns including shadow-mode validation, alert-rate calibration, and integration with existing manufacturing execution systems for the closure of the detect-investigate-respond loop.