Adaptive Fault Diagnosis in Wafer Production Lines Fusing Knowledge Graphs and Deep Reinforcement Learning

Abstract

The semiconductor manufacturing industry faces increasing  challenges in maintaining high yield rates due to the extreme complexity of modern wafer production processes. As fabrication technologies scale down to nanometer levels, the interactions between processing tools, sensor readings, and production recipes become highly non-linear and stochastic. Traditional fault diagnosis methods, ranging from statistical process control to standard supervised deep learning, often struggle with the dynamic nature of production environments and the lack of interpretability in diagnostic decisions. This paper proposes a novel framework that fuses Knowledge Graphs with Deep Reinforcement Learning to establish an adaptive, explainable fault diagnosis system. We construct a dynamic knowledge graph that semantically models the topology of the production line, capturing causal relationships between equipment, process parameters, and defect classes. Subsequently, a Deep Reinforcement Learning agent is designed to navigate this graph, learning optimal diagnostic paths to identify root causes efficiently. By treating fault diagnosis as a sequential decision-making process over a structured knowledge base, our approach not only improves diagnostic accuracy but also adapts to concept drift caused by recipe changes or equipment wear.Experimental results demonstrate that the proposed methodsignificantly outperforms baseline approaches in both precision and adaptability, offering a robust solution for Industry 4.0 
semiconductor foundries.