Skip to Main Content
Diagnostic ambiguity caused by limited observability of sensors is a significant challenge in real-world diagnostic applications, such as gas turbine engines. Traditional data-driven clustering, classification and fusion techniques based on single fault (class) assumption result in large diagnostic errors. Thus, we solve this problem by diagnosing the inherent ambiguity as multiple faults. The proposed primal-dual optimization framework for classifier fusion improves the correct fault isolation rate, while minimizing the false alarm rate. The key points of primal-dual optimization framework, viz. multiple fault diagnosis and classifier parameter optimization, are extended to the error correcting output code (ECOC)-based weighted voting method and were found to significantly increase correct fault isolation rate compared to the single class assumption at the cost of false alarms. The primal-dual optimization framework also performed better than any traditional fusion technique when it was forced to give a single fault decision; this is due to the fault clustering effect made possible by the dual solution of the multiple fault diagnosis problems.