Recurrence-Based Techniques for Data Driven Fault Diagnosis and Monitoring in Neutral-Point-Clamped Inverters
de Oliveira, Lincoln Moura, Menaouar Berrehil, El Kattel, dos Santos Filho, Esio Eloi, Vasquez, Juan C.; id_orcid 0000-0001-6332-385X, Guerrero, Josep M.; id_orcid 0000-0001-5236-4592, Marcelo Antunes, Fernando Luiz
Abstract
The high cost and operational complexity of inverter technology make system reliability a key concern, which can be enhanced through advanced monitoring techniques. This paper introduces a novel data-driven condition monitoring (CM) approach for Neutral-Point-Clamped (NPC) inverters based on Recurrence Plot Analysis (RPA) and Recurrence Quantification Analysis (RQA). Unlike conventional deep learning-based approaches, the proposed method requires minimal data, offers high interpretability, and enhances the feature extraction process by capturing nonlinear dynamics and hidden patterns in the time-series data, which are often not accessible through traditional statistical or spectral techniques. The study investigates both normal operation with varying load conditions and open-circuit faults in power switches. A detailed theoretical background on RPA is provided, emphasizing the impact of parameter selection on performance. RQA metrics are extracted from PWM voltage and current <br/>signals and employed as diagnostic features to identify operating conditions and fault states. The proposed approach is experimentally validated using a laboratory prototype, confirming its effectiveness in fault detection and classification. Three operating conditions were evaluated: an RL load, an induction motor, and an open-circuit fault scenario. The experimental results demonstrate that the proposed method achieves high diagnostic accuracy (up to 100%), exhibits strong robustness to noise, and maintains consistent classification performance across varying load conditions. By employing a decision tree classifier applied to RQA metrics, the approach effectively distinguishes between healthy and faulty states, providing an interpretable solution for real-time fault diagnosis. These innovations position the proposed technique as accurate, and hardware-friendly alternative to complex neural-network-based solutions. These findings highlight the method's potential for integration into practical data-driven fault diagnosis frameworks for power electronic converters.