Optimal active unsupervised fault detection in cascaded h-bridge inverters based on machine learning

Abstract

Multi-Level Inverters (MLIs) are commonly used in high-voltage, high-power industrial applications. In this regard, their reliability, and health optimal performance are in the first priority. However, as the number of switches in a multilevel inverter increases, it comes so common to occur faults within the system. Ensuring the reliability of MLI is an important concern in power industries, making effective fault detection methods essential. Developing precise physics-based, model-based, and hardware-based models for fault detection is challenging, largely due to unknown parameters and incomplete understanding of the physical processes within the system. At this end, the proposed paper presents a highly efficient hyper-tuned machine learning (ML) model known as Isolation Forest (IF). This algorithm is an unsupervised method used for anomaly detection, which isolates outliers by recursively partitioning data points, as an effective way for identifying faults or rare events in large datasets with minimal computational complexity of the MLI system. To test this algorithm, a 17-level Cascaded H-Bridge (CHB) inverter is simulated with several faults, the proposed IF model is tested. In the next phase, the proposed model compared to the others, based on the performance indicators of F1-Score, Precision, Recall, and Accuracy, which the highest results retained for IF to have an accurate unsupervised fault detection model, that smoothens the way for a fully automated, and self-healing industrial application system.