Study on Open-Circuit Fault Diagnosis of Three-Level Inverter Based on AO-DKELM

In response to the challenges posed by the frequent loss of local time-frequency features in diode-clamped (NPC) three-level inverter fault diagnosis and the inadequate prediction accuracy of fault classifiers, this study proposes a novel fault diagnosis method that integrates spectral difference with a deep kernel-extreme learning machine (AO-DKELM). This method employs ensemble empirical modal decomposition (EEMD) on phase voltage signals under fault conditions to derive the marginal spectrum via Hilbert transform. Subsequently, it applies time-frequency spectral transformation to terminal voltage signals (the voltage between the three-phase line and the reference point <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{O})$</tex> to construct spectral difference eigenvectors. The spectral difference serves as the input for pre-training the Extreme Learning Machine Auto Encoder (ELM-AE). Building upon this foundation, the semi-supervised Deep Kernel Extreme Learning Machine (DKELM) fault classification prediction network model, optimized using the Skyhawk algorithm, is employed for diagnosing open-circuit faults in NPC three-level inverters. The results indicate that the proposed method not only effectively addresses the loss of significant components in the hidden fault features extracted by traditional algorithms, but also markedly enhances the accuracy of fault diagnosis.