A Lightweight Physics-Guided Feature Fusion Network for Fault Waveform Classification in Three-Phase Inverter Circuits
TL;DR
Experimental results show that MFF-Net achieves stable training, high classification accuracy, and clear feature separation, offering a feasible solution for lightweight online fault diagnosis in power electronic systems.
Abstract
Three-phase inverters are widely used in renewable energy conversion, industrial drives, and energy storage systems. Their IGBTs and other power switching devices often work for long periods under high-frequency and high-current conditions, which makes fault diagnosis an important issue for reliable operation. Traditional diagnosis methods usually depend on manually designed features obtained from Fourier transform, wavelet analysis, or related signal-processing tools. Although these features are interpretable, their performance is closely tied to expert experience. Purely data-driven deep learning models can learn features from raw waveforms, but they often show limited physical interpretability and may overfit when fault samples are insufficient. This paper proposes a lightweight Multi-view Fault Feature Fusion Network (MFF-Net) for fault waveform classification in three-phase inverter circuits. The model contains a lightweight one-dimensional convolutional neural network branch for temporal waveform representation and a physics-guided branch for extracting single-phase statistics, three-phase imbalance indices, zero-sequence components, and multi-band harmonic energy ratios. A gated fusion module is then used to combine the two feature groups according to sample-specific fault characteristics. A simulated dataset with 3,200 samples is built under normal, overcurrent, phase-loss, and IGBT bridge-arm open-circuit conditions. Experimental results show that MFF-Net achieves stable training, high classification accuracy, and clear feature separation, offering a feasible solution for lightweight online fault diagnosis in power electronic systems.