Open-circuit fault diagnosis of T-type three-level inverters: An intelligent approach based on a deep fusion network

Traditional fault diagnosis methods for T-type three-level (T 2 3L inverters are often hampered by their reliance on manual feature engineering and insufficient noise immunity. To overcome these limitations, this paper proposes CEBANet, a novel deep learning network that integrates a one-dimensional convolutional neural network (1D-CNN) backbone with three key mechanisms: Efficient Channel Attention (ECA), a Bidirectional Long Short-Term Memory (BiLSTM) network, and a final attention module. Methodologically, the network first processes augmented time-series samples of three-phase currents generated via a sliding window and random sampling. The ECA-enhanced CNN backbone extracts critical spatial features, the BiLSTM captures long-range temporal dependencies, and the final attention layer pinpoints the most salient segments for classification, enabling the identification of various open-circuit (OC) fault modes. Extensive experimental results demonstrate that the proposed CEBANet model effectively identifies 19 distinct operating conditions and maintains good robustness under various signal-to-noise ratio (SNR) conditions.