Stacked Deep Transfer Learning with EfficientNet for Fault Recognition in IoT-Based Power Systems

In recent years, the rapid adoption of Internet of Things (IoT)-based monitoring in smart power systems for identifying electrical faults in real time has become critical for system reliability and safety. However, the existing models struggled to generalize across varying fault patterns and failed to exploit the temporal nature of power signal data. Hence, this research proposes Stacked Deep Transfer Learning with EfficientNet (SDTL-ENet) for efficient fault recognition in IoT-enabled power systems. Initially, the data as collected from the IEEE Power and Energy Society's Real-Time Digital Simulator (IEEE PREDIS) dataset. Further, the data is preprocessed with Z-score normalization and Gramian Angular Field (GAF) transformation to standardize signal ranges and convert time-series signals into temporal-structure-preserving images. After that, the feature extraction is done by using EfficientNetV2B0 for spatial features and a Bidirectional Long Short-Term Memory (Bi-LSTM) for temporal dynamics. Then, the TL is employed by fine-tuning the top layers of EfficientNetV2B0 for capitalizing on what was learned beforehand and adapting to features specific to the faults. Finally, the predictions from the CNN-only models and CNN-Bi-LSTM models are combined as a Multilayer Perceptron (MLP) meta-learner in a stacked ensemble architecture to improve classification robustness. The proposed SDTL-ENet established better results in terms of accuracy (93.75%) than existing LSTM model.