Hybrid Feature Extraction and Deep Learning Framework for Power Transformer Fault Classification: A Real-World Case Study

Reliable fault diagnosis in power transformers is paramount for ensuring grid stability and safeguarding critical assets. This paper proposes a novel deep learning-based diagnostic framework integrating hybrid feature extraction with advanced classification models. The framework is trained and evaluated on a comprehensive real-world dataset of 3,000 labeled samples, each characterized by 80 features. A dual-stage feature extraction methodology is employed, utilizing Discrete Wavelet Transform (DWT) for multiresolution time–frequency analysis and Matching Pursuit (MP) for adaptive sparse decomposition, followed by Principal Component Analysis (PCA) for dimensionality reduction to mitigate redundancy and enhance class separability. Four state-of-the-art deep learning architectures are comparatively evaluated: a Deep Ensemble of Multilayer Perceptrons (MLPs), a Time Series Transformer, a CNN–LSTM hybrid, and a Conv1D–BiLSTM network. Experimental results demonstrate that the ensemble MLP architecture achieves superior diagnostic performance, attaining peak accuracy of 99.50% coupled with a minimal false alarm rate of 0.33%, and consistently outperforms the comparative models across a comprehensive suite of evaluation metrics. These findings substantiate that ensemble methods and attention-based transformer models offer significant advantages over traditional recurrent-convolutional hybrid architectures for this task. The proposed framework demonstrates substantial potential for enabling robust, high-precision fault classification in real-time transformer condition monitoring systems.