Smart Grid Fault Detection and Classification With Ranked Phase Current Spectrograms Feature Selection and Using Compact Neural Models

Abstract

Accurate fault detection and classification in power transmission systems are critical to maintaining grid reliability and minimizing outage durations. Traditional methods, often reliant on threshold-based algorithms or manual signal analysis, struggle with dynamic system conditions, high noise interference, and varying fault characteristics. This study proposes a robust framework for fault detection and classification in the IEEE-9 bus system by leveraging image-based feature extraction and machine learning. Phase currents are converted into spectrogram images to capture transient fault signatures in a 2D time-frequency domain. Four statistical metrics-Gini decrease, chi-square, information gain, and information gain ratio-are employed to rank and select high-impact features, reducing dimensionality while preserving discriminative patterns. These features are extracted using SqueezeNet, a lightweight convolutional neural network, and subsequently classified via neural networks and logistic regression. The method is rigorously evaluated under diverse scenarios, including variable fault inception angles (0-360 degrees), high-resistance faults (up to 200 Omega), and white Gaussian noise with signal-to-noise ratios (SNRs) ranging from 10 dB to noise-free (infinite SNR). The results demonstrate 99.5-99.7% precision in distinguishing 10 fault types (LG, LL, LLG, LLLG) and normal operation, with a 99.6-99.8% F1-score at 10 dB SNR, outperforming existing artificial intelligence approaches. The system's immunity to noise and adaptability to unexpected fault locations underscore its practicality for real-world deployment. By integrating spectrogram visualization with explainable feature selection, this work fills the gap between signal processing and interpretable machine learning, offering a scalable solution for modern power systems.