Litcius/Paper detail

Physics-inspired time-frequency feature extraction and lightweight neural network for power quality disturbance classification

Zhiwen Hou, Boyu Wang, Jingrui Liu, Yumeng He, Yuxuan Yao

2025Frontiers in Physics23 citationsDOIOpen Access PDF

Abstract

This study proposes a lightweight and efficient classification method for Power Quality Disturbances (PQDs) using the PowerMobileNet model, which combines the S-transform for time-frequency feature extraction and the MobileNetV3-CBAM neural network for enhanced classification performance. Extensive experiments demonstrate that PowerMobileNet achieves a prediction accuracy of 99.33%, significantly surpassing traditional Convolutional Neural Networks (CNNs) at 97.07% and MobileNetV3-SE at 98.58%. Compared to other state-of-the-art models, PowerMobileNet outperforms KELM (97.4%), SqueezeNet (99.0%), ShuffleNet V2 (98.6%), and AlexNet (98.3%) in terms of classification accuracy. Additionally, it exhibits superior robustness under various signal-to-noise ratio (SNR) conditions, maintaining high accuracy even at low SNR levels (e.g., 90% accuracy at 20 dB). The model’s parameter count is drastically reduced to 374,632 (1.43 MB), compared to the traditional CNN’s 112,094,345 (427.61 MB), making it highly suitable for resource-constrained environments. Furthermore, PowerMobileNet demonstrates the shortest runtime, with a training duration of 925 s and a classification time of 0.57 s. These results validate the effectiveness and efficiency of PowerMobileNet for real-time PQD classification, offering significant potential for practical power quality monitoring applications.

Topics & Concepts

Robustness (evolution)Convolutional neural networkFeature extractionComputer sciencePattern recognition (psychology)Artificial intelligenceArtificial neural networkNoise (video)Image (mathematics)ChemistryBiochemistryGenePower Quality and HarmonicsMachine Fault Diagnosis TechniquesPower Transformer Diagnostics and Insulation