International Journal of Research and Technology in Electrical Industry

International Journal of Research and Technology in Electrical Industry

Transfer Learning-Based Open-Circuit Fault Detection Using Time-Frequency Analysis on Small Datasets for Voltage Source Inverters.

Document Type : Research Article

Authors
Department of Electrical Engineering, Abbaspour School of Engineering, Shahid Beheshti University, Tehran, Iran
Abstract
Open-circuit (OC) fault identification in Voltage-Source Inverters (VSIs) is a critical challenge for the reliability of power systems and motor drives. While Deep Learning (DL) technologies offer automatic feature extraction, they typically suffer from high computational costs and the requirement for massive labeled datasets, which are scarce in real-world industrial scenarios. This paper provides an innovative lightweight diagnostic approach utilizing Transfer Learning (TL) to address current issues of data scarcity and training inefficiency. The pre-trained SqueezeNet model, an efficient Convolutional Neural Network (CNN) with a lower parameter count, is fine-tuned to properly categorize various fault states. In the proposed methodology, the three-phase output current signals are first converted into time-frequency scalograms using the Continuous Wavelet Transform (CWT) to capture rich transient fault features. Subsequently, these visual representations are processed by the network. The proposed method achieves 99.90% accuracy on a small dataset (2000 samples) with considerably reduced training time relative to training deep models from scratch. These findings demonstrate the effectiveness and robustness of the suggested methodology for real-time fault diagnosis in inverters.
Keywords
Subjects

Volume 4, Issue 2
2025
Pages 588-601

  • Receive Date 09 February 2025
  • Revise Date 06 December 2025
  • Accept Date 20 December 2025