CFENet: A contrastive frequency-sensitive learning method for gas-insulated switch-gear fault detection under varying operating conditions using acoustic signals
Published in Engineering Applications of Artificial Intelligence, 2024
Gas-insulated switchgear is an essential protection and control equipment in power systems, with its security and reliability playing a vital role in ensuring safe operation. However, mechanical faults during the operation of gas-insulated switchgear can result in abnormal vibrations, leading to insulation aging, breakdown, and gas leakage. The acoustic-based method has gained popularity for non-contact fault diagnosis in gas-insulated switchgear. However, the diverse acoustic characteristics and negative impacts of high-frequency representations pose challenges for robust mechanical fault detection, especially considering the varying voltages at which gas-insulated switchgear operates. To overcome these challenges, this paper proposed a novel contrastive frequency-sensitive learning method to enhance the robustness of fault detection in gas-insulated switchgear. The proposed method comprises four components: feature encoders, embedding encoders, a classifier, and a new contrastive method. The feature encoders extract valuable representations guided by the contrastive method, while the embedding encoders learn similarities between different features based on the outputs of the feature encoders. The classifier performs fault detection based on the extracted information from the embedding encoders. The effectiveness of the proposed method is validated by experimental data obtained from a real gas-insulated switchgear, achieving an accuracy of 99.65%, which demonstrates competitive performance compared to baseline methods.
Recommended citation: Zhang, Z., Liu, H., Shao, Y., Yang, J., Liu, S., & Yuan, G. (2024). CFENet: a contrastive frequency-sensitive learning method for gas-insulated switch-gear fault detection under varying operating conditions using acoustic signals. Engineering Applications of Artificial Intelligence, 135, 108835.
Download Paper | Download Slides | Download Bibtex