Lightweight transformer-based generative adversarial network for acoustic anomaly detection in converter valves

Area of Science:

• Electrical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Acoustic analysis for anomaly detection in high-voltage direct current (HVDC) converter valves is crucial but faces challenges.
• Infrequency of anomalies, computational limitations, and class imbalance hinder practical deployment of current methods.

Purpose of the Study:

• To propose a novel unsupervised anomaly detection framework, the lightweight transformer-based generative adversarial network (LT-GAN), for HVDC converter valves.
• To address data imbalance and resource constraints in acoustic anomaly detection.

Main Methods:

• Developed LT-GAN incorporating MobileNet V2 and D-MobileNet V2 for Mel-spectrogram processing.
• Integrated a K-ViT block to improve global representation learning and reduce network parameters.
• Utilized unsupervised learning on real-world acoustic datasets.

Main Results:

• LT-GAN achieved an Area Under the Curve (AUC) of 0.9806, significantly outperforming baseline methods.
• Achieved high partial AUCs (p-AUCs) in low false-positive regions (0.9295, 0.9122, 0.9559).
• Demonstrated exceptional lightweight characteristics: 9.626M parameters, 0.506 GFLOPs, 37.48 MB model size.

Conclusions:

• The proposed LT-GAN effectively detects anomalies in HVDC converter valves using acoustic data.
• LT-GAN offers a resource-efficient solution for practical deployment, balancing high detection performance with low computational overhead.