Few-Shot Lightweight SqueezeNet Architecture for Induction Motor Fault Diagnosis Using Limited Thermal Image Dataset

• Published in: IEEE access 2024, Vol.12, p.50986-50997
• Main Authors: Siraj, Farhan Md, Ayon, Syed Tasnimul Karim, Samad, Md. Abdus, Uddin, Jia, Choi, Kwonhue
• Format: Article
• Language: English
• Subjects: Alternative energy sources; Condition monitoring; Costs; Data models; Datasets; Deep learning; Fault detection; Fault diagnosis; Feature extraction; Few-shot learning; induction generator; Induction motors; Infrared cameras; Infrared imagery; Infrared imaging; infrared thermal imaging; inter-turn faults; Lightweight; limited data; Mathematical models; Model accuracy; Monitoring; Offshore; offshore wind turbine; Parameters; prototypical network; Thermal analysis; Thermal imaging; Wind farms; Wind power; Wind turbines
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3385430

In the realm of renewable energy, wind turbines play a pivotal role in efficiently harnessing wind power, especially in offshore environments, where their significance is amplified. These turbines require vigilant monitoring due to the elevated risk of operational faults. Moreover, vast labeled data is scarce in industrial settings due to the cost associated. Hence, there is a need for fault diagnosis methods that can diagnose precisely with minimal data. This research addresses that problem by proposing an architecture built using prototypical network, few-shot learning, and a modified ultra-lightweight SqueezeNet model specifically made for fault diagnosis. Central to our approach are thermal image datasets captured through infrared (IR) cameras, which enable the detection of subtle temperature variations indicative of faults. The proposed architecture excels in data scarcity. It can swiftly generalize from limited samples, thus reducing the dependence on extensive labeled data and reducing training time. Moreover, the modified model stands out for its highly efficient architecture, featuring 16x lower trainable parameters than SqueezeNet. Despite being ultra-lightweight, our model outperforms the original SqueezeNet by achieving 98% accuracy, 10% higher than the original model, and achieves similar or greater accuracy than other models with significantly more trainable parameters. The proposed architecture achieves optimal computational efficiency while maintaining precise diagnostics. The potential of this technology lies in its ability to be used in real-time fault diagnosis applications on lightweight devices.