Transfer Learning-Based Feature Extraction Using CWT and 2D CNNs for Fault Diagnosis in PMSMs

Permanent Magnet Synchronous Motors (PMSMs) are widely used in electric vehicles, manufacturing systems, and renewable energy applications, where system reliability is vital for performance and safety. This paper proposes an intelligent condition monitoring framework that integrates Continuous Wavelet Transform (CWT) with deep Transfer Learning using a pretrained GoogLeNet 2D Convolutional Neural Network (2D-CNN) for advanced diagnosis of common PMSM faults. The method converts d-axis and q-axis voltage signals into high-resolution CWT scalograms, enabling rich time-frequency feature representation of motor behaviour under both healthy and faulty conditions. These spectrograms are used to fine-tune only the final layers of GoogLeNet, significantly reducing training time and computational complexity compared to training from scratch. Moreover, since TL requires far less labelled data, it reduces the cost of data acquisition-which is often expensive and time­ consuming in industrial settings. The proposed framework demonstrates strong adaptability across fault types, achieving high classification accuracy with 100%, 98.6%, 96.6%, and 100% for healthy, partial demagnetisation, static rotor eccentricity, and inter-turn short circuit (ITSC) respectively. By combining pretrained CNNs with wavelet-based voltage imaging, this approach delivers an efficient, cost-effective, and scalable solution for PMSM condition monitoring in complex operational environments.