Traditional fault diagnosis methods for Permanent Magnet Synchronous Motors (PMSMs) are limited by insufficient accuracy and poor real-time performance. Distinguishing between demagnetization and eccentricity faults with similar characteristics is particularly challenging, rendering them inadequate to meet modern industry’s stringent requirements for motor reliability. Machine learning technology, leveraging its powerful pattern recognition capabilities, has emerged as a promising solution for motor fault diagnosis. However, most existing machine learning algorithms are primarily applied to offline fault diagnosis and are unable to satisfy real-time demands. To address this limitation, this paper proposes a machine learning-based approach for diagnosing motor demagnetization and eccentricity faults, implemented using a Field Programmable Gate Array (FPGA). Specifically, stator current signals are collected, and Fast Fourier Transform (FFT) is utilized to extract distinctive fault features associated with demagnetization and eccentricity. Subsequently, fault diagnosis algorithms based on Support Vector Machine (SVM) and Random Forest (RF) are developed and their diagnostic performance is evaluated and compared. Finally, an FPGA-accelerated implementation of the selected machine learning algorithm is designed to enable online fault diagnosis. Experimental results demonstrate that the proposed method achieves high diagnostic accuracy and excellent real-time performance, providing a novel solution for the health management of high-reliability motors.