Fault diagnosis in electrical machinery is crucial for industrial reliability and efficient electric power utilization. This study presents a radar-based approach for non-contact motor vibration sensing for fault detection. Unlike conventional techniques, radar sensing eliminates the need for hardware installation and provides a non-contact alternative for vibration monitoring. Instead of considering only spectral-domain inputs, the vibration-related radar signal derived from the phase of the in-phase and quadrature channels is analyzed through multiple representation levels, including time-domain statistical features, Power Spectral Density (PSD), and 2D time–frequency images. To evaluate these representations, conventional machine learning classifiers are applied to time-domain and PSD-based inputs, end-to-end convolutional neural network models are used for direct classification of the 2D images, and a hybrid framework is constructed by coupling deep representations extracted from the 2D images with machine learning classifiers. This integration leverages the high-level abstraction of deep features, combined with the efficiency of machine learning, for robust and accurate classification. Experimental validation was conducted on an induction motor testbed for seven conditions. These are bearing defects, rotor bar breakage, stator winding fault, imbalance, horizontal misalignment, vertical misalignment, and a healthy state. The classification results are based on 659 valid radar recordings acquired under different operating conditions. Representative spectral comparisons of the internal fault datasets further indicated that different internal faults exhibit different frequency-domain organizations relative to the healthy condition, supporting the presence of condition-dependent spectral information beyond a single common motor-specific offset. The results show that the best time-domain and PSD-based baselines achieved 93.02 and 93.78% accuracy, respectively, whereas the proposed hybrid deep feature machine learning framework achieved 98.50% accuracy. These findings support radar as a practical alternative for machine monitoring and demonstrate that time–frequency representations combined with deep feature learning can further improve fault recognition performance, overcoming the limitations of contact-based sensors.