To address the challenges in synchronous motor fault diagnosis—including difficulties in modeling multi-channel temporal signals, inadequate extraction of local dynamic features, and poor generalization under small-sample and high-noise conditions—this paper proposes a Vision Transformer framework based on Multi-Scale Feature Aggregation (MSFA-ViT). The method incorporates a lightweight channel attention module to adaptively recalibrate input signals, enhancing the representation of inter-channel correlations. A multi-branch dilated convolution module is introduced to capture local fault patterns across multiple temporal scales. Furthermore, an efficient patch embedding mechanism is designed to compress long sequences and extract discriminative features. By integrating lightweight attention with multi-scale feedforward networks, the model effectively combines global dependencies with fine-grained local structures, enabling end-to-end feature learning. Experimental results on a public dataset of synchronous motor electrical faults show that MSFA-ViT achieves superior diagnostic accuracy and F1 scores compared to state-of-the-art methods, while demonstrating consistent performance under signal occlusion, noise interference, and baseline drift. Even in small-sample settings, the model maintains strong robustness and discriminative power. This work provides a lightweight, robust, and practical solution for intelligent online monitoring of synchronous motors.