To address the challenges of insufficient feature learning for tail classes and imbalanced gradient competition between head and tail classes in long-tailed motor fault diagnosis, this paper proposes a spatiotemporal message passing neural network with a cosine decay learning rate and warm restarts strategy (STMPNN-CDLR). First, a multi-node topological structure is constructed to model the spatiotemporal dependencies among sensors at different sub-time windows within the same global time window of multivariate time-series data. A dynamic spatiotemporal relation weighting matrix is designed to capture the temporal evolution patterns of sensor features, thereby enhancing the representation of potential spatiotemporal interactions between head and tail classes. Second, the message passing mechanism of the STMPNN is utilized to perform feature updates across sub-time windows, improving the model’s ability to perceive both local and global information. Finally, the CDLR strategy periodically restarts and decays the learning rate to mitigate gradient competition imbalance caused by long-tailed distributions, enhancing model stability and sensitivity to tail classes. Experimental results on four motor fault diagnosis tasks with varying long-tailed ratios show that the proposed method achieves superior diagnostic performance and stability for tail classes without compromising the accuracy of the head (normal) class, with overall accuracy exceeding 94.57%. These results verify the effectiveness and superiority of the proposed approach in addressing long-tailed motor fault diagnosis problems.