<p>Accurate fault diagnosis of rotating machinery is essential for ensuring reliability and safety in modern industrial systems. However, multisensor vibration signals typically exhibit strong spatiotemporal dependencies and nonstationary behavior, which limit the effectiveness of existing deep learning approaches. In particular, most prior methods process spatial and temporal features sequentially, leading to information loss and reduced robustness under time-varying operating conditions. To address these challenges, we propose a Multiscale Bidirectional Spatial-Temporal Gated Recurrent Unit (MB-STGRU) network for multisensor fault diagnosis under nonstationary conditions. The model introduces a spatial-temporal GRU cell that embeds lightweight spatial attention into temporal recurrence, enabling joint learning of inter-sensor correlations and bidirectional temporal dependencies. In addition, a multiscale convolution and adaptive fusion module is designed to capture fault signatures across both short-term and long-term horizons, enhancing robustness to complex operating variations. We evaluate MB-STGRU on two public benchmarks from Southeast University and Paderborn University, where it achieves accuracies of 99.17% and 99.04% on SU1 and SU2, and 98.82% and 98.79% on PU1 and PU2, respectively, outperforming strong baselines such as EGCN by up to 1.4%. All experiments were conducted on an NVIDIA Tesla T4 GPU using PyTorch 2.0.1 and CUDA 11.8, underscoring the importance of GPU-accelerated parallel computation for real-time and scalable industrial fault diagnosis.</p>