Nowadays, fault monitoring is no longer limited to simple abnormal alarms, but pursues in-depth diagnosis of the root causes of faults and autonomous decision-making. The working conditions of industrial equipment are complex and changeable. Variations in load and rotational speed can lead to shifts in data distribution, which requires the model to have online self-update capability to avoid performance degradation. Another major technical challenge is how to lightweight complex algorithm models and deploy them to the edge side with limited computing resources to meet real-time requirements (e.g., model inference latency < 50 ms). This paper proposes a Transformer model integrated with a multi-scale attention mechanism (Multi-Scale Attention Transformer, MSAT) for real-time fault monitoring and localization of engineering equipment. The specific innovations include: Designing a dual-channel input module that fuses vibration signals and temperature signals through adaptive weights to enhance cross-modal feature correlation; Proposing a local-global attention mechanism to simultaneously capture local high-frequency vibration features and global long-term temporal dependencies in the Transformer encoder; Introducing a fault localization guidance loss to achieve accurate localization of fault locations through attention weight visualization. Experiments on three public engineering fault datasets (CWRU bearing fault, XJTU-SY gearbox fault, and SEU hydraulic system fault) show that the average fault recognition accuracy of MSAT reaches 98.7%, which is 3.2–5.8% higher than that of existing SOTA methods (such as ResNet and TCN). Additionally, the average advance warning time for early faults is shortened by 1.2 s, and the localization error is controlled within 0.5mm.