Abstract: With the accelerated deployment of sodium-ion energy storage on the grid and user sides, battery cabins are prone to multiple hidden faults under complex electrochemical-thermal environments and high-power cycling, making it difficult for traditional detection methods to timely identify potential risks. To address this, this paper develops a deep learning-based diagnostic framework tailored to the characteristics of sodium-ion battery cabins. A multi-scale feature enhancement model and an anomaly evolution probability inference method are proposed to enable dynamic capture of weak-varying signals and accurate inference of fault trends. Validation conducted on a sodium-ion energy storage test cabin demonstrates that the proposed method significantly improves early detection capabilities for issues such as sodium plating, gas generation, and abnormal temperature rise, while maintaining high robustness under complex operating conditions and multi-source disturbances.