Spatio-Temporal Attention Gated Transformer for Memory Uncorrectable Error Prediction

Uncorrectable Errors (UEs) in memory are a critical trigger for hardware failures and service disruptions in modern data centers. Existing prediction methods typically rely on rule-based feature engineering or traditional machine learning models, which struggle to effectively capture the complex spatio-temporal evolution patterns of memory errors while ignoring the spatial correlation of memory errors in bit-level physical topology. To address this, this paper proposes an end-to-end deep learning model named Spatio-Temporal Attention Gated Transformer (STAGT). Firstly, the model adopts decoupled positional encoding to explicitly preserve physical topology information. Secondly, it constructs a spatio-temporal decoupled attention architecture, leveraging spatial attention and adjacency masks to accurately capture local correlations between error bits. On this basis, a spatial global gating mechanism is introduced to adaptively adjust feature weights according to the overall spatial state. Finally, temporal attention is integrated to model the progressive evolution patterns of failures. Experimental results on real production environment datasets demonstrate that STAGT outperforms existing mainstream methods in all performance metrics.