Interpretable Transformer-Based Hardware Health Assessment and Anomaly Detection for Servers in Large-Scale Data Centers

Abstract

Reliable hardware health assessment is important for cloud servers, where latent degradation may gradually evolve into service anomalies. This paper proposes an interpretable Transformer-based framework for multi-sensor hardware health scoring and anomaly detection. Multivariate monitoring data are first normalized and segmented into sliding windows, and then encoded to capture temporal dependencies and cross-sensor correlations. Based on the learned representations, the model outputs both an anomaly score and a 0-100 health score. In addition, an occlusion-based contribution analysis is introduced to quantify the influence of different hardware subsystems on the final prediction. Experiments on 10 cloud servers show that the proposed framework can identify abnormal operating states, capture different behavioral patterns between compute and storage nodes, and reveal the relative importance of network and power-related features. These results suggest that the method is effective for interpretable server hardware health monitoring and is potentially useful for EMI-related anomaly awareness in cloud environments.