A Hybrid Temporal-Graph Deep Learning Framework for Intelligent Predictive Maintenance in Large-Scale Data Center Infrastructures

Modern data center infrastructures generate large volumes of heterogeneous operational logs from servers, hypervisors, storage systems, and network devices. While these logs contain valuable early indicators of equipment degradation and potential cascading failures, conventional threshold-based monitoring remains largely reactive and unable to capture complex temporal dynamics and inter-component dependencies. To address these limitations, this paper introduces a Hybrid Temporal–Graph Transformer (HTGT) framework for log analytics and predictive failure detection in large-scale data centers. The proposed architecture combines Transformer-based temporal sequence modeling with Graph Neural Networks (GNNs) to jointly learn long-range temporal dependencies and structural relationships among infrastructure components. The framework is evaluated on large-scale real-world datasets, including the IBM BlueGene/L (BGL) and Hadoop Distributed File System (HDFS) logs. Compared against state-of-the-art methods such as DeepLog, LogBERT, and LogGPT, HTGT consistently achieves superior prediction accuracy, ROC-AUC, and early detection lead time. Further analysis shows that the model yields lower false alarm rates and improved robustness across datasets. Ablation studies confirm the contributions of temporal attention, graph-based dependency modeling, and hybrid fusion to overall performance. The proposed framework offers a scalable, deployable solution for AIOps-driven predictive maintenance, enabling more proactive reliability management and reduced operational risk in modern data center systems.

کلیدواژه ها:

Predictive Maintenance ، Data Center Logs ، AIOps ، Transformer ، Graph Neural Networks ، Anomaly Detection ، Failure Prediction