Inferring network-level alarm influence pathways for cellular infrastructure reliability

Large-scale telecommunication infrastructure, such as Radio Access Networks (RANs), is a critical component of the modern digital industry. To ensure infrastructure intelligence and operational reliability, it is essential to understand how alarms influence one another across base stations. Existing alarm analysis approaches primarily focus on isolated temporal patterns at individual sites, overlooking cross-site alarm influence mechanisms at the infrastructure level. To address this gap, this paper proposes a Co-Occurrence and Time-Aware Graph Attention Network (COTA-GAT). It infers alarm influence pathways without relying on explicit physical or logical topology information. COTA-GAT constructs a directed influence graph from alarm co-occurrence patterns and incorporates time-decay weighting to represent both the recurrence intensity and temporal proximity of cross-site alarm interactions. Two time-aware graph attention layers are then used to learn delay-sensitive node embeddings by jointly considering node-level features and time-decay-weighted edges. Subsequently, a link-prediction decoder estimates potential alarm influence pathways. Additionally, a controlled Network Simulator-3 Long Term Evolution RAN simulation with independently generated fault-cascade ground truth is used to validate the true propagation edges beyond chance Experiments conducted on a large-scale real-world telecommunication alarm dataset demonstrate that COTA-GAT outperforms most graph-based baselines and provides an interpretable topology-agnostic influence graph. Considering a day-long propagation window, COTA-GAT achieves area under the receiver operating characteristic curve and average precision values of up to 0.9109 and 0.9235, respectively. The proposed framework provides a novel data-driven foundation for proactive reliability management in large-scale communication networks, even in the absence of information about the underlying network topology.