Graph Convolution Network – Long Short Term Memory based Reliable and Intelligent Fault Aware Fog Edge Computing Framework for Resource Constrained IoT Devices

Abstract

The rapid growth of Internet of Things (IoT) applications has increased the demand for predictable, high-performance Edge–Fog–Cloud computing to support seamless data acquisition, transmission, and processing across billions of resource-constrained distributed devices. Edge–Fog-driven IoT architectures, positioned between IoT devices and centralized cloud infrastructure, improve reliability and efficiency by extending network services to the Internet edge and supporting the storage and processing of sensor data. Fog nodes deployed near end users and devices offer mobility support, low latency, bandwidth conservation, reduced energy consumption, and location awareness. Therefore, an efficient computing framework for scheduling fog nodes for edge users and IoT end devices is essential. Existing fog-node scheduling approaches based on metaheuristic techniques and advanced artificial intelligence address load balancing, computational offloading, and service orchestration; however, they often fail to proactively manage dynamic workload fluctuations, network anomalies, and unpredictable node failures. This paper proposes a multi-tier deep learning framework based on a Graph Convolutional Network (GCN) and Long Short-Term Memory (LSTM) model for reliable, fault-aware fog-edge computing. The architecture supports intelligent task delegation and predictive fault management for lightweight edge offloading on resource-constrained IoT devices. The IoT ecosystem, consisting of resource-constrained devices, fog nodes, and end users, is first represented as a graph. In this graph, IoT devices, edge gateways, and fog servers act as nodes, while their relationships, defined by parameters such as bandwidth and latency, act as edges. An adjacency matrix quantifies link weights, and a feature matrix captures each node’s resource characteristics, such as CPU utilization and battery status, at a given time. The proposed model is evaluated in a Python environment using the Google cluster dataset, a benchmark trace-log dataset. Experimental results show that the model effectively detects dynamic workload fluctuations, network anomalies, and unpredictable node failures. Performance analysis confirms that the proposed framework achieves 98.9% accuracy in detecting node failures and workload fluctuations.