Industrial IoT systems generate massive telemetry streams, requiring intelligent predictive maintenance models to detect failures early, reduce downtime, and improve operational reliability and safety. Traditional approaches employ statistical analysis, sequence segmentation techniques, CNN-LSTM hybrids, and graph-based classification models to capture spatial-temporal dependencies and identify abnormal device behaviour patterns. These methods typically achieve high classification accuracy but often exhibit moderate RUL estimation performance, demonstrating strong fault detection capability across industrial, energy, and smart infrastructure applications. However, static graph structures, limited temporal attention, imbalanced fault distributions, and poor generalization under noisy conditions restrict robustness and real-world deployment scalability. This paper proposes a dynamic graph-based GAT-BiLSTM with cross-attention and gated fusion, achieving 99.50% accuracy and superior RUL prediction stability under noisy conditions. The framework incorporates adaptive adjacency learning and multi-task optimization to enhance predictive maintenance accuracy and robustness in IoT sensor networks.