Advanced AI Framework for Robust Fault Diagnosis in Industrial Systems

The paper proposes a novel advanced AI framework for robust fault diagnosis in industrial systems that experience missing data in sensor measurements. The approach integrates Diffusion Model-based Imputation, Multi-Path Transformer-Graph Neural Network (MPT-GNN), and Uncertainty-Aware Federated Learning (UA-FL) to restore missing sensor readings, enhance fault detection accuracy, and preserve data privacy across distributed industrial environments. The framework combines short-term temporal convolutional networks, Transformers for long-term analysis, and GNNs for inter-sensor connectivity, resulting in improved precision and interpretability of fault diagnosis. Additionally, Bayesian Neural Networks are incorporated for reliable uncertainty estimation, while Elastic Weight Consolidation provides memory-efficient edge device deployment. Experimental results demonstrate fault detection accuracy of up to 98.7% on industrial machinery datasets, minimizing the impact of missing data and facilitating real-time, scalable, and robust deployment of industrial AI systems for predictive maintenance applications.