11 / 2025-04-11 09:43:03

Multi-Sensor Information Fusion via Selective Spatio-Temporal State-Space Modeling for Machinery Diagnostics

Deep learning,Selective state spaces,Graph neural network,Intelligent fault diagnosis,Information fusion,Shapely additive explanations

Deep learning approaches have revolutionized fault diagnosis through their capacity to extract complex patterns from multi-sensor data. Traditional neural architectures like CNNs and RNNs process Euclidean data but often struggle to capture intricate relationships within multi-sensor systems. Graph Neural Networks (GNNs) offer advantages by modeling these relationships through topological structures, yet they typically focus on spatial interactions while overlooking temporal evolution essential for fault progression analysis. This paper introduces a novel Selective Spatio-Temporal Graph (SSTG) framework that integrates Selective State-Space Models with graph-based architectures to treat spatio-temporal graphs as unified dynamic systems. Our approach captures both temporal evolution and structural dependencies simultaneously, guiding fault feature selection through graph information. We develop a dual-path Bayesian fusion mechanism that adaptively weighs uncertainty from different information sources, enhancing diagnostic robustness. Experiments across multiple benchmark datasets demonstrate SSTG significantly outperforms existing methods in both accuracy and interpretability, particularly under imbalanced data conditions. Through SHAP-based analysis, we provide transparent explanations of model decisions, revealing how different sensor regions contribute to diagnostic outcomes. This work establishes new possibilities for advanced fault diagnosis by effectively combining graph-structured modeling with selective state-space techniques.