Hierarchical multi-attention neural networks for sensor fault diagnosis and mitigation in digital twins

Digital twin technology has emerged as a quintessential facilitator of industrial digitization within the paradigm of Industry 5.0. Nonetheless, its effectiveness is inherently dependent on the establishment of robust sensor fault management within expansive Internet of Things ecosystems. This paper focuses on the sensor fault diagnosis and mitigation in digital twins. To address this challenge, we propose a hierarchical architecture predicated on the parallel fusion of multi-level attention neural networks, comprising five distinct blocks: the input vector block, the virtual sensor block, the residual calculation block, the fault diagnosis block, and the decision block. In particular, the virtual sensor is realized through a neural-network-based estimator, which integrates a multi-head-attention-based Transformer encoder for modeling long-term dependencies and a globally-attention-optimized bidirectional gated recurrent unit network for spatial feature extraction. The fault diagnosis block is distinguished by a dual-branch convolutional neural network architecture that synthesizes squeeze-and-excitation operations with global attention mechanisms, thereby enabling the extraction of fault-sensitive features from the residuals that emerge between the actual sensor readings and those of the virtual sensor. Ultimately, an autonomous multiplexer system is introduced to dynamically substitute faulty sensors with their virtual counterparts, effectively preventing the propagation of faults. Experimental validation conducted using two publicly accessible datasets demonstrates noteworthy improvements in fault detection accuracy and system resilience.