Robust multimodal latent repair for fault-tolerant cyber–physical production systems

Multimodal fusion systems in industrial environments are critically vulnerable to sensor faults that destabilize downstream tasks. In capital-intensive sectors such as automotive manufacturing and aerospace, these unmitigated failures translate directly into substantial financial losses due to unplanned downtime and catastrophic equipment damage. To address this, we introduce a mathematically grounded framework for robust multimodal latent repair, which fundamentally re-engineers the Transformer block for fault tolerance. We derive a spectral decomposition proving that residual connections preserve fault locality while attention heads diffuse it. Leveraging this orthogonality, we propose the residual-contrastive attention module. This mechanism utilizes a novel local–global ratio to explicitly detect anomalies and employs an inverse-residual gate to suppress and inpaint corrupted tokens using global context. Validated on rigorous industrial digital twins, our model-based approach achieves state-of-the-art reconstruction fidelity, providing a deterministic architectural guarantee of error containment that effectively decouples physical sensor degradation from operational system stability without relying on implicit regularization. • We derive spectral proofs of fault diffusion and anchoring in Transformer blocks. • A novel Residual-Contrastive Attention module isolates industrial sensor faults. • The Local–Global Ratio enables precise, unsupervised latent anomaly detection. • We achieve state-of-the-art correction across three heterogeneous robotic datasets.