Generative adversarial fault-tolerance for robust multimodal sensor fusion in industrial systems

Modern industrial systems rely heavily on multimodal sensor data to handle random emerging faults. To address this, we propose a novel generative adversarial framework embedded within a multimodal autoencoder architecture for active latent restoration of compromised sensor streams. Our method leverages a two-stage training process: first, a multimodal autoencoder is pre-trained on clean data to learn robust latent representations; then, a generative adversarial network is optimized to map corrupted latent vectors to their correct distributions while preserving cross-modal consistency through a task-specific penalty. Additionally, we adopt the discriminator optimization process to identify anomalies in the underlying data. Therefore, the discriminator is designed to conditionally initiate the generator’s fault-correction mechanism. Experiments on multimodal industrial robotic datasets demonstrate that our approach significantly outperforms conventional methods in reconstructing accurate signals from corrupted inputs, even under several failure scenarios. The proposed system not only enhances fault resilience but also opens new directions for generative models in industrial automation, bridging the gap between robust data fusion and real-time error recovery.