Multimodal Zero-Shot Reasoning Framework for Real-Time Edge Intelligence in Safety-Critical Systems

Safety-critical systems increasingly require the transfer of intelligent reasoning to edge devices that have limited resources where the most significant aspects are real-time response, reliability, and robustness. The paper introduces a multimodal zero-shot reasoning framework for edge intelligence to operate in real-time under the constraints of safety-critical areas. The different modalities and methods often need task-specific training data or cloud inference but the presented work is completely contrary as it utilizes unseen scenarios while performing under strict computational and latency constraints. Fast and reliable inference at the edge is enabled by the combination of lightweight multimodal feature extraction, semantic alignment, and efficient reasoning methods. Safety-aware design principles are incorporated into the framework through fault-tolerant processing, confidence-aware decision handling, and graceful degradation strategies which assist in dealing with sensor failures and uncertain predictions. The architecture is purposely built to have a trivial memory footprint and requirements for processing power, thus it is the perfect candidate for embedded-edge platforms. A large number of experiments show that the proposed approach has a similar level of reasoning accuracy while at the same time cutting down the inference latency and computational cost significantly compared to the state-of-the-art models in multimodal and zero-shot edge intelligence. These results confirm the framework’s usefulness in facilitating robust real-time intelligent decision-making for safety-critical edge systems under various operational conditions and application domains that include autonomous, industrial, and cyber-physical systems in a global context.