Robustness of Deep Neural Networks in Engineering Systems：Frontiers and Outlook

Deep neural networks （DNNs） are catalyzing a profound paradigm revolution in traditional engineering systems. However，a sharp contradiction exists between the inherent vulnerability of mainstream DNN models and the extreme robustness requirements of engineering systems. This “robustness gap” has become one of the key bottlenecks restricting the widespread application of artificial intelligence in engineering systems. Currently，there is no comprehensive survey specifically addressing the robustness of DNNs in engineering systems. This survey aims to systematically introduce and analyze important research on the mechanisms of DNN robustness and methods for improving it within engineering systems. First，at the problem-discovery level，this survey deconstructs the connotation and extension of the robustness gap issue. Subsequently，at the root-cause analysis level，it introduces theoretical advancements in analyzing the causes of robustness defects in DNNs and the relationship between robustness and multi-scale network architectures. Furthermore，at the level of existing countermeasures，this survey explores general methods for enhancing the robustness of DNNs in engineering systems，as well as specialized methods for improving DNN robustness tailored to the characteristics of key engineering fields such as industrial manufacturing，power grid systems，and autonomous driving. Finally，from a future outlook perspective，this survey focuses on cutting-edge directions，including novel intrinsically robust DNN architectures，new learning paradigms，and new strategies for embedding engineering semantic constraints. It seeks to provide a valuable reference for constructing robust DNNs for engineering systems.