Rare event discovery technologyFeatured Technologies
May 11, 2018
For example, when designing products like automobiles, the product is modeled on a computer in the design phase and repetitively simulated to discover faults. Any fault found after production can result in recall or redesigning, or any other consequence that can similarly incur enormous costs, which makes validation for fault discovery prior to production a very important process.
However, exceptionally rare faults are very difficult to detect―experienced experts repeat simulations in the search for such cases, but this has the problem of taking too much time for the validation. On the other hand, while the demand for reliability increases with advancements in the social systems, the design becomes more complex, giving rise to an increased risk of rare faults being overlooked.
NEC and the National Institute of Advanced Industrial Science and Technology (AIST) are partners in R&D at the NEC-AIST AI Cooperative Research Laboratory (Note 1), established in June 2016. Here, we jointly developed a rare event discovery technology that can find multiple fault conditions quickly by having the AI efficiently repeating simulations while it learns. This new technology can handle events that are difficult to detect due to their extreme rarity caused by numerous simulation conditions and the immense combinations thereof.
Details of this technology
Features of the technology
Shortens design validation time
AI searches for faults by learning the criticality and frequency of faults from the simulation results. In this process, based on the learned results, AI focuses its search on the vicinity of conditions that suffered insufficient validation of faults due to their low frequency while only sporadically searching frequent, oft-validated conditions. As described here, we developed an algorithm that intentionally performs an inhomogeneous search depending on the frequency of occurrence. As a result, this technology can efficiently narrow down occurrence conditions for rare events and shorten validation time.
Reduces the risk of multiple critical events being overlooked
Focusing the search in the vicinity of the occurrence conditions of the event discovered early in the search process can increase the risk of overlooking other faults in case there were multiple critical events. We mathematically derived the optimal condition for reducing the risk of overlooking events and verified that the most efficient ratio of search in the vicinity of the event and search in other areas is 50/50. Based on the results learned from the criticality and frequency of faults, AI calculates and adjusts the concentration ratio of event-vicinity search. This reduces the risk of multiple critical events being overlooked.
This technology can be applied to optical design, structural design of bridges and buildings as well as fluid design of engines. NEC will continue to generate design innovations with this technology together with AIST.
- (Note 1)NEC-AIST AI Cooperative Research Laboratory
This research laboratory was established on June 1, 2016 inside AIST's Artificial Intelligence Research Center (AIRC). Through a continuous flow of development of technologies that integrate simulation and AI spanning from fundamental principles into their industrial applications, the Laboratory aims to establish a new realm of "decision making in unknown circumstances" and contribute to further acceleration of AI research and the industries in joint effort.