Abstracting Failure Case Database Information for Detecting Failure Mode

Abstract

Industrial accidents continue to happen despite rapid technological advancement and they are often caused by triggers similar to those of past accidents. If we turn our eyes to the world, especially to the emerging industrial players, we hear news about accidents caused by phenomena that have already caused similar accidents elsewhere. Industries, as they emerge and grow over hundreds of years, learn their lessons throughout their histories and build rules, regulations, and common knowledge to avoid accidents. Each industry is probably well aware of accidents that took place in its own country, especially when the accident led to enforcement of a new law. Nevertheless, we hardly have any knowledge of accidents in foreign countries unless they were of huge sizes. Japan had a national project of building a database of knowledge and lessons learned from past accidents. Failure Knowledge Database (FKDB) went on the Web in 2005. As of today it still attracts a large number of readers with its over 1,600 failure cases. Our research is targeted at making use of this FKDB by abstracting the knowledge, especially what triggered the accidents, and comparing the knowledge with functional and structural elements used in new designs. Design Record Graph (DRG) is a graphical representation of the designer’s intension starting from the left with the product functional requirement which iteratively divides into sub-functions to reach a set of functional elements (FE). Each FE maps to a structural element (SE). Then the SEs iteratively combine to form assemblies and finally the product at the right end. A failure starts from one of the FE-SE pairs and propagates the DRG in both left and right directions to reach the two ends. The propagation leaves a trace of how the point of failure led to disabling the product. For each failure case in FKDB, we identified the origin of failure, the FE-SE pair that started the accident. An FE is abstracted by a verb phrase and a set of noun phrases, and similarly an SE with some noun phrases. By limiting the phrases to use, similar concepts are described by the same abstracted phrases. A new design has a number of FE-SE pairs and their propagations in the DRG to reach the two ends. The designer can then compare all propagations in the design, without the knowledge if any of them are dangerous, with those in FKDB that are known to have led to accidents. We developed quantitative operators to evaluate the similarity between two traces. Our results offer a way of warning the designer about possible flaws in a new design similar with causes of past accidents that the designer has no idea about. Our method of preventing design failure can apply to other fields for novice planners in avoiding failure while still in the planning stage. We can further develop the use of knowledge into overseas countries by mapping the limited number of verb and noun phrases into foreign language.