Knowledge-based seismogram processing by mental images

Abstract

The impact of pictorial knowledge representation is demonstrated for two examples of time series analysis in seismology. The approaches perform a) automated recognition of known event signatures and b) high-resolution onset timing of later phases. Both methods work well under extreme conditions of noise and achieved human-like performance in recognizing known situations. Crucial for this success of pictorial knowledge representation was the design of suitably scaled images. They must be simple and robust enough to transform the complexity of “real life” data into a limited set of patterns. These patterns differ significantly from the initial data; they correspond more closely to the non-linear weighting of recognized impressions by an experienced scientist. Thus the author addresses the pictorial presentations as mental images. For both reported applications, part of their power comes by model-based image modifications. However, this enhancement is far from demanding a complete theory. Any fractional model already enhances the image adaptation, so mental images are best suited to deal with incomplete knowledge like any other artificial intelligence approach. Cognitive plausibility was found for both the non-linear image scalings and the model-based image modifications. In general, the author's method of pictorial knowledge representation conforms to the concept of mental images by Kosslyn. Any new task will demand the composition of new, dedicated image transformations where some generalized design criteria are derived from the author's applications.