Distributed recognition of patterns in time series data

From a manufacturing plant to a rocket launch pad, to a hospital intensive care unit, arrays of sensors are producing large volumes of information on critical systems. These environments require data analysis to reduce and summarize the data. People often do the data analysis and large amounts of data often require legions of people. Engineering specifications can be relatively difficult to monitor and verify automatically by examining noisy sensor data. Pattern recognition provides a needed layer of abstraction between the raw data and decision making, offering a powerful diagnostic and troubleshooting capability that is relatively easy to develop and apply. By contrast, one could employ mathematical models to verify that the data looks right. However, in real systems such models can be very complex and nonlinear, a model might either take too much time and money to develop or require a depth of understanding that does not exist. This technology still requires someone with knowledge of the pattern to sit down and describe it in detail. It is conceivable for such systems as patterns to include a tool for automatically building the templates in limited cases by looking at a set of training examples, eliminating the need for the human expert, but at this time this is still a research problem.