|
RAP Seminar Series
A Novel Approach to Outlier Detection: by
Andrew Weekley
Wednesday, 4 April 2001
|
|
The analysis of times series data has a fundamental role in science. The most commonly used methods make certain assumptions about the times series, e.g. local stationarity, or that the data represents a specific model with a superimposed random signal having a fixed statistical distribution. In practice these assumptions are often violated. The random signal may change in time or could be represented by more than one distribution. For example, in the case of signal interference or in certain failure modes, there could be more than one distribution. Quality control and outlier detection algorithms must diagnose these situations and find the “correct” signal in spite of additional spurious signals. In addition, it is desired that a quality control, or confidence, index is available which indicates how reliable each data point is. IODA (Intelligent Outlier Detection Algorithm) is an algorithm that has been developed at the National Center for Atmospheric Research (NCAR) to perform quality control on time series data. IODA is unique in that image processing techniques are used to detect outliers in the time series. This algorithm, however, has wider applications than simply one-dimensional time series data. For instance a similar methodology could be used on multi dimension time and spatial series data. This seminar will introduce the fundamental concepts behind IODA, and compare the performance of IODA with several other outlier detection/mitigation techniques. |
|
|