Anomaly Detection Using Self-Organizing Maps-Based K-Nearest Neighbor Algorithm

Self-organizing maps have been used extensively for condition-based maintenance, where quantization errors of test data referring to the self-organizing maps of healthy training data have been used as features. Researchers have used minimum quantization error as a health indicator, which is sensitive to noise in the training data. Some other researchers have used the average of the quantization errors as a health indicator, where the best matching units of the trained self-organizing maps are required to be convex. These requirements are not always satisfied. This paper introduces a method that improves self-organizing maps for anomaly detection by addressing these issues. Noise dominated best matching units extracted from the map trained by the healthy training data are removed, and the rest are used as healthy references. For a given test data observation, the k-nearest neighbor algorithm is applied to identify neighbors of the observation that occur in the references. Then the Euclidean distance between the test data observation and the centroid of the neighbors is calculated as a health indicator. Compared with the minimum quantization error, the health indicator extracted by this method is less sensitive to noise, and compared with the average of quantization errors, it does not put limitations on the convexity or distribution of the best matching units. The result was validated using data from experiments on cooling fan bearings.