Parameterized entropy filter for time series anomaly detection

In view of the time series obtained from mechanical systems, which performs low signal-to-noise ratio and nearly Gaussian distribution, a parameter-adjustable entropy filter is designed for anomaly detection. In order to detect the statistical anomaly caused by changes of mean value and variance, the parameter setting strategies is discussed with a sliding-window based Shannon entropy filter. A monotonic factor K1 is introduced to obtain different smoothing results well the monotonicity of the filter is maintained. In order to adjust the tolerance range of normal signal, a scale factor K2 is introduced. In such a way, the anomaly detection of the time series can be implemented in a variable scaling way. With the entropy filter's assessment criteria which achieved by computing the improvement ratio of overlap ratio of anomaly and normal in the time series after and before being filtered, the rational value ranges of the two factors, in cases of both mean value drifting and variance variation detection, are analyzed by simulated signals. Experimental detection is performed with colored foreign yarn signals of an electronic yarn clearer. The results indicate that such a parameterized entropy filter performs good anomaly detection ability for nearly Gaussian distribution signals. © 2011 Journal of Mechanical Engineering.