Systematic distinction of events and errors in sensor data

Due to neglecting the importance of distinguishing sensor data in event/anomaly detection, similarities and differences among event samples and error samples are analyzed based on the sensor data uncertainty, and a systematic distinction framework is designed to partition the raw data set into event subset, error subset and ordinary subset through node-level temporal processing, neighbor-level spatial processing, cluster-level ranking and network-level decision fusion. Experimental results on real-sensed data show that the framework achieves a distinction ratio as high as 97% in different network cases. Comparisons with traditional methods show that the proposed framework reduces the false-alarm rate to 1/10 of the traditional methods and does not exceed the traditional miss-hit rate.