Sensor fault self-detection based on the mean shift method

Abstract

To accurately identify sensor faults caused by complex environmental conditions and ensure that structural health monitoring systems correctly perceive the structural state, a self-detection method for sensor nodes based on mean shift and sliding window techniques was proposed. The self-detection method comprises two stages, i.e., fault prescreening and fault self-detection. During the fault prescreening stage, the method rapidly identifies potentially abnormal data using the quartile method combined with the sliding window technique, significantly improving the efficiency of the method. During the fault self-detection stage, the method employs the mean shift algorithm to perform adaptive clustering of the abnormal data, effectively detecting various faults. Data from the Canton Tower were used to test the effectiveness of the method by setting four types of sensor faults, i.e., offset, drift, gain, and stuck. Then, the proposed method was compared with extremely randomized trees, random forests, support vector data description, and one-class support vector machines. Results show that the proposed method can detect the four aforementioned faults with high accuracy and computational efficiency.

Keywords

sensor fault detection / mean shift / sliding window / machine learning