Sensor Drift Detection Based on Discrete Wavelet Transform and Grey Models

Xiaojia Han,Yue Sun,Ataul Bari,Chao Pei,Aidong Xu,Jing Jiang

doi:10.1109/access.2020.3037117

Abstract

Drift detection has been a difficult problem in the field of sensor fault diagnosis. In this article, a sensor drift detection method using discrete wavelet transform (DWT) and a grey model GM(1,1) is proposed. DWT is used to separate the noise part from the trend part of the sensor data. Then, the GM(1,1) model is used for time series prediction in the trend part. Finally, residuals generated by predicted and current denoised sensor data are calculated and compared with a pre-selected threshold for drift detection. The residuals may not necessarily be Gaussian distribution. Therefore, the pre-selected threshold is chosen by using the kernel density estimation (KDE) method without Gaussian assumption. The effectiveness of the proposed method has been demonstrated using a simulated temperature sensor output from a sensor model on a continuous stirred-tank reactor (CSTR), as well as measurements from a physical temperature sensor in the nuclear power control test facility (NPCTF).

Highlights

S ENSORS are essential parts of technical processes that measure some physical variables in real time
Grey model is chosen for time series prediction in the trend part of the sensor signal in the proposed method
Sensor drift detection has been a difficult problem in the field of sensor fault diagnosis

Summary

Introduction

S ENSORS are essential parts of technical processes that measure some physical variables in real time. Many sensors can be exposed to harsh environments over a long period of time. This may lead to deterioration in some sensing elements and cause the entire sensor to malfunction. Sensor drift refers to the case where the difference between the sensor outputs and the actual value of the process variables diverge linearly with time [4]. Sensor fault detection refers to techniques to locate faulty sensors in a system. Faulty sensors can be identified so that targeted calibration can be performed. This is known as condition-based maintenance [5]

Objectives

Methods

Conclusion