Identification of Pipeline Intrusion Signals Based on ICEEMDAN-FE-AIT and F-ELM in the uwDAS System

Abstract

Aiming at the problem of low signal identification accuracy due to various noises in pipeline intrusion signals collected by the ultraweak fiber grating distributed acoustic sensors (uwDAS) system, we propose an identification method on top of a new denoising approach for pipeline intrusion signals in this article. The denoising approach uses improved complete ensemble empirical mode decomposition with adaptive noise, fuzzy entropy, and adaptive interval thresholding (ICEEMDAN-FE-AIT). The fisher score feature selection and extreme learning machine (F-ELM) are combined to identify the intrusion signals. We build a data acquisition platform in the laboratory to collect the pipeline intrusion signals, including chainsaw, mechanical vibration, excavator digging, artificial digging, and no-intrusion. Experiments show that the ratio of noise signal to noise reduction ( ${R}_{\text {DNSN}}$ ) of ICEEMDAN-FE-AIT is better than those of four other denoising methods, namely, variational mode decomposition and permutation entropy (VMD-PE) method; the ICEEMDAN-PE-AIT method; the ICEEMDAN, energy density and average periodicity, and AIT (ICEEMDAN-ET-AIT) method; and the ICEEMDAN-FE and wavelet soft threshold denoising (ICEEMDAN-FE-WSTD) method. The values of ${R}_{\text {DNSN}}$ for the five signals are 15.2043, 16.7654, 14.9815, 15.5541, and 13.5428 dB, respectively. The average identification accuracy is 93.27%, in subsequent identification experiments using F-ELM.