CRACK LOCALIZATION IN HYDRAULIC TURBINE BLADES BASED ON KERNEL INDEPENDENT COMPONENT ANALYSIS AND WAVELET NEURAL NETWORK

Hanling Mao,Hongwei Hu,Xianghong Wang,Zhiyong Zhang

doi:10.1080/18756891.2013.817065

Hanling Mao, Hongwei Hu + Show 2 more

Open Access

https://doi.org/10.1080/18756891.2013.817065

Copy DOI

Abstract

Abstract Hydraulic turbine runner has a complex structure, and traditional location methods can't meet its requirement. This paper describes a source location of cracks in turbine blades by combining kernel independent component analysis (KICA) with wavelet neural network (WNN). The research shows that the location accuracy of WNN combined with KICA feature extraction is the best comparing with the results of WNN and back propagation neural network (BPNN). The method decreases the dimension of input parameters and improves the accuracy of location as well.

Highlights

Almost all of hydraulic turbine blades crack after putting into operation
acoustic emission (AE) source location methods[9,10], such as time of arrival (TOA) location, energy location (EL) and modal analysis location (MAL), are to find spatial locations of crack sources according to the information on sensor location or time of sources occurred
It shows that the result of region recognition using nine feature parameters extracted by kernel independent component analysis (KICA) is the best in the two types of networks

Summary

Introduction

Almost all of hydraulic turbine blades crack after putting into operation. Cracks are seriously dangerous for operating stability and safety of power station. In order to monitor the state of the turbine runner and to locate crack sources using AE technique, it needs multi-channel acquisition because of its large scale and complex structure. AE source location methods[9,10], such as time of arrival (TOA) location, energy location (EL) and modal analysis location (MAL), are to find spatial locations of crack sources according to the information on sensor location or time of sources occurred These methods are unsuitable for the runner with complex structure. The information of ANN is distributed in connection weights, which makes ANN have high fault tolerance and robustness It realizes source location by nonlinearly mapping feature parameters of input signals into recognition space. Input dimensions and output patterns the large number of samples were used in this study

KICA Feature Extraction

WNN Theory and Algorithm

Crack Location in Turbine Blades

Region in the middle of blade

Findings

Conclusions