Resarch On SVM-Based Data Reconstruction Algorithm And Its Application In Structural Health Monitoring

Data acquisition and transmission is a very critical link in structural health monitoringsystem. If data loss or noise distortion occurs during the process of data acquisition andtransmission, the reliability of the data collected will decrease, thus jeopardizing the decisionmade in the link of performance evaluation.After analyzing the root causes of unreliable data in structural health monitoring systems,this thesis focuses on two major factors, specifically, data loss and noise distortion, and utilizesleast squares support vector machine to reconstruct unreliable data respectively.For the situation of data loss in structural health monitoring, three methods are proposed toreconstruct the lost part of data. The first one is based on least squares support vector machinestime series reconstruction algorithm, which is relatively simple, using only one channel data torealize the lost part of data; The next one is based on the analysis of correlation of same variableby least squares support vector machine reconstruction algorithm, which makes use of samevariable to reconstruct the lost part of data; The final approach is based on analysis of correlationof different variables by least squares support vector machine reconstruction algorithm, whichuses different variables to reconstruct the lost part of data. For the first method, its advantagesare also its disadvantages, when the sensor fails, the reliability of the reconstructed data willgreatly decrease or even becomes incredible; The second and third methods make full use of theredundancies between channels, with better reconstruction results, both methods have their ownengineering background, and therefore should be selected wisely in practical applications.For the situation of noise distortion in structural health monitoring, two methods areadopted to reconstruct the polluted part of data. The first method applies the relevant theory ofgray-scale correlation to analyze the correlation of channels, to reconstruct the noise-distorteddata from one or more channels that top the correlation sequence. The results of reconstructionby least squares support vector machine are compared with those by Kalman filter, and animprovement can be observed. The second method to reconstruct data is based on the analysis ofgray-scale correlation between different variables, and good reconstruction results are alsoachieved.Lastly, a software system is developed for structural health monitoring based on Ethernetcommunication, using VC++software development tool, together with the Windows built-in MFC (Microsoft Foundation Class). The software system is capable of handling10sets ofmonitoring instruments and80channels at the same time, displaying real-time dynamic responsecurve in the interface window, and preserving historical data.