Study on the detection of abnormal sounding data based on LS-SVM
- PDF / 286,019 Bytes
- 6 Pages / 595 x 842 pts (A4) Page_size
- 111 Downloads / 166 Views
Study on the detection of abnormal sounding data based on LS-SVM HUANG Xianyuan1,2∗ , ZHAI Guojun2 , SUI Lifen1 , CHAI Hongzhou1,2 1 2
Institute of Surveying and Mapping, Information Engineering University, Zhengzhou 450052, China Naval Institute of Hydrographic Surveying and Charting, Tianjin 300061, China
Received 29 July 2009; accepted 27 May 2010 ©The Chinese Society of Oceanography and Springer-Verlag Berlin Heidelberg 2010
Abstract A new method of detecting abnormal sounding data based on LS-SVM is presented. The theorem proves that the trend surface filter is the especial result of LS-SVM. In order to depict the relationship of trend surface filter and LS-SVM, a contrast is given. The example shows that abnormal sounding data could be detected effectively by LS-SVM when the training samples and kernel function are reasonable. Key words: LS-SVM, trend surface filter, kernel function, abnormal sounding data
1 Introduction The multi-beam sonar is a full coverage sounding system. Multi-beam data always include a small quantity of outliers, due to many factors such as ambient noise, instrument noise and unreasonable setting of multi-beam sonar parameters, etc. (Huang et al., 2001). In order to improve the accuracy of seabed topography, outliers should be removed (Dong and Ren, 2007). To detect the outliers of multi-beam data, trend surface filter is a kind of commonly used method at present. This approach constructs seafloor surface by polynomial function (He et al., 2004;Yang et al., 2004; Zhao, 2002; Lirakis and Bongiovanni, 2000; Bisquay et al., 1998), but in the seabed, topography is a more complex case, trend surface filter is not good any more, even lose valid terrain information. Based on the above, a new method of constructing seafloor surface by LS-SVM is presented. Seabed surface structure problem will be converted to high-dimensional least square problem by LS-SVM, and the model parameters are Lagrange factors. This approach seeks to optimally compromise between the model complexity and learning ability (Chen et al., 2004; Yan et al., 2001). Once the train samples and kernel function are chosen reasonably, then seafloor surface would be constructed properly and outliers could be removed effectively.
2 Choose training samples and kernel function The key of LS-SVM is the appropriate selection of training samples and kernel function. Commonly, the quality of median-beam data is better than the edgebeam (Li et al., 1999). In other words, the edge-beam contains more outliers. In the course of seafloor surface conformation, consider the contribution of edge-beam data, so the training samples are composed of large number of median-beam data and a small amount of edge-beam data. Selecting the correct kernel function is also very important, and different kernel function means different criterion for the similarity of multi-beam data. The commonly used kernel functions include overall and localized kernel function. When the submarine is fairly flat, select the overall kernel function such as polynomial kernel, and the ex
Data Loading...
