Support vector machines (SVMs) are used for condition trend prediction of mechanical equipment. The influence of cost functions, kernel functions and parameters on prediction performance of SVMs is studied. Cost functions play a very important role on prediction performance of SVMs. Experiments show that the prediction performance of e insensitive cost function is superior to that of least square cost function. At the same time, analysis of the experimental results shows that four kernel functions have very close prediction performance in short term prediction, and radial basis function kernel has better prediction performance than other kernels in long term prediction. In comparison with traditional Back Propagation (BP) neural network, Radial Basis Function (RBF) network and Generalized Regression Neural Network (GRNN), experiments show that SVMs, which implement the structure risk minimization principle, obtain the best prediction performance.
[1]
Johan A. K. Suykens,et al.
Financial time series prediction using least squares support vector machines within the evidence framework
,
2001,
IEEE Trans. Neural Networks.
[2]
Johan A. K. Suykens,et al.
Weighted least squares support vector machines: robustness and sparse approximation
,
2002,
Neurocomputing.
[3]
Lutgarde M. C. Buydens,et al.
Using support vector machines for time series prediction
,
2003
.
[4]
F. Tay,et al.
Application of support vector machines in financial time series forecasting
,
2001
.
[5]
Vladimir Vapnik,et al.
Statistical learning theory
,
1998
.
[6]
Tae Yoon Kim,et al.
Artificial neural networks for non-stationary time series
,
2004,
Neurocomputing.