Reducing the Training Samples for Damage Detection of Existing Buildings through Self-Space Approximation Techniques