As the reliability test data of complicated mechanical products is rare in quantity on the system-level and difficult to determine the accurate composition of the life distribution unit as well, the traditional reliability evaluation method based on evolutionary theory has been of little use. And the Statistical Learning Theory begins to be widely focused on as a novel small sample statistic method, which has been mostly applied to pattern recognition, fault detection, time series prediction and so on. This paper creates a new method for reliability evaluation derived from Statistical Learning Theory. By constructing Support Vector Machine with analog reasoning, and solving linear operator equation, the probability density of product can be evaluated directly and then the product reliability index can be obtained. Compared with the traditional way, this method can apparently increase the accuracy and generalization ability of reliability evaluation within limited samples. Finally, this paper presents the bridge of a certain heavy special vehicle as an example to testify the efficiency of this method, and uses the accelerated life test of the vehicle bridge to estimate its reliability.