A product manufacturing process starts from parts processing, then assembles parts into components, and forms the product finally. To obtain an expected product quality, the quality characteristics in the part level, component level, and product level must be controlled, and the deviations between their actual and target values are required to keep in specified tolerances. All these tolerances form a product tolerance system, and the quality characteristics of product levels contain the geometric and the non-geometric parameters, which are interrelated and form a complex system. General tolerance is the total amount the actual parameters are permitted to vary, which not only include the geometric parameters in machining, but also includes physical, chemical, electrical, and other parameters. What most concerns the product users is whether product quality characteristics meet their requirements, rather than a component quality characteristic or part quality characteristic, and the product quality characteristics are generally not only geometric quantities but also include many non-geometric quantities. The product tolerance system optimization design from part level to product level cannot be achieved, as it includes geometric and non-geometric quantities. In this paper, a product tolerance system model is developed on the basis of determining the quality characteristics of product levels, and the information of the product levels quality characteristics is excavated from data recorded in the product testing process using data mining methods through the support vector nonlinear regression relational model between parts quality characteristics deviations and product quality characteristics deviations. Then, the product manufacturing cost model is set up, which includes the machining dimensional tolerances and non-geometric tolerances, and the product tolerance system optimization model is developed by minimizing the product manufacturing costs as the objective function and the quality characteristics tolerances of product levels as constraint conditions. Finally, a micro-motor product is used as an example to optimize its tolerance system, and its manufacturing costs are decreased by 13.14 %. The results show that the developed method is effective and provides a new way for the product tolerance system optimization.