The invention provides a method for predicting cured piece smoke NNK on the basis of robust regression modeling. A model from a physicochemical index item to the smoke NNK is built according to existing cured piece physicochemical data and smoke NNK data, and a cured piece smoke NNK value of an unknown cured piece smoke NNK sample can be directly predicted through physicochemical component data. By means of the method, the steps of rolling, burning, smoke capturing, detection and the like of a traditional chemical mode are omitted. Meanwhile, a robust regression model is adopted so that defects caused by singular values in the physicochemical data or in the smoke data can be effectively avoided, robustness of the model is guaranteed to a large extent, and compared with a common linear regression modeling, robust regression modeling has the advantage of better guaranteeing robustness. As is proved in practice, the cured piece smoke NNK value can be effectively predicted through the model, detection efficiency is greatly improved, and detection cost is lowered.