The judgment of screw locking results is the key point of intelligent screw locking machine. In order to better distinguish the results of screw locking, aimed at the characteristics of unequal length and category imbalance of screw locking data, and the problem of misjudgment of similar categories, a discrimination model was established for the screw locking data in two-stages based on random forest. In the first stage, the characteristics are constructed according to the physical characteristics of the original data. Moreover, the data is under-sampling and the random forest algorithm is used for feature selection. In the second stage, first, the probabilistic principal component analysis variance of each physical property is used as the feature to cluster, the similar categories are grouped in the same cluster, and then the random forest algorithm is used on each cluster to establish a classification model. Finally, the result of screw locking is distinguished by deter mining the cluster of data first and then classifying by classifier in the cluster. The experimental results show that compared with traditional screw locking method and classic machine learning classification algorithms, the model has better accuracy, recall rate and F value.