This research aims to propose an innovative smart system we developed for early failure detection of Automatic Tool Change (ATC) systems. Input data is the system's tool magazine door open/close signals. Then, 41 indicators from 26 machines are obtained from statistics-based feature extraction methods. Under the guidance of predefined risk levels, nine high-ranking top level indicators are selected using correlation and regression analysis. In addition, some lightweight supervised learning algorithms are used to build and train the model to solve the classification problem of the system states, such as Normal, Caution, and Danger. The experimental results confirm that the high-ranking indicators can achieve the most prominent and stable performance under a series of tests. Under 10-fold cross-validation, the average accuracy is 89.43 %, which is 19~38 % higher than those of other feature groups. Among them, the Naive Bayes algorithm obtains the best accuracy of 94.2 %. This proves that the proposed smart system can effectively grasp the health status of the ATC systems.