Companies leverage plenty of monitoring tools collecting performance metrics of Telecom BSS to assure it is in good status. The influence of the system failure is variance, depending on the length of time to finish the system reparation. The metrics collected by monitoring tools may have the indication of the system failure, and the maintainers have chances to foresee a system failure from those metrics. However, the metrics collected by the monitoring tools are too much, and some hints may hide in combinations of multiple metrics. We leverage machine learning approaches to address this problem. We used several machine learning tools and algorithms to explore the configuration of the machine learning models to obtain the model performing the best to our dataset. We compared many algorithms like linear SVM, SVM with RBF kernel, random forest and fully connected neural network. We also introduced an anomaly detection learning technique to see if better performance can be achieved. We found SVM with RBF kernel can achieve the best performance to our dataset, and we conducted a comprehensive grid search of the hyperparameters of the RBF SVM to found the best configuration to our dataset. We achieve F-score 21 in the final explored result and the model can predict 15% of the system failure 60 minutes in advance.