Traffic incidents detected in real time reduce both human injuries and damage to property while allowing better traffic control and more informed decision-making. Traffic incident data imbalance strongly reduces the effectiveness of detection methods. We introduce a new detection system called FA-WRF which combines factor analysis with weighted random forest to identify traffic incidents. We use traffic flow changes to create our first incident variables. We use factor analysis to transform the initial variables into smaller dimensions. We use Bootstrap version 3 to define data selection rules for our training set. The evaluation method of classification trees assigns MCC weight values to determine the optimal weights for each decision tree. The model produces superior results by weighting strong decision trees to optimize decision quality for unbalanced data problems. We evaluate the system performance by examining detection and false alarm rates alongside classification performance and AUC of ROC curves. Experimental results have shown that the FA-WRF-based model produces better classification results and has high competitiveness when it comes to processing unbalanced data classification in comparison with other methods such as KNN, Random Forest, XGBoost, Extra Trees Classifier, Stacking Classifier, and LGBM.