Industrial process safety has always been a concern for engineers and researchers. Fault diagnosis frameworks based on data-driven methods are prevalent and play a vital role in guaranteeing industrial process safety. However, the data collected in actual industrial production regularly exhibits high-dimensional and complex timing characteristics. In this research, a new framework for fault diagnosis is constructed on the strength of dynamic L2-norm normalized fisher discriminant analysis (FDA) integrating with AdaBoost. Firstly, timing characteristic in industrial process is taken into account so that a dynamic fault dataset is constructed. Then, the FDA vectors are normalized by L2-norm and utilized to reduce data dimension which can learn fault patterns in feature extraction. In addition, an ensemble learning method named AdaBoost is adopted for pattern classification. To verify the effectiveness of the proposed method, simulation experiments based on Tennessee Eastman process are carried out and satisfactory results are obtained.