English Version


EE论坛(350期):A hierarchical deep neural network for fault diagnosis on Tennessee Eastman process

发布时间:2017-10-27 来源:学生科 705330

讲座时间:11月13日 14:00


讲座题目:A hierarchical deep neural network for fault diagnosis on Tennessee Eastman process

主讲人:Dr. Li Bai


报告简介:With recent advancement in using Deep Learning Neural Network (DLNN) for classification problems, various companies and academic institutions devoted resource and time to develop open source DLNN packages to expand research applications in challenging classification problems. For example, there are numerous facial recognition applications developed by using Tensorflow and Torch created by industry and academic institutions respectively. The classification results are over 95% accuracy which are much superior method compared to the methods using conventional data-driven methods, i.e., Artificial Neural Network (ANN) and Support Vector Machine (SVM). For this research, we need to classify chemical plant process faults. A benchmark system is called Tennessee Eastman Process (TEP) problem was studied in the research. The TEP is a challenge classification problem because the TEP system is a nonlinear time-variant system with 20 different process faults. The faults can be injected in any specific time so that the plant process data will be collected then classified with fault types. To the best of our knowledge in this research, we made the following two main contributions that i) to develop DLNN methods to classify the faults using Tensorflow framework, ii) to design a Hierarchical Deep Neural Network (HDNN) to classify the faults into fault clusters to further improve the classification accuracy. The proposed method has much better performance in terms of i) average True Positive Rate (TPR) and ii) the average classification time delay. The TPR are 96.47%, 93.75%, 81.15%, and 80.30%. for HDNN, DNN, SVM and ANN method respectively.  Also, the average classification time delay is 6.78, 6.88, 12.13 and 12.63 sample time for HDNN, DNN, SVM and ANN methods respectively. The experiment results demonstrated that our proposed method HDNN outperforms SNN, SVM, and DNN. Most importantly, the method has a rapid computation time.


主讲人简Dr. Li Bai is the Chair and Professor in the Electrical and Computer Engineering Department at Temple University. He has extensive research experience and expertise in distributed software computing, robotics, wireless sensor networks, system and software integration using commercial-off-the-shelf products and computer network security. He published over 80 peer-reviewed international journals and conference papers in the related areas. In addition, he was a core organizer in the 8th International Conference on Information Fusion held in Philadelphia in July 2005. He is the Chair of IEEE Philadelphia section in 2007. Currently, he is the chair of IEEE Philadelphia Computer Society and award committee member in IEEE Philadelphia section.