HMM-based On-line Fault Diagnosis And Multi-step Ahead Fault Prediction For TE Process

Nowadays, the improvement of science and technology has motivated more and more high level of modernization. And modern control systems are moving towards large-scale, complicated and automated direction. With the increase of scale and complexity of system, the requirement of system security further improves. Fault diagnosis and prognosis are effective methods to improve system reliability and reduce the risk of accidents. If we detect a fault of system and identify its type timely, or predict the development trend of a fault when the system has a small abnormity, we will make an effective maintenance decision accurately and timely to prevent the system fault, then avoid unnecessary loss caused by the development of the fault. For that reason, this thesis studies on-line fault diagnosis and multi-step ahead fault prediction methods based on Hidden Markov Model(HMM), which are applied to Tennessee Eastman(TE) chemical process successfully.On-line fault diagnosis studied in this thesis including two aspects: on-line fault detection and on-line fault identification. Fault detection is an important part of fault diagnosis. The thesis presents a new HMM-based on-line fault detection method. This method regards a new real-time statistic as a quantitative index of on-line fault detection, firstly adopts the principal component analysis approach to take feature extraction of system variables, and uses the variable moving window technology to track dynamic data to get the real-time threshold, then determines whether the system has a fault by comparing the real-time statistic value to the real-time threshold. The on-line fault identification method firstly trains HMM fault model base of all system fault case, then tests the matching degree between system operating data and each fault model to judge the system fault belongs to which category. Then the on-line fault detection and identification methods are applied to TE chemical process for simulation.Finally, the thesis proposes a multi-step ahead prediction method using HMM based on the structure and basic algorithms of HMM combined with the HMM forecasting model. Then the method is applied to TE process to forecast the system health condition of fault 6 after a period of time and predict the developing trend of the fault, which verifies the effectiveness of the multi-step ahead fault prediction method.