Fault Detection And Prediction Based On Multi-model Decision Fusion For Air Separation Process

Air separation is widely used in modern industries, such as metallurgy, chemical engineering, petroleum, municipal engineering, medical industry and aerospace. To improve the stability of operation and reduce failures in the air separation process is important for the enterprises to improve the economic efficiency effectively. Fault detection and prediction is critical for reducing accidents in production, improving production performance, reducing economic losses and enhancing the competitiveness of enterprises. Therefore fault detection and prediction system of the air separation process can ensure the stable and efficient production of air separation effectively by detecting the faults in time.In this thesis, the recent research of fault detection and prediction, especially that in air separation field, is reviewed first. Then several commonly used methods of non-parameter model monitoring, including principal component analysis (PCA), dynamic principal component analysis (DPCA), cluster and neural network ensemble are introduced and applied on the data from a real air separation process. The simulation results of the above-mentioned methods demonstrate that these different methods describe different aspects of the operating characteristics of air separation process and have their own advantages and disadvantages in fault alarm rate, false alarm rate and advance time of forecast respectively.In order to enhance the reliability and accuracy of fault detection and prediction, data fusion technology is adopted to combine the decisions of different methods. Multi-model decision fusion method is proposed and three multi-model decision fusion strategies based on the batch estimations and the weighted average, Bayes reasoning, multi-model space transformation are presented. Simulation results on the real data of these three multi-model methods show that process monitoring models based on multi-model decision fusion method can take full advantage of complementarity and redundancy between the various models and thus are more reliable, accurate and efficient than the models based on single non-parameter monitoring method.