Research On Outlier Detection Methods For Process Control System

With the rapid development of technology, there are more and more needs for systems in the field of process control, which usually refer to control precision, response speed, stability and robustness. So the scale of process control systems in the real world has became bigger and bigger, and the complexity has became higher and higher. In addition, most industry processes have the following features:mechanism complexity, nonlinearity, time varying of parameters, big lag, strong coupling and so on, which bring great difficulties to modeling and bring a great challenge for making control strategy at the same time. For this reason, scientists have turned their sights to the process data of process control systems. They try to find the methods of modeling and control strategy based on the process data, and they have made some results, which has improved the position of process data in the field of process control. With the improvement of the process data, however, the quality of process data has been more welcomed than before. Because a set of data with higher quality will bring a accurate base to modeling and strategy making. On the other side, a set of data with lower quality will even result in the failure of control, which will make a terrible result.On this background, after analyzing the characteristics of process control systems and process data, this paper proposed a special method for outlier detection in the process control systems. Main research results are summarized as follows:(1) aiming at the structural characteristics of process control system and the application of process data, this paper gives a special definition for the outlier in process control system, and makes a detection strategy based on this definition, which refers to detection based on model.(2) modeling is a very important step in the detection strategy. This paper proposed that we can use time-series to model the process data according to the characteristics of process data. After analyzing traditional method, this paper proposed dynamic neural network which has the function of association and memory to model the time-series. After simulation, we can find that the dynamic neural network works better than the traditional method at the aspect of veracity and efficiency, which has proved that DNN has a better effect.(3) based on the DNN model, we can get the fitting residual error. This paper proposed thought of wavelet analysis for analyzing this fitting residual error more efficiently. After transforming fitting residual error with wavelet analysis, we can detect the outliers in the process data based on the theory. Also, after simulation we can find that wavelet analysis works efficiently at detecting outliers.(4) because of a problem about the wavelet analysis, which is threshold, this paper connected hidden Markov model with wavelet analysis. As HMM is a statistical model, it can detect the outliers directly after analyzing the wavelet coefficient. This connection solves the problem of threshold efficiently. And the practicality and effectiveness of this method has been proved through simulation experiments.