Research On Fault Diagnosis Methods Of Industrial Process Valve Based On Statistical Learning

As one of the important executive components of industrial process control loop,the performance of the regulating valve will directly affect the performance of control loop,therefore it is very necessary to monitor the performance status of regulating valve in industrial processes.However,there are still the following problems in traditional methods of control valve fault detection and diagnosis:(1)The actual industrial production process is complex and changeable,and the data is redundant,it is difficult to directly carry out statistical learning modeling.(2)Traditional extreme learning machine algorithms,are prone to overfitting,and it is difficult to perform reliable data modeling for complex and changeable industrial production processes.(3)Traditional statistical learning methods have a large demand for data samples,and are powerless for real-time monitoring of the state of regulating valve in industrial production process.(4)There is still a gap between the data generated by simulation platform and actual industrial process,and it is difficult to fully verify the actual reliability of invented algorithm.On the basis of reading relevant literature,corresponding solutions are proposed for the above four problems.The main work is as follows:In view of the low efficiency of traditional statistical learning methods in dealing with complex,high-dimensional,nonlinear,and non-Gaussian industrial data,this paper extracts key features of data through manifold learning,and uses local preservation projection method to reduce the data dimension.In this case,the local information and global structure between data are still retained,and the feature extraction effect of the method is verified on the DAMADICS platform,which improves the efficiency of subsequent fault diagnosis of the control valve using the statistical learning method.A locally reserved projection regularization extreme learning machine method is proposed,which combines the structural risk on the basis of the extreme learning machine considering the empirical risk.The proposed method solves the problem that the extreme learning machine method is prone to overfitting,and the data feature extraction improves the model training efficiency.The experiments on the DAMADICS show that the comprehensive diagnosis rate of the method for 19 typical control valve faults exceeds 96%,and the diagnosis rate of 5 faults reaches 100%.A data description method of locally preserved projection support vector is proposed.Data feature extraction improves the efficiency of the model.A hypersphere model is constructed by using normal data.The model achieves real-time fault detection by classifying data at a single moment,and is verified on the DAMADICS platform,and the experiment show that in the diagnosis of 19 typical control valve faults,the method has the highest diagnostic rate on 12 faults,6 of which reach 100%,and the comprehensive diagnosis rate exceeds 96%.Aiming at the problem of the difference between simulation data and actual production process data,two typical control valve faults,the leakage and the stuck valve,are manually set on the Intelligent Process Control Test Platform(IPC-TF)of Wuhan University of Technology.The two proposed algorithms are further verified on IPC-TF,which proves the effectiveness of the proposed algorithms.