Fault Diagnosis Method Research Of The Control Valve Based On Data Driven

The process control is one of the most important branch during the development of industrial automatization. The control valve is the most important terminal part of the process control system. And, it is also the main source of the breakdown in the process industry. In the modern complex process system, a fail of one of the thousands of control valve could result in the stop of the whole production line or even a catastrophic accident.Preventive maintenances are conducted regularly to ensure the safety and reliability of control valve. However, such actions have resulted in "excess maintenance", which cause the high maintenance cost, low reliability and failure rate rising. On the other hand, as the Distributed Control System (DCS) and computer technology are widely used in process control, a large amount of data is collected and stored, which contains a wealth of information about the system’s status. Unfortunately, these data has not been effectively used and there is the phenomenon which is called "data is rich, information is poor". Therefore, it has therealistic significance and application value to study the fault diagnosis of control valve based on the data driven using the good methods of machine learning, pattern recognition, signal processing and data mining.With the Support of the National High-tech Research and Development Program (2008AA04Z130), the Specialized Research Fund for the Doctoral Program of Higher Education (20110131110042), the National Science Foundation (51305234) and the BinHua group, this paper firstly analyzes the characteristics of the control valve system in morden process control and the development status of the fault diagnosis method for control valve. And then the research direction based on Least Square Support Vector Machine (LS-SVM) is put forward. The main work and results are summarized as follows:Simulation fault test:Sample data is the reseach basis of the fault diagnosis based on data driven. In order to obtain the sampes of control valve, a fault simulation testing platform is designed and built referring to the working condition of a control valve from chlor-alkali industry of Binhua Gropes. Eleven kinds of working conditions including the normal situation, leakage and obstruction are simulated and the sample data is collected. Test results show that the default sample has good predictability and separability.Research of the LS-SVM prediction model and its parameter optimization method:Building an accurate model is the key for the fault diagnosis based on model. The mechanism model of control valve is not applicable because of its complex structure, nonlinear and time variation. Therefor, models based on LS-SVM are built for the flow and downstream pressure predicting of control valve. Then, the prediction accuracy with different input parameters is analized and the optimal input feature vectors for the flow and downstream pressure predicting are determined respectively. The influence of LS-SVM parameters C and a are studied using the actual production data of a control valve. A Fruit flies Optimization Algorithm (FOA) is put forward to optimize parameters of LS-SVM. The test results reveal that FOA has equal accuracy compared with the particle swarm optimization algorithm and the grid search method. Nonetheless, it has absolute advantage in computing speed and is more suitable for the online fault diagnosis of control valve.Research of LS-SVM and Hammerstein integrated prediction model and its denoising method:Control valve for precision control requires a more accurate model. In order to improve the accuracy, an integrated model of LS-SVM and Hammerstein is put forward. The nonlinear module of Hammerstein and related parameters are solved using LS-SVM. Compared with LS-SVM model, the prediction accuracy of the integrated model is improved. And compared with SVM model and BP neural network model, the presented method shows the highest precision and the least consuming time. What’s more, SVM is improved to have superior preformance than BP network for the small sample problem. Noises sensitivity of the integrated model of control valve is analyzed and the results show that the model is sensitive to noise when the signal noise ratio (SNR) is smaller than 40dB and the noise robustness is poor. In order to further improve prediction precision, the wavelet denoising method is used. The test results show that the noise robustness is effectively improved.Research of fault detecting method based on the residual control chart: Residual control chart is put forward with reference to the quality management control chart method. Residual is defined as the predicting percentage error and the residual threshold value is detimined according to the principle of 3σ The residual distributions of the flow and downstream pressure predicting under the eleven kinds working condition are analysed. The fault detection is tested using the residual control chart. The results show that more than 90% testing samples are detected correctly in addition to three kinds of defauts including the leakage of top flange and the light leakage of the cover flange. To solve this problem, some more rigorous fault detection rules are designed with reference to the definition of abnormal patterns in the quality control chart. The fault detection accuracy is greatly increased.Research of data cleaning method in the fault classification:Fistly, the effect of abnormal samples on classification is analysed. And then, a method to clean the abnormal data based on LS-SVM classification is proposed according to the characteristic that the abnormal sample has higher miss classified rate. And anothor method to clean the abnormal data based on local outlier factor (LOF) of Mahalanobis Distance is proposed according to the characteristic that the abnormal sample has bigger LOF. The default sample data of control valve is tested using "one to one" LS-SVM multiple fault classification and the proposed data cleaning methods. The results show that the proposed methods has a high accuracy for the fault classification of control valve.Implementation of the fault diagnosis system for control valve based on slide window:On the basis of the theory and methods proposed in previous chapters, the online intelligent diagnosis of control valve is realized by integrating the default detecting based on LS-SVM regression, the default isolation based on LS-SVM classification and the technology of sliding windows. And a satisfactory fault diagnosis accuracy is obtained. The research of this dissertation has important theoretical significance and engineering practical value to promote the online intelligent diagnosis of control vlave based on data driven.