Diagnosis Of Pneumatic Control Valve Based On Statistical Learning

With the expansion of the production scale and continuous improvement of industrial automation, Pneumatic control valve is more and more widely used in industrial process control. As the terminal actuator of control loop, Pneumatic control valve is often used to control the flow and pressure of all kinds of media, and it plays an important role in keeping production stable, process safety, optimization control and so on. Therefore, as an important part of process monitoring system Pneumatic control valve fault diagnosis study is of great importance for process safe, stable and efficient production in industries such as petrifaction, food and so on. This thesis focuses on the multi-fault diagnosis of Pneumatic control valve based on statistical learning. The main contents of the study are as follows:1. Study on the standard platform of fault diagnosis methods of Pneumatic control valve - DAMADICS(Development and Application of Methods for Actuator Diagnosis in Industrial Control Systems). Through the study of simulation model of Pneumatic control valve in DAMADICS, this thesis realize the simulation of multi-fault of Pneumatic control valve, thus solving the problem of data deficiencies of valve faults in real industrial application.2. Support Vector Machine(SVM) and Extreme Learning Machine(ELM) which are popular methods in statistical learning are applied to the multi-fault diagnosis of Pneumatic control valve. Based on the simulation data of multi-fault in DAMADICS, the diagnostic efficacy evaluation of the two methods are compared, the results show that ELM is more efficient than SVM in model training efficiency and model prediction accuracy.3. Propose the method of Sparse Bayesian Extreme Learning Machine(SBELM), the main idea of which is to apply the Bayesian thought to multi-classification of ELM to train the output weight of multiple classifiers of ELM. The hidden parameter of SBELM is random generated like the traditional ELM, thus keeping the features of original ELM. The model trained by SBELM can prune the repeated or disturbed training samples by predefined criteria, thus achieving model sparsity. SBELM can also give the class membership probabilities of test sample, which has important meanings for practical applications. Furthermore, considering the restriction of training model size in industrial application of fault diagnosis, this thesis propose the method of pruning hidden nodes based on SBELM. Finally apply the methods to the multi-fault diagnosis of Pneumatic control valve in DAMADICS and obtain a good diagnosis result.4. The industrial application is focused on the sugar factory Cukrownia Lublin SA, Poland. Based on the pneumatic control valve controlling the thin juice inflow into the first evaporator, applying SVM, ELM, SBELM to the fault diagnosis of three kinds of real valve faults, respectively. The results show that SBELM is better in prediction accuracy and testing time, ELM is better in model training time, while SVM performs general.