Fault Diagnosis And Prediction For The Imperial Smelting Furnace Based On Data-driven Technique

The smelting process of the Imperial Smelting Furnace (ISF) involves various complex physicochemical reactions, characterized by properties like multivariate, nonlinearity, strong coupling, large delay and uncertainty, etc. The complex and unstable component of the raw materials lead to the furnace running unsteadily, which consequently cause the poor output and quality of Lead-Zinc. Therefore, research on the new effective methods for the ISF fault diagnosis and prediction has an important and practical significance for guaranteeing the production safety and stability and improving industry economy profits.In view of the characteristics of Imperial Smelting Process (ISP), this thesis focuses the research on new methods of fault detection, fault diagnosis and prediction based on the data-driven technique. These methods presented are as follows: the improved fault detection method based on principal component analysis (PCA); the fault diagnosis approach based on PCA and multi-class classifiers of SVM; a novel fault diagnosis strategy for the incomplete samples based on Case-Based Reasoning (CBR) technique and SVM; the new fault prediction method based on Hammerstein model using weighted least squares support vector machines (WLS-SVM). All the methods are applied to the fault diagnosis and prediction system in the melting process of the imperial smelting furnace (ISF), and achieve good performances. The major innovation research achievements include:(1) Considering the characteristic of the ISP data in abnormal distribution, the novel fault detection method based on the improved principal component analysis (PCA) is presented. Firstly, the data are preprocessed by the exponentially weighted moving average (EWMA) and the improved method for the abnormal value removal, and the PCA model for the ISF is obtained. And then, the control limit of the PCA model is calculated by the method of multivariate kernel density estimation (KDE), which helps to reduce the false alarm rate or missing alarm rate of the traditional PCA model, and increase the sensitivity of the monitoring process and improve the detection of the ISF. (2) A fault diagnosis approach of the ISF based on PCA and multi-class classifiers of SVM is proposed. Firstly, the PCA approach is adopted to extract the feature, and the monitoring model is established. And then the reconstructed data are analyzed by the Q and T~2 statistics to determine if the supervised process exceeds the normal control limits. If fault occurs, then the supervising program will alarm to prompt that the process has abnormal states. Secondly, the SVM multi-class classifiers with 'one to other' algorithm are used for classification with the input of the feature. Experimental results have demonstrated that the proposed methods achieve excellent performance with high diagnosis accuracy and rapid speed.(3) For the incomplete samples for ISF, this paper presents a novel fault diagnosis strategy based on Case-Based Reasoning (CBR) technique. Firstly, the similarities between the retrieval feature vector and the complete samples in the historical database are measured by the method of distance measure or the relevance measure, the optimal case is retrieved and the imputation is obtained, finally the generated complete samples are inputted to the fault diagnosis models to diagnose the ISF system. The presented strategy is validated and has achieved satisfying results.(4) The ISP has characteristics such as nonlinearity, strong coupling, time-varying parameters. It is hard to predict its condition precisely by traditional method under fault condition. To solve this problem, the new fault prediction method based on Hammerstein model using weighted least squares support vector machines (WLS-SVM) is presented. Firstly, the error variables of the least squares support vector machines are weighted according to their distributions, and then the nonlinear function of Hammerstein model is constructed by the weighted least squares support vector machines regression. A numerical algorithm for subspace system (singular value decomposition, SVD) is utilized to identify the Hammerstein model. Finally, the fault prediction model for ISF is used to predict the key parameters and dynamic behavior of the furnace, which can reduce the failure rate and make the furnace steady run smoothly. (5) A data-driven fault diagnosis and prediction system of the ISP is designed and developed. It is able to forecast running state, track the trend of the running furnace, and give alarm and diagnose signals for abnormal parameters and states.