Complex Industrial Process Data Mining Method And Its Application In Copper-matte Converting

The complex industrial processes represented by non-ferrous metallurgy generally have such characteristics as multi-variable, non-linear, long time-delay, strong coupling and so on. Therefore, they are difficult to be controlled optimally using system principle model. Decision and operation of domestic complex industrial processes rely on human's experience to a great extent, so some relevant production process consume too much energy and raw material, run unstably and have great potential in saving energy, improving product yield and quality and so on. On the other hand, with the development of industrial automation, lots of production process data are accumulated, these data maybe contain some useful information such as system's operation law, operator's experience, optimal operation pattern and so on, but they have not been made the best used because of the restriction of data analysis technology level. Therefore, studying the Complex Industrial Process Data Mining (CIPDM for short) is of great theoretical significance as well as application value. Some problems of CIPDM methods and their applications are studied in this dissertation, the main research contents and results are as follows:1. Based on analysis on structure and basic optimization problems of complex industrial process, a basic framework of CIPDM is proposed. This framework normalizes the definition, basic tasks, general realization process and algorithm structure of CIPDM, stresses the roles of industrial process principle analysis in improving the efficiency and validity of CIPDM, and is of great guidance value for the implementation and application of CIPDM.2. In view of the validity of results of data mining is directly impacted by the data quality, an approach based on wavelet analysis for detecting and amending anomalies in data set is proposed in order to improving the data quality. Taking full advantage of wavelet analysis' abilities of "high-pass filtering" and "time-frequency analysis", this approach detects and amends the anomalies according to the values of their wavelet transformation coefficients; fast computation algorithms for low dimensional wavelet transformation coefficient are proposed, which can be directly applied to process anomalies in low dimensional data set; integrating the means of attribute reduction, a method based on wavelet analysis and non-linear mapping is proposed to detect the anomalies in multi-dimension data set. Simulation experiments show that the approach is accurate and practical.3. Based on the depth analysis on the relation between model performance and data quality, an idea of "optimal modeling restricted by data quality" is proposed. Through qualitatively and quantitatively analyzing the influence on the accuracy of the model due to data quality factors such as noise strength, sample size and so on, it is pointed out that training error of model has optimal value (be called as expectation training error), which can be estimated according to data quality information and be used as the criterion of optimal modeling. Therefore, a new optimal modeling idea, "estimate expectation training error firstly, then adjust model structure according to the difference between the expectation error and real error", is proposed, and the keystones and difficulties to realize this modeling idea are analyzed. This idea provides a new criterion not relying on the test data, so it can significantly improve the time efficiency of optimal modeling. In this dissertation, this idea is used to optimize artificial neural network (ANN) model and support vector machine (SVM) model, and achieves good performance.4. In view of the performance of ANN is very sensitive to its structure and training method, a new optimal modeling method based on ANN with double-net structure is proposed. This method inherits the advantages of two kinds of traditional methods, those are "structure pruning/growing" and "early stopping", and overcomes their defects:the ANN is of parallel two sub-nets with same structure, which are both trained by "early stopping", therefore, "over-fitting" can be avoided as well as the "incline problem" of "early stopping" be solved; the structure of two sub-nets are adjusted according to the idea of restricted optimal modeling, therefore, the structure of ANN can be optimized with high time efficiency. Simulation experiments verify that the method has better performance than relational traditional methods.5. In view of the problem that the parameters of SVM are relatively many and the optimization of which lacks theoretic basis, a new efficient and accurate optimal modeling method based on SVM is proposed. Based on the analysis on the coupling degrees among the three parameters of SVM, the parameters optimization problem is divided into two relatively independent optimization problems:kernel parameter optimization and structure parameters (that is intensive parameter and regularization parameter) optimization. A new Kernel Alignment Coefficient based on distance relation is introduced into kernel parameter optimization in this paper. Two structure parameters are proposed to be optimized synchronously according to the idea of "restricted optimal modeling". In addition, a method to evaluate the reasonability of the SVM model by distribution characteristics of the model errors in training set is proposed. Simulation results show that the method proposed in this paper is nearly as accurate as "cross validation" method, while much more rapid than "cross validation".6. Process mechanism, operation technology regulation of copper-matte converting and application conditions of data mining are comprehensively analyzed, based on these, optimal decision methods based on data mining are proposed, two new evaluation indexes, named as support degree and confidence degree, are introduced into model evaluation. Using some research results of this dissertation in basic theory and methods of data mining, and according to historical data accumulated in production process of matte converting of a factory, optimal decision making models for flux adding amount and blasting time are built. Simulation results show that these two decision-making models can significantly improve the converting quality of S1 period, and have great popularization value. The modeling process reflects the means of analyzing and solving problems of optimal decision making by data mining method, which is of guiding significance for optimal decision making of other complex industrial processes.This dissertation's research shows that data mining method is of great application value and broad application prospects in optimal decision and control of complex industrial processes which have accumulated a large number of historical production data.