Application Of Data Mining Methods To High-conductivity And High Flexibility Copper Alloys And Other Metallic Materials

Generally, data mining (sometimes called data or knowledge discovery) is the process of analyzing large amounts of data from different perspectives and summarizing it into useful information and finding the laws of nature. As a multidisciplinary technology, data mining has become the focus of research in the fields of database system and machine learning. It has also aroused wide interest of the academic community and industry because of their broad prospects for application.The goal of this work is to develop a number of new data mining methods and applied these approaches to empirical design of high-conductivity and high flexibility copper alloys and other metallic materials. As traditional methods such as the error back-propagation (BP ) neural networks, radial basis function neural networks (RBFN) and genetic algorithms (GAs) still suffer from some weaknesses, several novel data mining techniques have been proposed for the purpose of improvements through the comprehensive use of a variety of algorithms. In particular, support vector machine (SVM) developed by Vapnik, due to its rigorous theory background and remarkable generalization performance, is intruduced to empirical design of high-conductivity and high flexibility copper alloys. We will show the capability of SVM in forecasting the properties and the process optimization of copper alloys, and its potential utilities to solve problems which can not be solved efficiently by using traditional methods. Therefore, these researches on the establishment of the accurate relationship between material components, process and properties could provide new practical methods, and then enrich and improve existing methodology in material design theory.This study mainly involved the following aspects:(1) Least squares support vector machine (LSSVM), as a fast and efficient method, is successfully applied to the optimization of alloy components, thermomechanical treatment processes parameters and mechanical and electrical properties of Cu-15Ni-8Sn-XSi alloys. A knowledge repository on the domain knowledge or the processes is established via sufficient data mining by LSSVM and expressed directly through the use of three-dimensional map and/or its contour lines. The results show that the best Si content of Cu-15Ni-8Sn-XSi alloys is about 0.3. The Cu-15Ni-8Sn-0.3Si alloy can obtain preferable hardness (more than 412HV) and good conductivity after a 50% prior cold works and ageing at 400℃for 1.5h～3h. Additionally, the alloy can obtain a higher hardness (more than 416HV) and good electrical properties after a proper prior cold deformation and ageing at 450℃for 0.9～1.7h. The function curves predicted by the LSSVM model are similar to the analytical results of the principle of physics and metallurgical.(2) Traditional training algorithms for optimization of BP neural network are only based on gradient information and usually convergence to a local minima. To explore a new training strategy of BP neural network, a novel data mining approach, based on BP neural network using differential evolution (DE) training algorithm has been proposed, using Bayesian complexity regularization for weight decay in the error back-propagation learning procedure, the differential evolution algorithm improves the convergence performance and predictability of BP neural network. For the first time the differential evolution neural network was successfully applied to forecasting the mechanical and electrical properties of Cu-15Ni-8Sn-0.4Si alloy.(3) In order to overcome the shortcomings of standard SVM, the LSSVM, for the first time, is used to establish the empirical models capable of reflecting the complex interaction relationship between alloys composition, prior cold deformation, aging process parameters and the hardness and conductivity properties of the high conductivity and high flexibility Cu-Ni-Sn and Cu-Ti alloys. The models could be applied to the optimization of industrial copper alloys compositions and processing parameters and the improvement of properties.(4) The different models based on neural networks and LSSVM methods and forecasting capable of the hardness and conductivity of Cu-Ti alloys has been proposed and compared. The performance comparison results of the different models show that the forecasting performance of the proposed LSSVM model is in most cases superior over that of conventional ANN approaches. While the neural network models adopt the empirical risk minimization (ERM) principle which has adverse effects on the further improvement of forecast accuracy; the LSSVM model, due to embodying the structural risk minimisation (SRM) principle, has the potential abilities of finding the best compromise between the complexity of the model and learning ability, leading to the forecast accuracy closer to the fitting accuracy and achieving better generalization performance, even only based on limited information in the samples.(5) The design and training of the RBF networks is still in the exploration stage. In this study, the partial least square regression (PLSR) is applied to determine the weight of the RBFN. The PLSR picks-up orthogonal components form the primary independent variable data matrix, ignores the components of very little variance. So it eliminates the multicollinearity between the primary independent variables and ensure the regression process being steady. The new hybride model (RBF-PLSR) is successfully applied to Cu-Ti alloy performance prediction, showing better predictive performance and stability than the traditional RBFN.(6) Upon aging after solid solution for Cu-1.5Ti and Cu-4.5Ti alloys, there could be a linear relationship between the volume fraction of precipitates and the electrical conductivity. Based on the linear relationship, Avrami phase transformaion kinetics equation and electrical conductivity equation at different aging temperatures are described for Cu-1.5Ti and Cu-4.5Ti alloys. The function curves of the conductivity predicted by the LSSVM model are very similar to those which are discribed by the physical equations.(7) The mechanical properties of 7005 Al alloys were qualitatively analyzed by partial least squares (PLS) method and quantitatively calculated by using back propagation artificial neural network (BPN) with the same processing parameters as features. The calculated results are in agreement with experimental ones basically. In order to solve the overfitting problem, a novel method hybridizing PLS and BPN to forecast the mechanical properties of the alloys was proposed and tested. PLS method can compute the scores for the principal components according to the sorts of components and thus compress the input data for BPN with linear arithmetic. Consequently, the hybrid model provides the highest accuracy of forecasting.(8) A novel model (GASVR) that combined the improved genetic algorithms (GAs) and support vector machine regression (SVR) are proposed. The use of GAs in selecting efficiently SVR parameters for forecasting atmospheric corrosion depth of zinc or steel improves the forecasting performance of the proposed model due to embodying SRM principle and its advantage of converging to the global optimal solution. Consequently, the generalization performance of GASVR model is significantly superior over that of BP artificial neural network.