Multiple Kernel-based Classification Research And Its Application In The Performance Recognition For Copper Flotation

Abstract:Classifier design is one of the main research directions in the field of pattern recognition. More efficient classifier has been pursuing, and the situation changed a lot until the kernel method was introduced. However, the traditional method is based on a single kernel method, a single kernel function cannot effectively describe complex data distribution and pattern features. Multiple kernel method combined with several kernel functions offers a new idea for that problem. Because of the complex of the working environment and the nonlinear relationships between variables, the pattern recognition method without considering the feature selection and heterogeneous data has a limit in the copper flotation process. Main research and innovative achievements are as follows:(1) Focused on the nonlinear separability of features, which has long been ignored by traditional F-score, a multiple kernel F-score (MKF-score) feature selection method is proposed. A new sample distance has been defined, thus an effective evaluation criteria is obtained. Simulation results on the UCI database demonstrate that the proposed method is able to select the optimal feature subset and has achieved satisfactory classification result.(2) Focused on the heterogeneity of the target data distribution and lack of the non-target data, a classification method based on the multi-kernel SVDD (MKSVDD) is studied. Experimental results on the Banana database indicate that the method is applicable to heterogeneity data.(3) Focused on the problem that single kernel Gaussian process regression (SKGPR) has received comparatively large deviation while dealing with heterogeneous data, an expert classification method based on the multi-kernel GPR (MKGPR) is studied. Experiments have applied to data in different distribution and the results demonstrate that the method has achieved better performance when dealing with heterogeneous data.(4) Focused on the problem that a lot of nonlinearity even heterogeneous data in the copper flotation process, MKF-score, MKSVDD and MKGPR methods are applied to recognize the copper flotation performance and the method of multiple kernel-based copper flotation performance recognition is proposed. Industrial data experimental results verified the effectiveness of the proposed method, which can be well used to guide the actual flotation production operations.