Concentrate Production Rate Prediction In The Mineral Process Based On Multiple Models And Particle Swarm Optimization

Mineral process is an industrial procedure to concentrate the valuable minerals in the raw ore via a chain of physical treatments and chemical transformations, and it is usually classified into following unit-processes:minerals liberations (crushing, grinding and size classification), separation processes (floatation, magnetic or gravimetric separations, sorting, leaching etc.) and concentrate pre-treatment (drying, agglomeration, sintering, etc.), in which a set of indexes are extracted to evaluate the quality or productivities of each process, denoted as technical indexes, while global production indexes are to evaluate the entire mineral process of one day or a operational team, including concentrate production rate, grade and metal recovery rate. It is of great significance to establish model between global production indexes and technical indexes in terms of both theoretical and practical aspects. In general, global production indexes cannot be measured on-line, and are got by statistic analysis algorithm based on the laboratory analysis data, but it takes about more than2hours to get the proposed value. Therefore, those values cannot be taken as real-time data to evaluate the present process on whether the proposed object is met, that is, whether global production indexes reach the proposed target, at the same time, the engineer adjusts technical indexes on the basis of global indexes by the mean of statistic analysis algorithm, which cannot guarantee the real-time optimization and control of the entire mineral process. Once the predictive model established, the technical engineer can cooperate and control the entire mineral process. Besides, this type of models makes it available for plant-wide optimization and control of the whole production process, which guarantees the proposed objective including the maximum production rate, proper grade of the pulp etc..Based on "Close-looped optimization algorithm on technical indexes of complex industrial process in dynamic environment" of National Nature Science Foundation Project of China, the dissertation mainly establishes the model between production rate and technical indexes in two algorithms, the former is based on LS-SVM and Adaptive cPSO, in which Adaptive cPSO is proposed to optimize the kernel parameter, penalty coefficient and select feature, whist multiple models approach is introduced to address the characteristic of multiple modes in the mineral process. The details are listed as follows:(1) Adaptive Chaotic Particle Swarm Optimization (Adaptive cPSO) is proposed to apply to SVM to optimize its kernel parameter, penalty coefficient and select input featureAiming at solving problems of continuous parameter optimization and discrete input feature selection of LS-SVM, we propose two novel strategies to modify the standard PSO. At first, chaotic operators are introduced to replace random operators, these operators can iteratively generate ergodic, non-repeated, and uniformly distributed solutions, which guarantees the search capability and stability. Besides, we put forward a novel adaptive search strategy, in which we optimize parameter in a manner of standard PSO unless the fitness value of the global solution (gbest) lessens in certain iterations, or adaptive lattice search will be introduced to create lattice particles to replace the worst PSO particles. The proposed algorithm is applied to SVM to its kernel parameter, penalty coefficient and select input feature.(2) Production Rate Prediction Based on Least Square Support Vector Machine (LS-SVM) and Adaptive Chaotic PSOMineral process is a complex process with many uncertain factors and multi-variable coupling. And it has the characteristics in terms of large-range continuity, nonlinearity and large-time delay etc., so there is no reported explicit physical model between technical indexes and production rate so far. However, the rich real-time and laboratory analysis data makes it possible to establish a data-driven nonlinear model by LS-SVM, in which adaptive cPSO is adopted to optimize kernel parameter as well as penalty coefficient and select the input feature. The results validate the proposed modeling algorithm an efficient tool for the prediction of concentrate production rate.(3) Production Rate Prediction Based on Multiple models and PSOA multiple models approach is proposed to predict production rate in the mineral processing plant, due to its characteristic of multiple modes. Fuzzy Maximum Likelihood Estimates (FMLE) is introduced into the structure of input partition owing to its characteristic of the ability to partition the data set of different shapes, sizes and densities. Next, Least Square Support Vector Machine (LS-SVM) with mixtures of kernels is proposed to build local models for each region. To enhance the performance of local models, Particle Swarm Optimization (PSO) is employed to optimize the kernel parameters, penalty coefficient and weighting coefficients. The practicality of the identification scheme presented here is demonstrated by application to mineral process for multiple prediction models between production rate of the concentrated ore and the technical indexes. Compares with the original methods of Takagi-Sugeno-Kang (TSK) fuzzy model and multiple Neural Network (NN) models, the proposed method is more efficient.