基于自适应自然梯度法的高斯过程磨矿粒度软测量建模

针对现有的磨矿粒度测量仪表检测周期长,难以满足实时检测的问题,结合典型两段式磨矿回路的特点,提出了基于高斯过程(Gaussian process,GP)的磨矿粒度软测量建模方法,将自适应自然梯度(adaptive natural gradient,ANG)法应用到对高斯过程超参数优化过程中,构建基于ANG-GP磨矿粒度软测量模型,并分别与BP神经网络和支持向量机软测量模型进行仿真试验的比较研究。结果表明：基于ANG-GP的磨矿粒度软测量方法优于其他2种方法,且具有较高预测精度,能有效地对磨矿粒度进行在线检测,表明了该方法的有效性。

Modeling of Grinding Particle Size Soft Sensor Based on the Adaptive Natural Gradient Method in the Gauss Process

The online detection of the particle size is of great significance to realize the optimizing control of the grinding process and to improve the grade of concentrated ore and metal recovery rate. However, the problem of the present instrument is that the particle size cannot meet the real-time detection due to the long measurement period. Based on the characteristics of the typical two stage grinding circuits, this paper puts forward the grinding particle size soft sensor modeling method based on Gaussian process ( GP) , and the adaptive natural gradient ( ANG ) is applied to the super Gaussian process parameter optimization of the process. Then, the model of grinding particle size soft sensor was built based on ANG-GP. Soft sensor simulation experiment was carried out comparative study with the BP neural network and support vector machine model, respectively. Results show that this method is superior to the other methods, this method has high prediction accuracy, and it is effective to online detection of grinding particle size, which shows the effectiveness of this method.