PSO-BP神经网络的分级过程建模

The modeling of classification process based on PSO-BP network

分级条件直接影响分级效率和精矿品位,为建立分级效率与分级条件之间的关系,首先通过机理分析建立分级效率模型结构,再采用BP神经网络建立模型结构参数与分级条件之间的关系。在网络训练中,考虑到基于梯度的优化方法易陷入局部极小的缺陷,采用PSO算法优化网络权值和阈值。实验表明,与基于梯度的动量BP算法相比,PSO算法训练和测试网络的精度和稳定性均优于前者。最后,将训练好的网络用于实际分级效率模型进行分级效率预测,预测的结果为实际值与估计值的相对误差在7%以下,这表明预测精度能达到给定的工业指标。

Given that the direct effect of classification conditions on classification efficiency and concentrate grade, it is necessary to establish the relation between the classification efficiency and classification conditions. The model structure of the classification efficiency is firstly built through mechanism analysis; however, another relationship of structure parameters and classification conditions are hard to set up. Considering the good capacity of artificial neutral network to approximate nonlinear systems, the BP network is adopted to build the relationship. Nevertheless, the optimization algorithms mostly used in BP network training are methods based on gradient, and they are apt to get trapped into local optimum and cannot achieve an ideal precision under certain network structure. To improve the performance of the training in modeling, PSO algorithm is presented to train the weights and biases. Compared to momentum BP algorithm（based on gradient and mostly used）, the experiment results show that the PSO algorithm has better performance in both precision and stability, and the results also indicate that not any network structures are appropriate. At last, the trained network is applied to classification efficiency prediction, and the results show that the relative error of real and predicted values can be lower than 7% with a high probability, which can meet the industrial demand.