基于LSTM的浮选钼精矿品位软测量模型

战略性稀有金属钼矿品位低，组分复杂、嵌布粒度细等特点，其有价金属分离回收难。浮选作为微细粒钼矿分离回收的主要选矿方法之一，其浮选钼精矿品位一直是选厂的关键性产品指标。国内大多数选厂采取轮班制采样，人工化验得到精矿品位结果，但此方式严重滞后于浮选工艺，难以满足对生产过程进行实时监测和操作指导。LSTM是一种特殊的循环神经网络，引入门机制有效的传递或选择性遗忘长时间序列中的信息，解决RNN中的长期依赖、梯度消失和爆炸问题。本文分析整理东坡选厂中各平台源数据，结合选厂浮选工艺及机理，筛选出多个影响浮选钼精矿品位的变量作为模型输入；将输入变量进行异常值判定，缺失值填充和数据降噪等数据预处理，建立高质量浮选钼精矿品位数据库；软测量模型采用PyCharm软件编码，使用BatchNorm批量规范化处理样本数据，加入Dropout正则化防止过拟合，建立基于LSTM的浮选钼精矿品位软测量模型，通过前向传播算法更新神经网络结构参数，并于Linear模型和CNN模型的预测性能指标结果比较。结果表明：基于LSTM的浮选钼精矿品位软测量模型预测准确度高，样本数据误差波动平稳，浮动范围小，模型泛化能力强，模型平均绝对百分比误差MAPE为1.13%，均方根误差RMSE为0.7049%，决定系数R2为0.8763，实现了浮选钼精矿品位的在线预测。

Soft measurement model of flotation molybdenum concentrate grade based on LSTM

The strategic rare metal molybdenum ore has the characteristics of low grade, complex fractions and fine embedded grain size, with its valuable metals difficult to separate and recover. The flotation, as one of the main beneficiation methods for separation and recovery of microfine grain molybdenum ore, and Its flotation molybdenum concentrate grade has been the key product indicator for the processing plant. The majority of domestic processing plants adopt a shift system sampling, manual assay to get the concentrate grade results, but this way seriously lags behind the flotation process, it is difficult to meet the production of the process for real-time monitoring and operational guidance. LSTM is a special kind of recurrent neural network, and the introduction of the gate mechanism effectively pass or selectively forget information in long time sequences, solving the long-term dependency, gradient disappearance and explosion problems in RNN. This paper analyses and collates data from various platform sources in the Dongpo processing plant, combines with the process and mechanism of flotation in the processing plant, screening out a number of variables that affect flotation molybdenum concentrate grade as model inputs; the input variables are subjected to data pre-processing such as outlier determination, missing value filling and data noise reduction to build a high-quality molybdenum flotation concentrate grade database, the soft measurement model adopts PyCharm software coding, uses BatchNorm batch normalization to process sample data, adds Dropout regularization to prevent overfitting, establishes LSTM-based soft measurement model for flotation molybdenum concentrate grade, updates neural network structure parameters by the forward propagation algorithm, and the predictive performance index results of the Linear model and CNN model are compared. The results show that the LSTM-based soft measurement model for flotation molybdenum concentrate grade has high prediction accuracy, smooth sample data error fluctuation, small floating range, strong model generalization ability, the average absolute percentage error MAPE of 1.13%, the root mean square error RMSE of 0.7049% and coefficient of determination R2 of 0.8763, realizing the online prediction of flotation molybdenum concentrate grade.