Abstract:

When the working condition of wet ball mill is changed, the distributions of real time data and modeling data are inconsistent, and the i.i.d assumption of traditional soft sensor method would not be satisfied, which leads to the inaccuracy of soft sensor model and the deterioration of the system performance. Therefore, by introducing the transfer learning method, a soft sensor model based on transfer variational autoencoder - label mapping strategy is proposed to achieve the precise measurement of wet ball mill load parameters under multiple working conditions. Firstly, the parameters of hidden variables distribution obtained by encode with target domain data are transferred to fit corresponding hidden variable of source domain data. For then, acquiring transfer data by decoding. Moreover, the similarity measure is used to select similar samples to construct the label mapping model, and the mapped labels are obtained. Then the final soft sensor model is constructed according to the transfer data and mapped labels. The experimental results show that the proposed soft sensor method is significantly better than the existing method, and the proposed method is suitable for soft sensor modeling under the situation of multi working condition.

Key words: transfer learning, variational autoencoder, label mapping, wet ball mill load parameters, process control, prediction, experimental validation

摘要：

在工况改变时，湿式球磨机的实时数据和建模数据分布不一致，不满足传统软测量建模方法要求的数据同分布假设，导致模型失准和性能恶化。为此，引入迁移学习思想，提出一种基于迁移变分自编码器-标签映射的软测量模型，实现多工况下湿式球磨机负荷参数的准确测量。首先，迁移目标域数据编码得到的隐变量分布参数，对源域数据对应隐变量进行拟合，再解码得到迁移数据；然后采用相似性度量选取相似样本构建标签映射模型，并得到映射标签；最后使用迁移数据和映射标签构建出最终的软测量模型。实验结果表明，该软测量方法显著优于现有方法，适用于多工况下的软测量建模。

关键词: 迁移学习, 变分自编码器, 标签映射, 湿式球磨机负荷参数, 过程控制, 预测, 实验验证