Abstract: The accurate prediction of the load parameters of the ball mill plays a key role in the monitoring and control of the grinding process,and the data drift in the multi-working environment leads to deep learning and other methods that have limited effect on the load prediction of the ball mill.For this reason,a ball mill load forecasting method based on adaptive network was proposed.A classification model of grinding and grading working conditions based on deep correlation alignment was established;then the relative position encoding was introduced into the Transformer,and the attention mechanism was decoupled to directly encode the position information into the attention mechanism,thereby improving the prediction performance.Furthermore,an adaptive network was proposed,which applied the distribution matching regularization term to the hidden layer features of Transformer model to learn the common parameters of the hidden state of the model by reducing the distribution difference between different working conditions,so as to improve the generalization ability of the model.A boosting-based approach was employed to learn the importance of hidden states.The test results showed that the proposed adaptive prediction network could significantly improve the accuracy of ball mill load parameter prediction,and the prediction performance was also ahead of the comparison method in the face of unknown working conditions.

Key words: ball mill load parameters, multiple working conditions, adaptive network, working condition division, Transformer model

摘要： 鉴于多工况环境下的数据漂移限制了深度学习方法预测球磨机负荷的效果,提出一种基于自适应网络的球磨机负荷预测方法。首先建立基于深度相关对齐的磨矿分级工况划分模型;然后将相对位置编码引入Transformer,对注意力机制进行解耦来将位置信息直接编码进注意力机制,进而提高预测性能;进一步提出一种自适应网络,将分布匹配正则化项应用于Transformer模型的隐层特征,通过减少不同工况之间的分布差异来学习模型隐藏状态的共同参数,提高模型泛化能力;最后采用基于Boosting的方法学习隐藏状态的重要性。试验结果表明,所提自适应预测网络可以明显提高预测球磨机负荷参数的准确性,而且在面对未知工况时预测性能也领先于对比方法。

关键词: 球磨机负荷参数, 多工况, 自适应网络, 工况划分, Transformer模型