Study of the Load Forecasting of a Wet Mill Based on the CEEMDAN-Refined Composite Multiscale Dispersion Entropy and LSTM Nerve Net

Abstract

To address the problem of wet ball milling in a strong noise environment, it is difficult to accurately detect the internal load parameters of the cylinder during grinding. In this paper, a mill load parameter prediction method is proposed based on complementary ensemble empirical mode decomposition (CEEMDAN)-refined composite multiscale dispersion entropy (RCMDE) and-long and short-term memory (LSTM) neural networks. Using this method, the vibration signals of the mill barrel under strong noise were decomposed using the CEEMDAN algorithm, sensitive modal components with strong correlation with the original signal were selected for reconstruction using the correlation coefficient method, and features of the reconstructed signals under different load parameters were extracted through RCMDE. The load characteristic vector of an RCMDE mill was used as the input of LSTM neural networks, and the filling rate, material and ball ratio, and grinding concentration were used as the output to establish the internal load prediction model of wet mill. Experiment results show that the prediction method has a high accuracy, with average absolute percentage errors of the filling rate, feed-to-ball ratio, and grinding concentration of 6.08%, 3.50%, and 3.47%, and average absolute errors were of 0.0167, 0.0146, and 0.0146, respectively.