Mill chatter is one of the most common problems in cold rolling. Thus, it is important to investigate the mill chatter phenomenon to ensure a high-speed and stable rolling process. The traditional mill chatter mechanism model cannot meet the monitoring and rapid diagnosis needs of the rolling process in the field. In this paper, a data-driven mill vibration analysis method is proposed. The main objective of this study was to develop a mill vibration monitoring method and an intelligent algorithm for mill chatter early warning. Rolling experiments showed that the proposed monitoring method could be a promising and effective technique for assessing the chatter phenomenon. The mill vibration acceleration amplitude prediction performance of a support vector regression, neural-network-based method, and extreme gradient boosting method were evaluated. The results proved that the prediction performances of the proposed extreme gradient boosting method were highly reliable with the highest determination coefficient value of 0.779, lowest mean absolute percentage error of 9.7%, and better forecast robustness under all of the dataset ratios. Meanwhile, the contribution rates of the variables on the mill vibrations were investigated, and results showed that an effective way to eliminate the mill chatter was to control the rolling speed, cumulative rolling strip length, tension and roll radius.

轧机颤振是冷轧中最常见的问题之一。因此，研究轧机颤振现象以确保高速稳定的轧制过程很重要。传统的轧机颤振机制模型无法满足现场轧制过程的监控和快速诊断需求。本文提出了一种数据驱动的轧机振动分析方法。这项研究的主要目的是开发一种用于轧机颤振预警的轧机振动监测方法和智能算法。滚动实验表明，所提出的监测方法可能是一种评价颤振现象的有前途和有效的技术。评价了支持向量回归，基于神经网络的方法和极限梯度提升方法的轧机振动加速度振幅预测性能。结果证明，所提出的极限梯度增强方法的预测性能高度可靠，在所有数据集比率下，最高确定系数值为0.779，最低平均绝对百分比误差为9.7％，并且具有更好的预测鲁棒性。同时，研究了变量对轧机振动的贡献率，结果表明，消除轧机颤振的有效方法是控制轧制速度，累计轧制钢带长度，张力和轧制半径。