基于强化学习的减少烘丝过程中烟丝 “干头” 量的方法

针对烘丝开始阶段存在的烘丝温度超调、过干烟丝较多等问题, 提出一种基于强化学习 (Reinforcement learning, RL)的减少烟丝“干头” 量的方法. 该方法利用生产实时数据作为输入特征向量感知烘丝生产过程的状态变化, 以烟丝含水率检测值为依据来评价、优化烘丝温度控制策略, 实现对烘丝机温度设定值的在线修正, 优化烘丝开始阶段的温度控制, 有效改善烟丝过干问题. 与烘丝机的自动控制模式和人工干预模式相比, 烟丝含水率的标准偏差比自动控制时降低了44.7%, 比人工干预时降低了14.3%. 实验结果表明烟丝含水率的稳定性有较大提高, 烟丝“干头” 量明显减少, 验证了所提方法的有效性和可行性.

A Method for Reducing Over-dried Tobacco at Head Stage of Drying Process Based on Reinforcement Learning

To solve the problem of high overshoot of drying temperature and too much over-dried cut tobacco at head stage of drying process, a method for reducing over-dried tobacco based on reinforcement learning (RL) is proposed. The presented model detects dynamic performance of tobacco drying system relying on real-time production data, evaluates and optimizes the temperature control according to the amount of moisture content in tobacco, and performs real-time correction for the set value of dryer temperature. The control strategy optimizes the temperature control and effectively improves the over-dried problem. The proposed method is compared with the automatic control mode and manual intervention mode of dryer. The standard deviation of the moisture content in dried tobacco is reduced by 44.7% compared with automatic control, and decreased by 14.3% compared with manual intervention. The experimental results show that the stability of the moisture content level is improved, and the amount of over-dried tobacco is significantly reduced, which verify the effectiveness and feasibility of the proposed method.