基于RAGAN的工业过程运行指标前馈-反馈多步校正

RAGAN Based Feedforward-feedback Multi-step Correction of Operational Indices for Industrial Processes

针对工业过程运行指标反馈校正存在滞后及一步推理校正模型可解释性差的问题,提出了基于递归注意力生成对抗网络(Recurrent attention generative adversarial networks, RAGAN)的运行指标前馈-反馈多步校正方法.该方法采用基于负相关正则化的集成随机权神经网络,建立综合生产指标预报模型,为校正提供前馈信息补偿反馈校正的滞后性.提出的RAGAN校正,采用多步校正实现一次调整的思想,将当前时刻运行指标映射到低维潜变量空间简化数据复杂度,利用长短期记忆(Long short-term memory, LSTM)模型实现数据的分步输入,提高模型可解释性;采用分布式注意力(Distributed attention, DA)机制构建数据读入网络,使校正环节获取任务相关性更高的数据,降低任务复杂度,减小噪声干扰,利用校正后的运行指标,保证系统的综合指标尽可能地跟随设定值运行.采用中国西部地区最大选矿厂实际数据的仿真实验,验证了所提方法的有效性.

To address the problems of time delay in feedback correction and weak interpretability of the correction model of the operational indices,this paper proposes a feedforward-feedback multi-step correction method for operational indices based on the recurrent attention generative adversarial networks(RAGAN).A global production indices prediction model is adopted to provide feedforward information for the feedback correction and compensate the time delay,which is established by the integrated stochastic weight neural network based on negative correlation regularization.The proposed RAGAN based correction approach adopts the idea of multi-step correction to realize one-time adjustment,mapping current indices to low-dimensional latent variable space to simplify data complexity,the long short-term memory(LSTM)is used to realize the step-by-step input of data which can improve the interpretability of the proposed correction model.In addition,a distributed attention(DA)mechanism is adopted to construct a data sensing network,which enables the correction model to obtain more task-related data and reduce the complexity as well as noise interference.The corrected operational indices are able to ensure that the global production indices follow their target values as far as possible.Simulation experiments based on actual data collected from the largest mineral processing plant in western China verify the effectiveness of the proposed approach.