Froth Image Bias Based Reagent Control For The First Zinc Rougher Flotation

A well controlled reagent dosages is the key to improve the zinc rougher flotation performance.Because there are no effective online feed grade testing methods and working condition perception approaches,as a result,reagent dosages in real industry are manually adjusted by observing the froth surface behavior.Machine vision-based reagent dosage control is known as a promising technique for replacing manual control.However,froth images captured at the flotation plant are often affected by noises,which make it difficult to extract handcrafted features consistent with the operator’s experience.Also,some highly abstract visual attributes cannot be extracted by designing handcrafted features,which hinders the effective implementation of subsequent control strategies.Therefore,to stabilize and improve the flotation performance,it is of great importance to study the froth image bias-based reagent control methods.This paper mainly focuses on "feed grade estimation","setpoint image generation","froth image modeling" and "reagent control",and explores momentum memory network-based feed grade estimation,Setpoint GANbased setpoint froth image generation,and memory-augmented generative adversarial network-based froth image modeling.On these basis,a reinforcement learning reagent control method is formed,which takes the bias between the real-time sampled froth image and the generated setpoint froth image as the instantaneous return.The main research results and contributions are as follows:(1)Aiming at difficulties in real-time feed grade testing and the online updating of a data-driven model,a momentum memory network-based feed grade estimation method is proposed.Specifically,a memory network with momentum key updating mechanism and uses a cascade module of convolutional neural network and self-attention mechanism as image sequence processing module is explored.By fully utilizing characteristics of memory network,an access frequency-based memory updating method is designed which enable the soft-sensor learn new feed grade estimation knowledge from newly collected data without performance degradation on old training data.Experiment results show that the newly developed method can largely improve the accuracy of the feed grade estimation results of zinc,lead,and iron.(2)Aiming at the problem of adaptive calculating references for froth image bias-based reagent control when feed grade fluctuates greatly,a generative adversarial network-based setpoint froth image generation method is developed.Specifically,it studies basic requirements in setpoint froth image generation from the perspective of reagent control,designs consistency loss function by combining froth flotation kinetics and metallurgical index requirements in flotation plants,and constructed generative adversarial networks to study the setpoint generation relationship between the sampled froth image,feed grade and setpoint froth image which representing optimal flotation conditions from historical industrial data.Experiment results demonstrate that the newly developed method can effectively generate setpoint image and lays a foundation for the subsequent froth image bias-based reagent dosage control.(3)Aiming at the problem of improving data efficiency in reinforcement learning which learns reagent control strategies from historical industrial data,and providing feedback for policy evaluation,a memory-augmented generative adversarial network-based froth image modeling method is proposed for the first zinc rougher.Specifically,a multi-head memory network is adopted to learn the dynamic mechanism of the froth state transition and hence improves the prediction accuracy and reliability of the froth image model.At the same time,an image generation module based on dual adversarial autoencoder is designed to further augment the controllability of the predicted image.On these basis,adversarial projection loss,adversarial content consistency loss and perception loss are designed for model training.Experiment results demonstrate that the newly developed froth image modeling method is effective in first zinc rougher froth image modeling.(4)Aiming at the high-dimensional problem in froth image bias-based reagent dosage control,a reinforcement learning control method based on memory augmented froth image model is proposed.The new method takes the bias between real-time sampled froth images and the setpoint froth image to construct instantaneous returns for reinforcement learning,and for this purpose,a froth image bias metric function based on convolutional neural network is designed.Aiming at the problem of action distribution shift in learning reagent dosage control strategy from fixed historical industry data,considering the predictive reasoning characteristics of memory augmented model,an action distribution shift detection method based the relevance degree of memory addressing is developed,and the relevant memory support degree is used to add a negative constraint to the image bias-based instantaneous return,so that reinforcement learning can utilize process data generated by the froth image model to eliminate its action distribution shift during training.Experiment results demonstrate that the proposed method can effectively realize the froth image bias-based reagent dosage control in first zinc rougher flotation,which has important practical significance for improving the flotation performance and flotation efficiency,and increasing the economic benefits of flotation plants.