Two-Stream Froth Video Feature-Based Reagent Optimal Control For The First Zinc Rougher

Zinc flotation uses the hydrophobic difference of mineral surface to separate useful zinc from gangue.It is an important procedure in the processing of zinc ore.In zinc flotation process,the first zinc rougher realizes the first separation of useful zinc particles and gangue,and its separation performance directly affects the production indexes of subsequent rougher,cleaner and scavenger operations.In order to improve the floatability of zinc particles,a variety of reagents need to be added into the first zinc rougher to change the physical and chemical properties of zinc particles and promote the formation and stability of bubbles.However,due to the complex process mechanism,strong nonlinearity and mutual correlation among variables existing in zinc flotation process,it is impossible to establish an accurate mathematical model,which makes it difficult for conventional control methods to achieve satisfactory results.All along,plant operators manually adjust the reagents by experience through naked-eye observation of the froth state in flotation plant.This "manual observation and operation" mode is highly subjective,which is difficult to ensure the stable operation of flotation production process,resulting in serious waste of resources and large reagent consumption.Therefore,this paper studies two-stream froth video feature-based reagent optimal control for the first zinc rougher.This study concentrates on three problems: "froth video feature extraction","flotation working condition recognition" and "flotation reagent optimal control".A two-stream froth video feature extraction method based on spatio-temporal convolutional neural networks is developed;A flotation working condition recognition method is developed based on the combination of knowledge distillation and supervised contrastive learning;A knowledge base that describes the setting rules between feed grade and two-stream feature setpoints is constructed using fuzzy association rule mining method and genetic algorithm;On these basis,a conservative double Q-learning based flotation reagent control method is developed.The main research work and innovative achievements of this study are as follows:(1)Since handcrafted features are limited by domain knowledge,weak working condition representation ability,low accuracy and poor portability,a two-stream froth video feature extraction method is proposed.The method includes a spatial convolution neural network to process static images of froth videos and a temporal convolution neural network to process consecutive optical flow images of froth videos.The spatial convolution neural network relies on the feature reuse and bypass setup mechanism of Dense Net to transfer the important local pattern of single froth frame to the deep feature space to obtain the static feature.The temporal convolution neural network uses the optical flow method to transform the pixel displacement of two consecutive froth frames into a static displacement field.Then the decomposed displacement field is sent into inception-v3 and convolutional long short-term memory unit model to obtain the dynamic feature.Industrial data experiments show proposed two-stream froth video features can completely represent froth surface.(2)Since low-discrimination feature encoding cannot accurately identify visually similar working conditions,a working condition recognition method based on knowledge distillation and supervised contrastive learning is proposed for the first zinc rougher.This method uses domain knowledge distillation and supervised contrastive learning as auxiliary tasks for classification.Distillation learning transfers the semantic knowledge of handcrafted features to the deep learning feature encoding network through the soft label generated by the softmax output layer under high temperature.Supervised contrastive learning algorithm performs similar encoding on images of the same working conditions and different encoding on images of different working conditions in the feature space.At the same time,memory queue-based negative sample augmentation and hard negative sampling are also designed to further improve the feature encoding discrimination.The average accuracy of proposed method in the working condition recognition experiment of the first zinc rougher is 88.82%,which is better than the existing flotation working condition recognition algorithms.(3)Considering different feed grades have different optimal froth states,an optimal setting method of two-stream features based on genetic algorithm and fuzzy association rule mining is proposed for the first zinc rougher.The optimal setting heavily depends on the feed grade.Since it is hard to measure the feed grade online,a multi-head nonlocal neural network model that uses two-stream features,the correlation of two-stream features in different subspaces and reagent dosages to calculate the feed grade is built.To reduce the redundancy of high-dimensional features,a stacked denoising autoencoder is employed to reduce the dimension of the two-stream froth video features.Finally,the optimal setting rule base of "feed grade-target two-stream feature" is established by genetic algorithm and fuzzy association rule mining method,which can be used as the optimized target for reagent control.(4)It is difficult to establish an accurate process model for industrial flotation,and online strategy exploration is likely to cause fault conditions.Hence,an optimal control method based on conservative double Qlearning is proposed.This method depends on actor-critic structure,and uses the conservative regularization term to guide the Q function to punish the control that shifts the dataset distribution in Bellman backup,so as to learn the optimal control strategy from fixed historical data.To improve the accuracy of policy evaluation,max-min double Q learning is used.For each state-action pair,its Q value is a convex combination of the outputs of two Q functions.After obtaining the double Q function,the optimal control strategy can be derived by solving "max Q".Experiments show that the proposed method can effectively use two-stream froth video features to realize the optimal control of reagents,which has important practical significance for improving the zinc concentrate,the utilization rate of mineral resources and the economic benefits of enterprises.