Monitoring Of Key Performance Indexes Based On Sequencing Features Of Multiple Videos In A Zinc Flotation Process

Froth flotation is a complex industrial process with long process and large delay.In the froth flotation,the key performance indexes directly reflect the quality of production process,and they are an important basis for timely formulating and adjusting the flotation operations or controls.However,the froth flotation process includes complex physical and chemical reactions involving the interactions of solid,liquid and gas.Meanwhile,the flotation cells are closely related and coupled seriously,and the working conditions are complex and changeable.The existing key performance indexes monitoring methods in froth flotation often use a single froth image or video,which cannot effectively capture the dynamic temporal information between multiple froth videos.Moreover,the froth video and key performance indexes are not synchronized,and the sampling periods of them are different.Thus,the froth video is difficult to match the key performance indexes accurately,resulting in poor monitoring results.Therefore,in this thesis,the key performance indexes monitoring method based on the time series of multiple video features is studied in a zinc flotation process,and a key performance indexes monitoring system is developed in a real zinc flotation process to provide accurate and timely production quality results.The research of this thesis has important research value and practical significance for the efficient production of flotation plant.The main work and contributions of this thesis are as follows.(1)To solve the bubble size extraction problem caused by uneven bubble size distribution and bubble adhesion,a bubble size extraction method using an optimal marker-based watershed segmentation is proposed.To improve the performance of bubble size extraction,a sub-image classifier is constructed and a lumped together marker removing strategy is designed.Then,based on the class information of each sub-image and the lumped together marker removing strategy,an optimal marker extraction method is discussed,which takes advantages of the adaptive threshold method and morphological reconstruction method.Experiments show that the proposed optimal marker-based watershed segmentation algorithm improves the accuracy of bubble size extraction.Additionally,a bubble burst feature extraction method based on weighted normalized cross correlation(WNCC)and Chamfer distance(CD)is proposed.First,the local motion correction is proposed and placed before the image similarity calculation to solve the problem of different bubble motions.Then,the similarity calculation based on WNCC and CD is designed to solve the problem of intensity changes when bubbles move.Compared with the existing bubble burst feature extraction methods,the proposed method is more accurate and robust.(2)In the froth flotation,the froth video and pulp grades are always not synchronized,and the sampling periods of them are different.It is difficult to match the froth video and pulp grades accurately.Therefore,an extraction method of temporal fusion feature based on the multiple froth videos and feed grades is proposed.First,the minimum time matching and up-sampling is introduced to integrate the froth video visual features and feed grades.Then,the time series of multiple video fusion features is constructed according to the relevance to the target grade.After that,the time series of multiple video fusion features is sent into a recurrent neural network to extract the feature vector representing the target grade.The proposed method effectively uses the dynamic change trend of multiple froth video features and solves the unmatched problem between the froth video and pulp grades.Compared with the single froth video features and the single froth image features,the time series of multiple video features have better representation ability for key performance indexes in froth flotation.(3)When the working conditions change rapidly,the representation ability of froth visual features is limited to the key performance indexes in froth flotation.To solve this problem,a Siamese time series and difference network(STS-D net)is proposed to monitor the tailings grade of the zinc fast rougher.First,a Siamese time series sub-network is designed to extract effective and uniform representation of the input time series at the current and previous moments.Then,a difference sub-network is introduced to predict the grade in an incremental way.After that,a dual task learning strategy with the temporal representation constraint of video features at the previous moment is proposed.The STS-D net effectively integrates the time series of multiple video features at two adjacent moments and the grade data at the previous moment,and it makes full use of the change trend of multiple video features and grade data at two adjacent moments.Experiments show that the proposed STS-D net improves the accuracy of the tailings grade in the zinc fast rougher.(4)In the froth flotation,the flotation cells are closely related and coupled seriously.It is difficult to use a single froth image or video to represent the concentrate/tailings grade.Therefore,an encoder-decoder and Siamese time series network(ES-net)is proposed to monitor the zinc concentrate/tailings grade.First,according to variation characteristics between the time series of video features in the zinc fast rougher and the time series of target grades,an encoder-decoder model is designed to predict the target grade.Meanwhile,according to variation characteristics between the time series of video features in the target flotation cell and the time series of target grade,a Siamese time series and difference network is constructed to predict the target grade.Then,a multi-task learning strategy is proposed to integrate the encoder-decoder model and Siamese time series and difference network.Because the proposed ES-net effectively integrates multiple froth videos from different flotation cells,it can obtain more accurate concentrate/tailings grades than the existing models.(5)Based on the above researches,a key performance index monitoring system has been developed and applied to a real froth flotation process.The system can measure the zinc tailings grade,zinc concentrate grade and tailings grade of the zinc fast rougher accurately and timely,which validates the effectiveness of the theory,method and technology proposed in this thesis.