Working Condition Recognition Based On Deep Visual Features Of Froth Video And Broad Learning System In A Zinc Cleaner

Zinc flotation is an important process of zinc mineral processing.As the last process,the working condition of zinc cleaner directly determines the grade of zinc-concentrate.At present,the working condition of zinc cleaner is usually recognized by workers via observing the froth appearance,which is highly subjective.However,the zinc concentrate grade detected by the fluorescence analyzer has long test period,and thus it cannot be detected in real time when the working condition changes drastically.Therefore,the image processing technology is of great significance for the real-time and accurate identification of flotation conditions.This paper focuses on the working condition recognition based on deep visual features of froth video and broad learning system in a zinc cleaner,the main innovative works are as follows:Aiming at the problem that the extraction of dynamic features from flotation videos is difficult,a deep neural network-based method to extract deep visual features from flotation video is proposed.This method uses Separable 3D(S3D)convolutional neural network to extract short-range spatiotemporal features between adjacent frames of froth video,which employs channel and spatial attention mechanism to give high weight for important features.Then Bi-Directional Convolutional Long Short-Term Memory(Bi Conv LSTM)network is used to extract long-range spatiotemporal features between distant frames of froth video based on the short-range spatiotemporal features.Lastly,a two-dimensional convolutional network is constructed to extract the multi-scale appearance features of the froth image.Aiming at the problem that the adaptability of working condition recognition methods is insufficient,a working condition recognition method with incremental learning based on broad learning system is proposed.Specifically,a broad learning-based working condition recognition model is established,in which the short-range and long-range spatiotemporal features and the multi-scale appearance features are fused;when the working condition of zinc cleaner changes hugely,or the fault of zinc cleaner occurs,leading to a decrease of recognition accuracy of the model,the model structure of broad learning network is adjusted by incremental learning.Specifically,the feature layer,enhancement layer and output layer are expanded.When the recognition accuracy of model by incremental learning fails to meet the requirements,the network will be retrained.Experiments were conducted on the industrial dataset of zinc cleaner,and the effectiveness of the method was verified.Taking the lead and zinc flotation process as the application object,the working condition recognition method proposed in this paper is applied to the industrial production of lead and zinc flotation,and a set of working condition monitoring and recognition system is established in the flotation plant.The application results of this system show that the proposed method can effectively recognize the working conditions of zinc cleaner.