基于改进慢快双流网络的锌快粗选工况识别

An Improved Slow-fast Dual-flow Network Based Working Condition Recognition of Zinc Fast Roughing Process

锌快粗选是锌浮选的第一个浮选工艺,其工况状态直接影响后续精选和扫选的性能效果。现有基于卷积神经网络的浮选工况识别方法直接构建泡沫图像和工况类别的关系模型,自动挖掘与工况关联的泡沫深度特征,取得了一定的效果,但忽略了不同采样率下泡沫视频的动态时序信息。为此,提出一种基于改进慢快双流网络的模型,通过泡沫视频对锌快粗选工况进行识别。首先,以慢快双流网络作为主干网络,引入P3D-A轻量化结构减少网络参数,提高模型推理效率。然后,融入时间注意力模块(Time Attention, TA)、空间注意力模块(Channel Spatial Attention, CSA)和空间时间聚合模块(Spatial Time Together, STT),对泡沫视频时空特征进行有效表征,促进双流网络融合。最后,设计辅助网络(Auxiliary Network, AN),降低模型过拟合,提高工况识别准确率。实验结果表明,所提方法能准确地识别锌快粗选工况,准确率达82.54%,与已有的慢快双流网络相比,准确率提高了11.98%。

Zinc fast roughing is the first process of zinc flotation,whose working conditions directly affect the performance of following cleaning and scavenging.The existing working condition recognition methods based on convolutional neural networks establish the relationship model between froth images and working conditions,and automatically mine the deep features associated with corresponding working conditions.These methods have acquired positive results,but ignore the dynamic temporal information of the froth videos at different sampling rates.Therefore,an improved slow-fast dual-flow network is proposed to identify zinc fast roughing working conditions based on froth videos.Firstly,the slow-fast dual-flow network is taken as the backbone network,and the P3D-A lightweight structure is introduced to reduce network parameters and improve the model reasoning efficiency.Then,Time Attention(TA),Channel Spatial Attention(CSA)and Spatial Time Together(STT)modules are integrated to preferably learn the spatial-temporal features of froth videos and promote the fusion of dual-flow networks.Finally,an Auxiliary Network(AN)is designed to reduce overfitting and improve the accuracy of condition identification.The industrial experiments show the effectiveness of the proposed method,in which the accuracy is 82.54%and is improved by 11.98%compared with the single slow-fast network.