基于双流Mobile Vit与通道剪枝的工业过程故障诊断

Industrial Process Fault Diagnosis Based on Two Stream Mobile Vit and Channel Pruning

摘要: 工业过程视频监控数据的时间连续性和空间连续性，但是已有的基于视频数据的故障诊断模型有规模庞大而难以部署。为此，本研究提出了一种基于双流Mobile Vit的轻量化视频分类模型用于故障诊断，并且利用权重剪枝技术来降低模型大小。首先提取工业视频的视频帧和稠密光流分别作为工业过程的空间特征和时序特征，再使用两条轻量化主干网络Mobile Vit提取视频的空间特征和时序特征。最终在双流模型的尾部使用卷积注意力融合机制使光流特征和特征充分融合用于最终的诊断。为使模型更加轻量化，模型权重剪枝被用来降低双流Mobile Vit的参数量。通过实验表明，相比于其他故障诊断模型,本研究所提出的模型和所应用的剪枝方法在取得较高的诊断精度的同时，模型大小也远远低于其他模型。

Abstract: Industrial process video surveillance data has the temporal and spatial continuity，however the existing fault diagnosis models based on video data is too large be deployed in real actual industrial process. Therefore, this study proposes a lightweight video classification model based on Two-stream Mobile Vit for fault diagnosis, and the weights of the Two-Stream Mobile Vit model are pruned. Firstly, video frames and dense optical flows of industrial videos are extracted as spatial and temporal features of industrial processes, respectively. Then, the backbones of the proposed model Mobile Vit is used to extract deep spatial features and temporal features. Finally, the Convolutional Attention Fusion Mechanism is used at the tail of the Two-Stream model to fully fuse spatial features and temporal features for final diagnosis. In order to make the proposed model more lightweight, model weight pruning is used to reduce the number of parameters of the whole model. It is shown through experiments that compared with other fault diagnosis models, the model proposed and the pruning method applied in this study achieve higher diagnostic accuracy while the model size is much lower than other models.