| The stator surface health in the generator chamber is the key factor affecting the normal operation of the whole generator.The traditional stator detection method of pulling out rotor is low efficiency,high cost and high risk.At present,the intelligent detection method of stator surface state without pulling out rotor mainly relies on the manual way of viewing video,so the detection efficiency and accuracy can not meet the industrial needs.In order to make stator surface state detection more intelligent and miniaturized,a two-level detection method of stator surface state based on machine vision was studied in this thesis,which was based on the actual intelligent detection platform with the actual project as the background.In the first level,panoramic mosaic of stator surface sequence images is carried out to replace lengthy video information,which is convenient for comparison of historical panoramic images and analysis of the quality change trend of stator surface.Firstly,the long sequence image mosaic effect of APAP algorithm is studied.In order to further improve the naturalness of mosaic,this thesis continues to study the long sequence image mosaic based on the improved AANAP algorithm.Aiming at the feature of weak texture region on the stator surface,a joint registration method of line feature and point feature was proposed to enrich the feature representation of weak texture region.The cumulative error and deformation of long sequence image mosaic are reduced by the binary tree mosaic method from sub-mosaic image to sub-mosaic image.The simulation results show that the improved algorithm is effective for the long sequence image mosaic of stator surface.In the second level,the stator surface lightweight defect detection network was designed.In this thesis,Mobile Net V3 was used as the backbone network of YOLOv4 network,and the 5×5 convolution operation in the backbone network was equivalent to two small convolution operations to make the model lighter.In order to improve the ability of the network to detect small size defects on the stator surface,a path aggregation structure composed of two feature pyramids was used in the detection network to enhance the sensitivity of the network to shallow features.By combining the convolutional channels in the cross-stage local connection structure and sharing the BN layer,the computational load of the network was further reduced.By combining the convolutional layer and the BN layer and replacing the Mish activation function with the hard-swish function,the reasoning speed of the network was further improved.Experimental results show that the improved network could ensure the accuracy and improve the detection speed significantly,but it still runs slowly on the embedded AI platform.Finally,channel pruning and residual block pruning were performed by taking the improved network as the benchmark network.Experimental results show that the speed of the lightweight network on the embedded AI platform is significantly improved after pruning,which can meet the requirements of real-time detection.At last,this thesis designs a two-level state detection system for stator surface in generator chamber,including hardware design and software design.And then,the thesis summarizes the whole work and determines the next optimization direction. |
