Research On Detection Of Metal Rolling Contact Fatigue Based On Swin Transformer

In the machinery,automobile and aviation industries,many mechanical parts such as rollers are in a rolling contact environment for a long time,which makes the surface of the material easy to evolve and produce various forms of defect fatigue.Defects not only affect the normal use of mechanical equipment,but also pose potential safety hazards.The rolling contact fatigue testing machine can simulate the working environment of the roller.The detection and quantification of the surface defects of the roller based on the testing machine is of great significance to the failure mechanism research and actual production application.Traditional vision methods rely on feature engineering,and the upper limit of accuracy is not high,and they are easily affected by external environmental factors such as lighting.Based on a more effective visual Transformer architecture,this thesis realizes defect instance segmentation and area quantification of pitting and peeling defect categories,and records the fatigue generation and evolution process under the testing machine.The main research work is as follows:(1)A visual inspection system based on a rolling contact fatigue testing machine was designed,including the construction of an image acquisition device,the establishment of a data set,and the optimization of the model.The collected images are preprocessed and labeled to complete the establishment of the data set.Due to the imbalance of pitting and peeling categories in the data,the method of cropping and mirror expansion is used to balance the category labels and reduce the impact of category imbalance on model performance.In order to make the model more robust to different lighting environments,the data is enhanced with random lighting and contrast.(2)To realize the precise location and quantification of fatigue defects,an instance segmentation model based on the Swin Transformer is modeled.In the original model,the window shift division strategy is single,and the patches in different windows cannot fully interact,especially based on the horizontal and vertical areas at the non-center of the feature map,which makes the model lack the ability to represent the characteristics of defects.Therefore,this thesis proposes a window shift partitioning strategy based on the horizontal and vertical directions of the center of the feature map,which aggregates patches towards the periphery of the center of the feature map into windows under horizontal and vertical paths for self-attention computations,enhancing model’s feature extraction abilities in these regions.Due to the inconsistency in the number and size of front and rear windows due to window shift division,an efficient matrix shift method is designed to maintain the same computational complexity.Finally,combined with the Cascade mask module,the entire model architecture is proposed: Cas-VSwin Transformer.The model has a stronger ability to extract defect features and also has excellent performance on general images.Subsequent experiments combined with the migration learning strategy showed that the performance on the established defect data set was 82.3%(AP box)and 80.2%(AP mask),and the average relative error of pitting and spalling area quantification was 2.26%.At the same time,experiments were carried out on the public defect data set to verify its applicability.(3)The proposed model is sufficient to complete the task of defect detection,but it fails to show sufficient superiority compared with other model performances.Deepening the network layer of its model is a simple way to improve the accuracy,but this will increase the running cost of the model.The optimal strategy is to improve its performance without increasing the number of model parameters.To achieve this goal,this thesis is based on the feature map distillation method MGD(Masked generative distillation),and designs the noncentral area constraint of the feature map,and proposes SMGD to reduce the invalid distillation of the non-defective area(smooth area)on the roller surface in the image.The SMGD strategy improved the model by 1.6% and 0.8% on AP box and AP mask,respectively.