Research On Online Detection Method Of Metal Parts Vibration Crack Based On Computer Vision

Metal parts are widely used in many fields of industrial scene,such as manufacturing and robot industry,because of their high mechanical strength,plasticity and durability.Metal parts are often used as load-bearing structures or connecting components in the system,which are subject to severe stress,and are prone to fatigue cracks and failure,seriously affecting the production efficiency and even threatening the safety of workers.At present,the fatigue damage detection of metal parts is mainly based on visual inspection,which has some problems such as strong subjectivity and high labor intensity.How to realize the automation of fatigue test of metal parts is an urgent problem to be solved.In this paper,the on-line detection method of metal parts fatigue crack is studied.The research content includes the following aspectsFirstly,the crack data set of metal parts in shaking table environment is constructed.The image data is collected in the actual production environment,and the data set is divided into time series fatigue crack data set and single frame fatigue crack data set,and the original image data set is annotated with image annotation tool.Then,a crack detection and segmentation algorithm based on traditional image processing method is proposed.On the premise of eliminating the influence of vibration,crack detection is realized step by step from the complex background,and then crack segmentation is realized by combining a variety of texture and geometric features.Finally,a deep learning algorithm for crack detection and segmentation based on mask r-cnn network is proposed.Aiming at the problem of network non convergence caused by less features in the crack annotation box,the original crack annotation box is enlarged to a certain scale,and the combination of crack and background is taken as a whole;a variety of data enhancement methods are used to enrich the data diversity;the ROI size in mask r-cnn network is increased to further improve the crack segmentation effect.In order to verify the effectiveness of the algorithm,the performance of the algorithm is tested on the self built fatigue crack data set.For the algorithm based on traditional image processing,the crack detection accuracy rate is 92.59%,the recall rate is 96.15%,the false detection rate is 7.40%,the missing detection rate is 7.40%,and the average absolute percentage error of crack length measurement is 14.40%;for the crack detection and length measurement algorithm based on deep learning,the improved mask With r-cnn algorithm,the average absolute percentage error of fatigue crack length measurement is reduced to 9.92%.The experimental results show that the two algorithms can complete the task of fatigue crack detection with high accuracy.