Research On The Defect Detection For Turbocharger Casting Based On Deep Learning

With the stable development of the automotive industry,the demand of turbochargers is increasing,and the detection of casting defects in the production process is becoming more and more important.However,most manufacturers still use manual inspections to identify and mark defects,which is inefficient and complicated.In addition,differences in the subjective perceptions of inspectors lead to the inability to guarantee long-term high-level inspection work,so it is difficult to meet the large-scale production requirements of turbochargers currently.Therefore,researching an efficient and accurate intelligent detection algorithm to replace manual detection has very important theoretical significance and application value for the turbocharger industry.This paper proposes a fast and accurate method of turbocharger casting defect detection based on deep learning technology.The technology involved in the research of turbocharger casting defect detection algorithm is analyzed theoretically.The performance and characteristics of different object detection algorithms are compared and studied.According to the requirements of defect detection tasks,YOLOv3 is used as the basic algorithm for defect detection,and the detection algorithm is studied based on the characteristics of casting defect detection.In order to reduce the influence of the noise of defect sample pictures collected in the production site,comparing the peak signal-to-noise ratio,structural similarity index and processing speed through experiments,the bilateral filtering algorithm is selected to preprocess the defective sample pictures.For purpose of meeting the data size requirements of deep learning algorithms,a random cropping translation method based on a priori labeling information is proposed,and a small sample data enhancement is realized in combination with color space transformation.The problems of long path and poor transmission of deep information in the detection part structure in YOLOv3 are analyzed.Considering the large difference in size of casting defects of the turbocharger and the variety of shapes,a detection network based on multi-path fusion is adopted to further enhance the fusion of features of different scales and reduce the loss of information during the transmission from shallow to deep layers.Based on the thought of transfer learning,a two-stage training program is designed,and the position regression evaluation criteria and category label values when calculating the loss function are adjusted to optimize the training process.The feasibility of the proposed method is verified through experiments.For all defect objects,the recall is 97.9% and the precision is 96.7.%.Finally,in order to facilitate the detection of subsequent defect records and data analysis,based on the appearance characteristics of different types of defects,post-processing methods are designed specifically.