Research And Deployment Of Small Object Detection Algorithm For Bearing Surface Defects Based On Deep Learning

Bearing surface defect detection is an essential process to ensure the safe operation of machinery.Damaged bearings can be found in time through defect detection to reduce economic losses.In complex industrial environment,there are often different degrees of overlap and occlusion between bearings,which greatly improves the difficulty of detection of small target defects on the bearing surface.The traditional method of feature template extraction can no longer meet the detection requirements of many hidden small targets,so this paper chooses the method based on deep learning to complete the detection of bearing surface small target defects.The current research methods on deep learning are generally real-time and still have a high detection rate for small targets,and most of them have not been deployed to mobile terminals for automatic detection.In order to solve the shortcomings in the task of defect detection,the research work of this paper is as follows:(1)The bearing images with multi-form and multi-surface defects between bearings are collected,the characteristics of various bearing defects are analyzed,and the sample number of bearing defect images is expanded by rotation,flip,affine transformation,mosaic and other methods.(2)Based on YOLOX,this paper proposes a small target defect detection algorithm based on weighted fusion of multi-attention features.The Res2 Block module with more finegrained feature extraction is introduced into the backbone network,and the self-attention mechanism is embedded to increase the regional features of hidden small targets and reduce the missed detection rate.A two-way pyramid feature fusion network with embedded coordinate attention as a weighting condition was designed to improve the interactive fusion ability of shallow detail features and deep high-level semantic features.In the detection head network,the Inception parallel information transmission module was introduced to parallelize high-dimensional features and improve the forward inference speed of the model.In the post-processing stage,the Focal Loss function is introduced to increase the learning of the model for positive sample targets and further reduce the missed detection rate.(3)In order to realize automatic detection,the improved YOLOX model is deployed to the Android mobile phone.The ONNX conversion tool was used to convert the trained parameter file,and then it was quantized.Then,NCNN architecture is used to extract and segment the quantized files,and generate model files and parameter files.Finally,the debugging and deployment of the APP are completed based on Android Studio and the dependencies of various libraries.Experimental results show that compared with the original YOLOX algorithm,the m AP of the improved algorithm on the self-made small train bearing surface defect dataset is increased by 4.04%,and the detection ability of small target defects on the bearing surface is significantly improved.Compared with the research methods in related fields,the improved algorithm has better detection accuracy and meets the detection speed of embedded devices.The YOLOX Demo APP deployed on Android still has a good detection effect and 29f/s real-time performance for bearing surface defects in the case of introducing interferers.