Research On Visual Positioning And Detection Algorithms Of Engine Crankshaft Bearing Caps Assembly Robot

Crankshaft bearing caps(CBCs)are vital components automobiles,which act as the role of strengthening the integration of crankshaft and cylinder block.The close integration of CBCs and cylinder blocks guarantees the smooth engine operation,so the assembly of CBCs is particularly important.In the assembly process of CBCs,intelligent feeding requires industrial robots to accurately locate and detect the placement direction and state of CBCs through image processing and analysis methods.In order to improve the efficiency of engine CBCs industrial assembly,reduce costs and improve reliability,this paper combines deep learning technology to propose a deep learning-based engine CBCs feeding robot visual positioning and detection algorithm.The main work of this paper is as follows:1.Establish and expand the CBCs data set.Perform data cleaning and manual labeling of data collected on the production site.After obtaining the coordinate information,it is transformed into an accurately labeled CBCs image data set.Since there are fewer defective samples in actual production,this paper expands the samples.The generated defect samples can be used for training of detection algorithms.2.Propose a CBCs visual localization and detection method based on attention mechanism.In the VGG part of Faster R-CNN’s feature extraction network,a two-layer attention network is introduced to make the network structure more focused on the characteristics of CBCs itself.In addition,the number of candidate frames and the aspect ratio are improved,which reduces redundant calculations.At the same time,the accuracy of candidate frame regression is improved.Finally,a loss function is designed based on GIOU,which improves the accuracy of positioning.3.An improved visual localization and detection method for fully convolutional single-stage CBCs is proposed.Based on the idea of pixel-by-pixel dense prediction,the network structure of the full-convolution single-stage target detection method is improved,and the characteristics of the corresponding layer in the feature pyramid network are strengthened.Besides,deformable convolution is introduced in the classification and regression branch,which expands the receptive field and extracts more suitable features for positioning and detection.In order to further improve the positioning accuracy,on the basis of GIOU,the degree of centroid overlap in the actual grabbing process is also taken as an optimization item,and a loss function is designed.The proposed visual positioning and detection algorithm of the engine CBCs loading robot finally realized can detect the target positioning and placement direction of CBCs.On the 606 CBCs data set established in this paper,the positioning accuracy can reach a cross-match ratio of 97%,and the detection accuracy and recall rate are both above 99%.