| The electric wires terminal is the conductive element that connects the wire harness to the external conductor.The electrical and mechanical properties of the wire harness depend on the tightness of the bite between the terminal and the wire harness.Crimping uses the compression force generated by the die to deform the wire and the connecting devices,so as to achieve a tight engagement.The image of terminal section can reflect the state characteristics of terminal and wire core after crimping,in order to avoid damage to a large number of raw materials due to poor crimping,cross-section analysis is an important task in the quality inspection of electric wires terminals.The existing terminal section analysis method mainly relies on manual use of image processing-based section analyzer for data extraction,the built-in algorithm of the instrument is backward and incomplete,resulting in poor measurement accuracy and low work efficiency,detection is time-consuming and labor-intensive.In recent years,deep learning has achieved remarkable results and has been widely used in industrial manufacturing,promoting the development of industrial intelligence.For the two subtasks of wire core counting and area extraction in the work of electric wires terminal cross section analysis,The research designed models for different tasks to fully exploit the practical value of deep learning in the image analysis of electric wires terminal cross section.The main research contents and conclusions are as follows:(1)An improved Efficient Det model based on a gradient harmonized mechanism is proposed for the wire core automatic counting task.The wire cores in the microscopic image of the electric wires terminal show a dense honeycomb shape and are difficult to count.In order to avoid overfitting the model by focusing too much on difficult samples during the training process,the gradient norm is used as a measure of the difficulty attribute of sample classification,the gradient harmonizing mechanism is introduced to coordinate the degree of contribution of samples with different attributes to the model,and the K-Means multidimensional clustering algorithm is used to optimize the prior bounding boxes generated by the initialization of the model to improve the matching degree between the prior bounding boxes and the external rectangular box of the wire core.The experimental results show that the improved Efficient Det wire core detection model in this paper has an average accuracy of 96.2% and a high accuracy of wire core counting,which is suitable for the task of automatic wire core counting in electric wires terminal section analysis.(2)An electric wires terminal cross section segmentation model with Mobile Net V2 and FCN co-representation is proposed.Considering the limitation of model size in real environment deployment,the lightweight Mobile Net V2 network and the fully convolutional network are fused to build the Mobile FCN semantic segmentation model,and the learning strategy of transfer learning fine-tuning is applied to migrate the model pre-trained on the Cityscapes dataset to the terminal section segmentation task.The mainstream semantic segmentation models are selected in the experiments to explore the model performance of different models for electric wires terminal cross section segmentation after transfer learning.The experimental results show that the mean Intersection over Union of the fused Mobile FCN model is slightly decreased,but the number of parameters is reduced by 80.4% compared with FCN,which is suitable for the subsequent task of extracting the geometric quantities in the electric wires terminal cross section.In this paper,there are a total of 35 figures,9 tables,and 85 references. |
