| Title |
Development of Crimp Harness Inspection Model for Consumer Electronics based on Multi-class Semantic Segmentation AI |
| DOI |
https://doi.org/10.5573/ieie.2024.61.1.43 |
| Keywords |
Wire harness; Crimp harness inspection; Multi class semantic segmentation; U-Net |
| Abstract |
Wire harness is defined as groups of wires providing interconnections for arbitrary electrical circuits that serve as the bloodstream in consumer electronics, electric vehicles, and autonomous cars. Poor wire harness quality is directly related to produce product failures, fires, and human casualties. This paper proposes a method that utilizes multi-class semantic segmentation techniques to determine the defects of the crimp harness, which is the result of a product of the wire harness crimp process. The problem of defect detection of crimp harness can be solved by accurately measuring the height and width of five consecutive segments from the preprocessed harness data. With this insight, we aimed to develop an AI model suitable for multi-segment identification based on U-Net, a representative semantic segmentation model. To identify multiple segments, we proposed a model that modifies the encoder part of U-Net using Resnet 34, EfficientNet B1, and Mix Transformer B0. For AI model development, images of crimp harnesses collected from a wire harness manufacturing plant were used to build the dataset. The developed multi-class semantic segmentation AI model showed a discernment accuracy of 95.14% on the test dataset. Through the method proposed in this paper, it is possible to maintain uniform high quality and reduce labor costs, which improves the shortcomings of the existing crimp harness quality inspection(manual, crimp sensor measurements, and rule-based image processing). |