| Title |
A Multi-Class Semantic Segmentation Ensemble Technique for Implementing a Crimp Harness Inspection Device |
| DOI |
https://doi.org/10.5573/ieie.2024.61.7.87 |
| Keywords |
Wire harnesses; Crimp harnesses; U-net encoder; Ensemble technique; Deep learning |
| Abstract |
Wire harnesses are a key component for organizing and protecting wires and cables in electrical and electronic equipment. Poor wire harness quality is directly linked to product failure, fire, and personal injury, so quality control is key to preventing this. In the wire harness manufacturing process, quality inspection is especially important for crimp harnesses, which are the product of the crimping process that crimp terminals onto stripped wire cables. In this paper, we propose a U-net encoder-based ensemble method for crimping process quality inspection. The ensemble method combines several different U-net structures to produce a more accurate final prediction. The ensemble weight calculation algorithm determines the ensemble weights, firstly, through the average values of IoU, F1 score, recall, and precision of the U-Net models, and secondly, by penalizing the models with lower average values of IoU. The analysis of the ensemble method using crimped harness images obtained from a wire harness manufacturing plant showed that the discrimination accuracy reached 95.41%. The proposed method overcomes the shortcomings of existing methods and offers the possibility of reducing costs while maintaining consistently high quality. |