딥러닝 기반 머신비전을 활용한 전기차용 PFC 자동 품질검사 시스템 개발
Development of Automated Quality Inspection System for Electric Vehicle PFC Using Deep Learning-Based Machine Vision
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study develops an automated quality inspection system for Patterned Flat Cable (PFC), a critical lightweighting component in electric vehicle batteries. PFC's unique physical characteristics—thin structure (0.1~0.3mm), high surface reflectance (35~55%), and flexibility—cause traditional manual and rule-based automated visual inspection methods to suffer from 10~15% and 5~8% false detection rates, respectively. To address these challenges, we constructed a dataset of 122,315 real manufacturing images with standardized defect classification guidelines and developed a hybrid deep learning system employing an EfficientNet-based classifier for circuit inspection and a two-stage pipeline combining binary classification with YOLOv5 for pad defect detection. Deployed on NPU-based cloud infrastructure, the system achieved F1-scores of 0.98 and mAP@0.5 of 0.618 with 152ms inference time, improving inspection throughput by 74.8× over manual inspection while reducing annual operational costs by 165 million KRW. Currently operating in two production lines, this validated framework demonstrates immediate industrial applicability and provides an extensible architecture for automated quality inspection in electric vehicle lightweight component manufacturing.
Abstract
This study develops an automated quality inspection system for Patterned Flat Cable (PFC), a critical lightweighting component in electric vehicle batteries. PFC's unique physical characteristics—thin structure (0.1~0.3mm), high surface reflectance (35~55%), and flexibility—cause traditional manual and rule-based automated visual inspection methods to suffer from 10~15% and 5~8% false detection rates, respectively. To address these challenges, we constructed a dataset of 122,315 real manufacturing images with standardized defect classification guidelines and developed a hybrid deep learning system employing an EfficientNet-based classifier for circuit inspection and a two-stage pipeline combining binary classification with YOLOv5 for pad defect detection. Deployed on NPU-based cloud infrastructure, the system achieved F1-scores of 0.98 and mAP@0.5 of 0.618 with 152ms inference time, improving inspection throughput by 74.8× over manual inspection while reducing annual operational costs by 165 million KRW. Currently operating in two production lines, this validated framework demonstrates immediate industrial applicability and provides an extensible architecture for automated quality inspection in electric vehicle lightweight component manufacturing.