Physics-Guided Two-Stage Anomaly Detection Framework for Industrial Connector Fitting Verification
物理誘導型2段階異常検知フレームワークによる産業用コネクタ嵌合検査
- Delivery
- Available on this site
- Format
- Price
- Non-members (tax incl.)：￥1,100 Members (tax incl.)：￥880
- Publication code
- 20266130
- Paper/Info type
- Proceedings (Autumn)
No.119-26
- Pages
- 1-6(Total 6 p)
- Date of publication
- Oct 2026
- Publisher
- JSAE
- Language
- Japanese
- Event
- 2026 JSAE Annual Congress (Autumn)
Detailed Information
| Author(J) | 1) 伊藤 龍翔, 2) 佐々木 茉子, 3) 市毛 弘一, 4) 松尾 弘達, 5) 宇賀治 篤史, 6) 羽田 松太郎, 7) 阿部 浩二 |
|---|---|
| Author(E) | 1) Ryusho Itoh, 2) Mako Sasaki, 3) Koichi Ichige, 4) Hirotatsu Matsuo, 5) Atsushi Ugaji, 6) Matsutaro Hata, 7) Kouji Abe |
| Affiliation(J) | 1) 横浜国立大学, 2) 横浜国立大学, 3) 横浜国立大学, 4) 日産自動車, 5) 日産自動車, 6) 日産自動車, 7) 日産自動車 |
| Affiliation(E) | 1) Yokohama National University, 2) Yokohama National University, 3) Yokohama National University, 4) Nissan Motor, 5) Nissan Motor, 6) Nissan Motor, 7) Nissan Motor |
| Abstract(J) | コネクタ嵌合検査の不良流出ゼロを保証する物理誘導型2段階異常検知を提案する．第1段階はCWTスカログラムを用いたCAEとMSE-SSIMスコアで検知し，第2段階は独自物理量のフェイルセーフ層で見逃しを排除する．5種コネクタで誤検知ゼロ，未知異常条件下でも加重平均Recall 99.97%を達成した． Translation |
| Abstract(E) | We propose a physics-guided two-stage anomaly detection framework for industrial connector fitting inspection to guarantee zero-defect outflow. Stage 1 feeds CWT-based log-scalograms into a linear bottleneck convolutional autoencoder, suppressing over-generalization and scoring anomalies with a hybrid MSE–SSIM metric. Stage 2 applies a deterministic fail-safe layer evaluating intersection magnitude (IM) with a modified z-score, eliminating remaining false negatives. On five automotive connector types, Stage 1 achieved F1 = 99.36% on the most challenging connector and 100% on all others; Stage 2 completely eliminated false negatives. Under open-set conditions, recall was 99.97% versus 83.37% for CNN baselines. |
