Open Access

ARTICLE

Research on Hybrid Unsupervised–Supervised Learning Fusion Method for Defect Detection of Stamped Parts

Zhikai Chi¹, Jiaxu Ning^1,*, Delong Zhang¹, Changsheng Zhang^2,3

1 School of Information Science and Engineering, Shenyang Ligong University, Shenyang, China
2 School of Computer Science and Engineering, Ningxia Institute of Science and Technology, Shizuishan, China
3 School of Software, Northeastern University, Shenyang, China

* Corresponding Author: Jiaxu Ning. Email:

Computers, Materials & Continua 2026, 89(2), 88 https://doi.org/10.32604/cmc.2026.085757

Received 17 May 2026; Accepted 24 August 2026; Issue published 15 September 2026

Abstract

Aiming at the problem that factory stamping parts have various types of defects, random locations, different sizes, and both known and unknown defects, it is difficult for traditional single inspection methods to achieve both accurate classification and generalized identification capabilities. To this end, the Hybrid Unsupervised Learning-Supervised Learning Fusion Defect Detection (HUSLFDD) model is proposed. The model adopts a dual-branch shared backbone network architecture, in which the supervised learning branch focuses on the accurate classification of known defects, and the unsupervised learning branch realizes feature capture and identification of unknown defects. The weighted fusion of the dual-branch results is completed through the image scorer, and fixed validation-calibrated parameters is achieved, ultimately taking into account detection accuracy and generalization capabilities. A large number of experimental results show that the proposed model has an F1 score of 91.1%–94.2%, a recall rate of 94.4%–100%, and an image-level AUROC of 80.6%–95.6% on four comprehensive test sets, which is significantly better than detection methods of single supervised learning, single unsupervised learning, and a simple combination of the two. At the same time, the model detection speed reaches 9.14 FPS, meeting the needs of industrial real-time detection. This research provides efficient and reliable technical support for quality control of stamping parts in the automotive manufacturing industry.

---

Keywords

---