Addressing Domain Shift in Insulator Defect Data: A Generalization Framework for Cross-Domain Detection of Broken and Self-Blast Insulator Defect

Accurate and timely detection of insulator defects is essential for the safety and stability of the power system. However, current detection faces challenges of domain shifts arising from insufficient data that do not encompass most inspection scenarios. To address this challenge, we propose a robust generalization framework for insulator broken and self-blast defect detection involving domain generalization (DG) and domain adaptation (DA) methods. First, we synthesize high-fidelity insulator defect data in 3-D space using domain randomization (DR) techniques to create diverse variations termed DR-Syn. For the DG method, we extract invariant features across domain data using a domain expansion method based on our proposed instance-reweighted image quality assessment (IR-IQA) model and a proposed discrepancy-constrained invariant learning (DCIL) model in the training process. For the DA method, we proposed a digital-twin-aided DR-Syn model that incorporates the target domain background information for specific-domain data generation. Extensive experiments validate the effectiveness of our framework in mitigating domain shift. The basic DR-Syn data can perform better than real-world intradomain data training. The DG method outperforms the real-world data training model in $\textbf {mAP}_{50}$ of 4.2% and 5.3% in intradomain training and 13.9% and 29.3% in cross-domain validation. The DA method achieves additional performance gains of 15.7% and 17.9% enhanced with digital-twin background modeling. Detailed ablation studies verify the validity of our method.