A transformer defect detection method based on state space model and cross-domain feature fusion
Power transformers serve as critical equipment in power systems, and appearance defect detection is essential for ensuring grid safety. This paper proposes a power transformer defect detection method based on state space model and cross-domain feature fusion. First, a multi-scale wavelet frequency enhancement module is designed, which extracts multi-scale frequency domain features through three-layer Haar wavelet decomposition and achieves adaptive complementary fusion of frequency-domain and spatial-domain features via cross-domain gating mechanism. Second, a dual-branch state space modeling module is proposed, which captures global long-range dependencies through multi-directional selective scanning of the Mamba architecture, combined with multi-scale convolution and ECA channel attention to extract local detail features, achieving complementary enhancement of global context and local information. Finally, a multi-scale linear attention module is introduced based on a dual positional encoding mechanism to enable efficient global context modeling. Experimental results demonstrate that the proposed method achieves an mAP50 of 84.7%, representing an improvement of 12.9 percentage points over the baseline method, validating the effectiveness of the proposed approach.
Authors
- Yunhan Li
- Pengfei Zheng (ORCID: https://orcid.org/0000-0003-3921-2468)
- Nan Wang (ORCID: https://orcid.org/0000-0002-4369-0115)
- Weihua Zhou (ORCID: https://orcid.org/0000-0003-2018-265X)
- Jingjing Lou (ORCID: https://orcid.org/0000-0002-8723-5931)
Institutions
- Taizhou Vocational and Technical College (CN)
- Huzhou College (CN)
- Shenyang Jianzhu University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1038/s41598-026-78578-z
- Primary Topic
- Industrial Vision Systems and Defect Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00