Multimodal collaborative perception for transformer oil leakage detection

In order to address the challenges of transformer oil leakage detection, such as complex background environments, severe interference from illumination shadows, and large variations in leakage scale, this study proposes a transformer oil leakage detection method based on multimodal collaborative perception. First, a channel similarity gated fusion module (CSGFM) is designed. By computing channel-wise similarity between dual-branch features and introducing a gating mechanism, CSGFM adaptively aggregates discriminative features while effectively suppressing background noise and redundant information. Second, a tri-stage cascade cross-attention module is introduced. Through a parallel image-text bidirectional cross-attention strategy, this module exploits prior textual knowledge about oil leakage locations and illumination shadows to achieve precise localization of leakage regions. Finally, a hierarchical multiscale feature fusion module is developed. By employing a wide-span feature pyramid and a secondary fusion strategy, this module effectively balances fine-grained feature extraction for small-scale leakage targets with global contextual perception for large-scale leakage areas. Experimental results on a self-constructed dataset demonstrate that the proposed method achieves a mean intersection over union of 69.86% and a mean Precision of 84.38%. Compared with existing methods, the proposed approach significantly reduces false detections caused by shadow interference and provides robust and reliable oil leakage detection performance, indicating its strong potential for intelligent inspection and condition assessment in substations.