YOLO-PAS: Detection of Mixed Impurity Particles in Transformer Oil Based on Lens-Free Holographic Imaging and Improved YOLO

Assessing the quality of transformer oil is critical to maintaining the insulation of transformers in the power system. However, identifying impurities can be challenging due to their small size and diversity. A new method for detecting mixed impurities in transformer oil based on lens-free holography and an improved YOLO-PAS algorithm is proposed in this article. A lens-free holographic system collected the hologram dataset for mixed impurities (bubble, copper, and cellulose). For YOLO-PAS, in the feature extraction stage, four C3_P modules based on partial convolution (PConv) are first adopted to reduce the extraction of irrelevant background features and concentrate solely on pertinent particles. Besides, the adaptive spatial feature pyramid network (ASFPN) is introduced to improve the model’s ability to adapt to multiscale features and focus on the target area quickly. This includes an adaptive spatial fusion (ASF) module and a parameter-free attention mechanism called SimAM. After cross-validation, YOLO-PAS showed significant improvements over YOLOv5s with an 8.6% increase in mean precision, a 7.4% enhancement in mean average precision (mAP), an 8.2% improvement in mean recall, and an 8.0% advancement in mean F1-score. After the Mann-Whitney U test, it was confirmed that the difference between YOLO-PAS and other target detection models was statistically significant (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$p \lt 0.001$ </tex-math></inline-formula>).