Abstract: Transformers are critical devices in power systems, and accurate online detection of transformer faults is of vital importance for ensuring the safe operation of power systems. In response to issues such as severe occlusion between components, small target sizes, and complex backgrounds in transformer acoustic image-based online detection methods, this paper proposes a multi-model fusion-based online detection method for transformer faults. The model consists of two parts: the Mask R-CNN (mask region-based convolutional neural network) and the improved Retina-Net network. The Mask R-CNN network is used to segment the main body of the transformer in the acoustic image, and the segmented transformer main body is further subjected to target detection using the improved Retina-Net network. The experimental results demonstrate that this method can be adopted to effectively segment the transformer's main body and accurately identifythe faulty areas, achieving a detection accuracy of 96.1% for transformer fault locations. The paper proposes a multi-model fusion algorithm to take full advantage of deep learning, significantly improving the accuracy and efficiency of online detection of transformer faults, and exhibiting high application value in the field of online detection and intelligent recognition of power grid equipment faults.