The precise segmentation of electrical equipment is a key step in infrared image fault diagnosis. An accurate segmentation method is proposed for infrared images of electrical equipment in complex backgrounds to address the issue of detail loss with mainstream semantic segmentation methods. Firstly, the PSPNet is improved by incorporating UNet network as the main structure to decode the multi-scale pyramid pooling of features extracted by UNet's top layer. Secondly, Convolutional Block Attention Mechanism (CBAM) is integrated into the feature extraction backbone network to incorporate channel and spatial attention mechanisms for gathering image context information from both dimensions, enhancing the network's focus on electrical equipment to improve its anti-interference capability. Finally, the PSPnet-CBAM-Unet network is constructed, and the features output by the CBAM are used as inputs for lower-level feature extraction and skip-connection features in the decoding layer. The effectiveness of this paper's method is tested with the segmentation of three types of devices in infrared images under complex backgrounds including voltage transformers, current transformers, and circuit breakers. Experimental results demonstrate that the proposed method achieves intersection over union and accuracy greater than 92% and 94% respectively, and the accuracy of segmentation is better than that of UNet, PSPNet, and Deeplabv3+ networks, and it is more accurate for the detail segmentation of infrared images of electrical equipment in a complex background.