With the rapid development of smart grid systems, image fusion technology has received wide attention in order to improve the accurate location of thermal faults. In this paper, the visible and infrared images of substation electrical equipment are used as the research object, and the network model is designed by deep learning method with auto-encoder as the backbone network, in which the encoder adopts the designed dual-branch network structure of densely connected branch and enhanced branch, one branch is densely connected branch using dense block connection and self-attention mechanism to extract edge and detail features, and the other branch is enhanced branch using a strengthened branch with an improved feature pyramid network (FPN) structure to enhance the global information. In this paper, two sets of corresponding features are obtained by the dual-branch structure, and the features are fused by the L1-parametric fusion strategy and input the decoder to reconstruct the fused image. After comparing with various methods, the method in the paper verifies the advancement of the algorithm in terms of subjective visual evaluation and objective image fusion evaluation indices, in which the objective evaluation indices QMI, SSIM and FMIpixel are 0.567 26, 0.593 47 and 0.887 60, respectively, reaching the highest value, which proves that the quality of fused images is improved and applicable to multi-source electrical equipment image fusion.