Traditional operation and maintenance knowledge base does not have the ability to identify the failure phenomena in the image. Therefore, the knowledge base cannot handle the problem of unstructured data. To tackle this issue, based on fault classification networks in deep learning, an improved capsule network feature extraction structure based Caps-DRFN algorithm is proposed, which can realize automatic classification of operation and maintenance images of electromechanical equipment. Firstly, aiming at the multi-noise problem of operation and maintenance images, the deep residual shrinkage networks (DRSN) are introduced to improve the feature extraction performance of the model on noisy data. Subsequently, for the multi-scale problem of actual shooting operation and maintenance images, through the combination of the feature pyramid networks (FPN) algorithm, the Caps-DRFN realizes image multiscale feature fusion and improves the accuracy of model classification. Finally, the vector neuron is constructed by using the capsule structure, and the digital capsule of the classification structure is obtained through the feature transmission method of dynamic routing. The model realizes the fault classification of electromechanical equipment. The experimental results show that compared with the traditional capsule network algorithm, the accuracy of the proposed Caps-DRFN algorithm based on feature fusion is increased by 15% and it is more robust.