Accurate fault diagnosis of electromechanical equipment under the condition of limited label samples is of great significance for improving the health management ability of complex electromechanical equipment. This article proposes a semi supervised bidirectional generative adversarial network (S-BIGAN) based on dual attention mechanism for mechanical and electrical equipment fault diagnosis. GAF is used to convert one-dimensional data into two-dimensional images, effectively utilizing a small amount of labeled data and a large amount of unlabeled data. Finally, bearing data is used as the validation object, and compared with algorithms such as CNN SVM and SGAN, the accuracy of fault diagnosis is greatly improved.