The winding looseness fault is one of the main mechanical faults in transformers, and there is still a lack of effective intelligent diagnosis methods. Therefore, a diagnosis method for such a fault based on Gramian angular field and transfer learning-AlexNet is proposed. The periodic characteristics of the vibration signals during steady-state operation of transformers make it difficult to construct a sufficient set of images with time correlation. Therefore, a sample construction method is proposed to generate the Gramian angular field image set of transformer vibration signals. The generated image set is sent to AlexNet for transfer learning to obtain the fine-tuned neural network model. The image set generated by the sample construction method is used as the training and validation set, and the experimental results are that training and validation accuracy of the convolutional neural network model established are both above 99%. The image set generated by the periodic vibration signal of the transformer is used as the test set, with a testing accuracy of over 99%, achieving accurate diagnosis of transformer winding looseness faults. It also provides a way for constructing a sufficient sample set using time-related image transformation methods for periodic signals.