Aiming at the problem that the feature quantity is difficult to select in the fault diagnosis of transformer winding looseness and relying on manual experience, a diagnosis method of transformer winding looseness based on automatic encoder noise reduction, gramian angle field (GAF) and depth residual network (ResNet) recognition is proposed. The method automatically learns effective fault features from GAF images without manually extracting the feature quantity. Firstly, the vibration signal is denoised through an automatic encoder to obtain a vibration signal with a higher signal-to-noise ratio. Then, the GAF method is used to convert the vibration signal into a two-dimensional image and generate an image dataset. Based on this, ResNet is trained to construct a network model suitable for classification and recognition of transformer winding looseness faults. Finally, a transformer winding looseness fault test platform is built to collect vibration signals of the winding under different looseness and experimental currents for analysis. The experimental results show that the proposed diagnosis method has an accuracy of over 95% in identifying transformer winding looseness, and can effectively identify the looseness phase and degree. It is suitable for identifying and diagnosing transformer winding looseness faults.