Abstract: Lack of fault data is always an important factor restricting the development of equipment fault diagnosis. Existing researches collect fault data by deliberately damaging equipment. In order to realize nondestructive fault diagnosis of seat motor, the fault mechanism of seat motor is analyzed to determine the possible types of faults in this paper. The occurrence of fault is simulated by sticking micro-speakers pasted on the surface of seat motor to play fault sound. On the basis of self-encoder system, the convolution operation is introduced, and the fully connected layer is replaced with convolution layer. The model structure is adjusted with the input data dimension, the size and number of convolution kernels, pooling and regularization. IDMT Isa Electric Engine data set is used as the source domain data to pre-train the model. Then, the learned data distribution in the source domain is transferred to the seat motor fault diagnosis task, and the detection results of various models are compared. The results show that under the condition of keeping the recall rate 1.00, the area under the curve (AUC) of the proposed method reaches 0.86. It manifests that the detection result is reliable, and the proposed method could be practically applied.