Abstract: A fault diagnosis algorithm based on motor noise signals and graph convolutional neural network (GCNN) is proposed for the needs of maintenance and noise monitoring and control of rotating machinery equipment. This algorithm transforms the time domain data into graph data by Fourier transform, and uses the proposed new graph convolutional neural network structure to train and classify the graph data. The experimental platform of motor fault is built, and the collection and experimental verification of motor noise signal in 6 different states are completed. The experimental results show that GCNN can effectively identify motor faults according to the limited motor noise signals, and has a certain small sample learning ability, and can classify faults under the condition of a small sample size. Comparative analysis shows that the accuracy of this algorithm is better than other algorithms such as K nearest neighbor graph algorithm, 1D convolutional neural network, automatic encoder and support vector machine, which provides a reference for practical engineering applications.