Objective To address the issue of bearing failure in fruit puree packaging machines operating in complex work environments.Methods This study proposes a method of bearing fault detection of fruit puree packing machine based on multi-channel depth residual graph neural network. First, it uses the Cmor wavelet to complete signal preprocessing, and converts the original one-dimensional bearing fault signal samples into two-dimensional time-frequency graph signals. Second, the study designs the multi-channel feature extraction module, which reduces the number of parameters and enhances the data interaction between each channel, so as to effectively extract the two-dimensional time-frequency spectrum information of the bearing fault signal of the fruit puree packaging machine. Third, it proposes the residual graph neural network to fully learn the bearing fault characteristics from the two-dimensional time-frequency domain data of the bearing fault signal of the fruit puree packaging machine. Finally, the study completes the sample classification of faulty bearings by using the bearing fault classification part. The accuracy, visualization, experimental comparison, and performance analysis under variable conditions and noise are conducted on the proposed method and the comparison method.Results The overall bearing fault accuracy of the proposed method reaches 96.95%, which is higher than 86.43% of WSVMD-net and 91.9% of OSCFD-net. When the signal-to-noise ratio (SNR) decreases from 12 to -4 dB, the bearing fault diagnosis rate of the proposed method only decreases from 95.24% to 91.35%.Conclusion Compared with WSVMD-net and OSCFD-net, the proposed method is more suitable for variable conditions and noisy environment, obtaining higher accuracy and lower loss value.