As one of the most widely used resonance demodulation methods in planetary bearing fault diagnosis technology, modulation signal bispectrum demodulates the fault feature information from the resonance band via inhibiting the background noise and interference frequencies. In addition, modulation signal bispectrum has the characteristics of preserving phase information, detecting secondary phase coupling signals, and removing Gaussian noise. Therefore, the ability to detect amplitude and phase modulation is crucial for fault feature extraction of modulation signal bispectrum. To address the challenge of analyzing non-Gaussian noise in modulation signal bispectrum, an autoregressive model filter based on non-Gaussian noise suppression is proposed to improve its performance in planetary bearing fault diagnosis and monitoring. Autoregressive model filters effectively capture key features of data series and are applied with superior performance in removing non-Gaussian noise. Therefore, the autoregressive model is considered as a pre-filter process unit to reduce non-Gaussian noise in the original signal to improve the accuracy of modulation signal bispectrum. The order of the autoregressive model is determined adaptively using an indicator called kurtosis, further improving the effectiveness of the autoregressive model. Finally, the non-Gaussian noise analysis model signal is processed using modulation signal bispectrum to remove Gaussian noise and decompose coupled modulation components, thereby accurately identifying the frequency components of planetary bearing faults. The simulation and experimental analysis results demonstrate that non-Gaussian model modulation signal bispectrum achieves higher accuracy in diagnosing planetary bearing fault characteristics than fast kurtogram and modulation signal bispectrum.