In the process of ethylene production by steam cracking, the coking diagnosis of the furnace tube of the cracking furnace is of great significance. Due to the existence of multimodal, nonlinear, non-Gaussian, and strong noise characteristics of the operating data of the cracking furnace, this paper proposes a soft sensor for coking diagnosis using bayesian t-distributed mixed regression modeling, which realizes the effective characterization of the multimodal, nonlinear, and non-Gaussian data through the hybrid model, and the model's parameter estimation is completed by the VBEM algorithm under the Bayesian framework to guarantee the anti-interference ability of the model. distribution to ensure the anti-interference ability of the model and the parameter estimation of the model is accomplished by the VBEM algorithm under the Bayesian framework. Finally, through simulation experiments and real industrial data experiments, as well as comparative analyses with PLSR, GMR, and GPR models, the model is verified to have good robustness, excellent prediction accuracy and robustness, which further confirms its potential application value in the diagnosis of furnace tube coking.