This study presents a cutting-edge soft sensing approach for coke-making diagnostics, aimed at tackling the challenges posed by multifaceted, nonlinear, non-Gaussian, and noisy operational data prevalent in coke-making ovens. Our proposed method leverages a Bayesian t-distributed mixed regression model, effectively capturing the intricate nature of multivariate, nonlinear, and non-Gaussian data. The utilization of the t-distribution ensures the model’s resilience to interference, with model parameter estimation achieved within a Bayesian framework. Conducting simulation experiments and real industrial experiments, as well as comparative analysis with PLSR, GMR, and GPR models, we demonstrate the model’s good robustness, excellent prediction accuracy, and robustness, further confirming its potential application in coking diagnosis.