A robust and mixture form of Bayesian latent variable regression with structural sparsity (BLVR-SS) is proposed to deal with an inferential sensing problem for the processes with multiple modes and noise components. In particular, a structural sparsity on the joint loading matrix is provided to automatically determine a subset of latent variables. By inducing the column-wise hyperparameter vectors, the dependency of quality variables on process variables can be reliably captured to make the estimation more accurate. In addition, the Student's-t distributions are placed on the latent variables for diminishing the adverse effect of outliers. A mixture form of BLVR-SS is then developed to separate the data into different operating modes. Subsequently, a variational Bayesian expectation maximization algorithm is developed for the parameter learning procedure of the model. The feasibility and efficiency of the developed inferential model is demonstrated through a numerical simulation, the benchmark Tennessee Eastman process, and a pilot scale 1.5 kg/h fluidized catalytic cracking setup.