A New Knowledge Gradient-based Method for Constrained Bayesian Optimization