In this paper we study the problem of incremental fault diagnosis and repair of mechatronic systems where the task is to choose actions such that the expected cost of repair is minimal. This is done by interleaving acting with the generation of partial conditional plans used to decide the next action. A diagnostic model based on Bayesian Networks is used to update the current belief state after each action. The planner uses a simplified version of this model to update predicted belief states. We have tested the approach in the domain of troubleshooting heavy vehicles. Experiments show that a simplified model for planning improves performance when troubleshooting with limited time.