Fault tolerance is a key element for reliable data delivery in practical applications of wireless sensor networks (WSNs). Sensor node failure can be caused by many reasons including environmental effects, transmission link instability, failure of hardware component and battery drainage. Fault tolerance is of particular interest in multi-tier heterogeneous networks where relay nodes are used to transport the collected data from sensor nodes. In this paper we present a new approach that uses Bayesian Network model to compute the failure probability of relay nodes. Sensor nodes learn about the failure probability of relay nodes through regular updates. Moreover, an algorithm is proposed to find two disjoint paths for each sensor nodes in the network. Simulation results are presented to analyze the fault tolerance in different network configurations.