This document proposes a neural network architecture called B-PR-F for fusing redundant location estimates from sensors on underwater robots operating in unmodeled noise conditions. B-PR-F models the underwater environment and sensor behaviors, uses contextual anticipation to predict sensor deviations, evaluates sensor reliability, and fuses estimates with weighted sums. Simulations in Gazebo demonstrate how B-PR-F can eliminate noise from virtual sensors by incorporating behavioral context and ordered neighborhood relationships between estimates. Experimental results show B-PR-F outperforms Kalman filters for localizing an underwater robot using sonar and DGPS sensors.