<p>This research addresses Multiple Uniform Linear Array sonars (MULA), where each Uniform Linear Array sonar (ULA) has several signal receivers that exist along a linear array. Using the receiver measurements, each ULA measures the conical angle of signal emanated from a 3D source. We consider the realistic case where the conical angle measurement noise can increase as one increases the distance between the source and a ULA. This paper considers a scenario where a ULA is attached to the body of an Autonomous Underwater Vehicle (AUV). All AUVs are sparsely deployed at various sea depths, and they are connected utilizing multi-hop communication links. All AUVs perform station-keeping maneuvers for measuring a 3D source signal. One cannot use Global Navigation Satellite System (GNSS) for localization of an underwater ULA. Due to sea currents, the location error of a ULA can increase as time goes on. It is assumed that the upper bound for location error of a ULA is known in advance. Considering the error upper bound, this research derives a 3D source localization based on a neural network, which is trained utilizing virtual measurements at various viable source positions. Since using a neural network does not require an initial source position guess, the proposed localization method is suitable for target localization in a workspace which is extremely large. To the best of our knowledge, our research is novel in addressing a 3D source localization based on conical angle measurements of MULA. In addition, this study is novel in computing a 3D source localization utilizing multiple ULAs with heterogeneous sensor position errors. The outperformance of the proposed 3D source localization scheme is demonstrated under computer experiments.</p>