Aggregation of Conflicting Data in Neural Network-based Sound Source Localization

Abstract

Sound source localization is an important task today, especially in the context of the active development of unmanned aerial vehicles. The rapid spread and cheapening of their technologies have created a critical need for the development of reliable detection and tracking systems. Although deep learning-based approaches demonstrate higher accuracy compared to classical methods in complex conditions with strong noise or reverberation, sound source location data obtained from different microphone arrays can be incomplete and contradictory. Given this, there is an urgent need to apply aggregation methods to combine these spatial probability distributions into a single consistent estimate, which will significantly increase the reliability of acoustic detection networks. This study focuses on solving the problem of determining the azimuth of a drone based on multichannel audio recordings. To achieve this goal, a convolutional neural network based on the ResNet-18 architecture was adapted, which is capable of processing spatial matrices generated using the generalized cross-correlation with phase transform method. The classification model achieved a high accuracy rate of 93.75%, confirming the effectiveness of using cross-correlation matrices as input data for spatial localization tasks. To solve the problem of conflicting predictions from independent expert systems, the following probability aggregation methods were tested in this study: linear, geometric, and multiplicative opinion pooling functions, the Dempster-Shafer rule of combination of evidence, and the Dezert-Smarandache proportional conflict redistribution rule. The experimental results showed that while linear pooling effectively preserves the multimodality of distributions, it may not assign the highest probability to the correct spatial sector in situations where the peak predictions of individual sources do not match. On the other hand, the multiplicative rule and the Dempster-Shafer rule successfully isolate a single maximum value during aggregation, but the latter can lead to false-positive conclusions when working with highly conflicting data. The Dezert-Smarandache method proved to be the most reliable solution for environments characterized by strong acoustic noise and correspondingly high model uncertainty, as it proportionally redistributes the conflicting mass without giving premature preference to a single incorrect class. Furthermore, geometric pooling demonstrated significant practical value by finding a compromise between competing estimates and providing better computational efficiency in scenarios with a moderate level of uncertainty. The obtained results form comprehensive recommendations for selecting appropriate data aggregation strategies adapted to specific environmental conditions.