Extended regression-based deep neural network with multi-task learning for fast source detection by sensor array

Aifei Liu

doi:10.1016/j.aeue.2025.155915

Extended regression-based deep neural network with multi-task learning for fast source detection by sensor array

2025-06-25

0

PRE

AI

DOI：10.1016/j.aeue.2025.155915

Original

Original request for help

Save

Abstract

En 中文

• By considering the direction of arrivals (DOAs) of sources are continuous, rather than discrete, we treat source number detection and DOA estimation as two extended regression problems. Thus, different from the conventional methods using two stages for the two tasks, we establish an extended regression model to simultaneously complete the two tasks, using the prior knowledge that an array of M sensors can identify (M−1) sources at most. The extended regression model is much more efficient than the other DNN-based solutions and classical subspace-based methods. • We implement the extended regression model with multi-task learning for the two tasks of source number detection and DOA estimation. Moreover, for the two branches of the two tasks in the multi-task learning, we design two specified labels with bounding values and arbitrary values, respectively. Then, we develop two specified loss functions that do not conflict with each other and thus avoid the negative transfer which is a curse in multi-task learning. • We develop the ER-DNN0 model with one branch for the detection of the number of sources and DOA estimation at the same time. Simulation results demonstrate the ER-DNN with two branches for multi-task learning performs better than the ER-DNN0 with one branch. Furthermore, it is illustrated that, different from existing solutions, the ER-DNN can identify the absence of source signals and provide better performance in terms of computational efficiency and robustness. • We generate each training data in the presence of array gain-phase errors which is randomly driven from a uniform distribution which helps the ER-DNN gain robustness against the array gain-phase error. Furthermore, we use the off-diagonal upper-right diagonal elements of the array covariance matrix as the input vector of the ER-DNN to gain immunity to nonuniform noise. • We use the real data collected by a uniform circular array (UCA) to further verify the generalization ability of the ER-DNN model trained with simulation data.

Extended regression-based deep neural network with multi-task learning for fast source detection by sensor array

Abstract

Journal

Organization