# Domain-Adaptive Direction of Arrival (DOA) Estimation in Complex Indoor Environments Based on Convolutional Autoencoder and Transfer Learning

## TL;DR

This paper introduces a new method for accurately estimating signal source directions in complex indoor settings using deep learning techniques.

## Contribution

The novel approach combines a Convolutional Autoencoder with Domain-Adversarial Neural Networks to adapt to new environments with minimal labeled data.

## Key findings

- The proposed CAE-DANN method outperforms existing methods in DOA estimation accuracy.
- The model effectively reduces domain discrepancies using Gradient Reversal Layer and MMD loss.
- High precision is achieved in new environments with minimal labeled target data.

## Abstract

Direction of arrival (DOA) estimation for signal sources in indoor environments has become increasingly important in wireless communications and smart home applications. However, complex indoor conditions, such as multipath effects and noise interference, pose significant challenges to estimation accuracy. This issue is further complicated by domain discrepancies in data collected from different environments. To address these challenges, we propose a deep domain-adaptation-based DOA estimation method. The approach begins with deep feature extraction using a Convolutional Autoencoder (CAE) and employs a Domain-Adversarial Neural Network (DANN) for domain adaptation. By integrating Gradient Reversal Layer (GRL) and Maximum Mean Discrepancy (MMD) loss functions, the model effectively reduces distributional differences between the source and target domains. The CAE-DANN enables transfer learning between data with similar features from different domains. With minimal labeled data from the target domain incorporated into the source domain, the model leverages labeled source data to adapt to unlabeled target data. GRL counters domain shifts, while MMD refines feature alignment. Experimental results show that, in complex indoor environments, the proposed method outperforms other methods in terms of overall DOA prediction performance in both the source and target domains. This highlights a robust and practical solution for high-precision DOA estimation in new environments, requiring minimal labeled data.