麦克风阵列缺陷下基于SAE-CNN的DOA估计

SAE-CNN based on DOA estimation under microphone array imperfections

对于高精度波达方向(DOA)估计方法来说,麦克风阵列缺乏对各种阵列缺陷的适应性是一个难以解决的问题.基于卷积神经网络(CNN)的DOA估计算法具有不依赖阵列拓扑结构先验假设的优势,且与基于模型的方法相比,能更好地适应阵列缺陷.首先利用SAE对麦克风阵列输出的协方差矩阵进行预处理,将信号分解为多个空间子域,这些空间子域具有比原始输入更集中的分布,有助于减少后续DOA估计的泛化负担;然后将空间子域作为CNN的输入来训练神经网络,建立学习特征与DOA估计的非线性映射关系.相比传统的DOA估计算法,基于SAE-CNN的DOA估计算法在麦克风阵列缺陷下具有更强的准确性和适应性.

The lack of adaptability of microphone arrays to various array imperfections is a difficult problem for high-precision direction of arrival(DOA)estimation methods.The DOA estimation algorithm based on convolutional neural network(CNN)has the advantage that it does not rely on the prior assumptions of the array geometry,and it can better adapt to the array imperfections compared with the model-based method.First,SAE was used to preprocess the covariance matrix output by the microphone array,and decompose the signal into multiple spatial subdomains,these spatial subdomains have a more concentrated distribution than the original input,which helps reduce subsequent generalization burden estimated by DOA.Then the spatial subdomain was used as the input of CNN to train the neural network,and the non-linear mapping relationship between the learned feature and the DOA estimation is established.Compared with the traditional DOA estimation algorithm,the SAE-CNN based DOA estimation algorithm has stronger accuracy and adaptability under microphone array imperfections.