Abstract: Subspace direction of arrival estimation algorithm and sparse reconstruction DOA estimation algorithm are driven by algorithm model, and have poor robustness to algorithm modelled errors such as antenna array position deviation and incident source statistical characteristic deviation. The DOA estimation algorithm based on deep learning is data-driven and can effectively improve the robustness of the algorithm to model errors. However, in the DOA estimation algorithm based on supervised deep learning strategy, a large number of labeled training data sets generated by known standard signal sources are required, which affects the estimation performance of the algorithm for unknown sources and is not conducive to practical application. To solve the above problems, this paper proposes a weighted eigenvector-based unsupervised DOA estimation algorithm (WEUDA) using unsupervised deep learning strategy. The WEUDA algorithm establishes the unlabeled training data set according to the weighted eigenvector of the multi snapshot data covariance matrix of the random incident source received by the antenna array of the unknown direction, so that the algorithm no longer depends on the known standard signal source, which is more convenient for engineering implementation. At the same time, it also improves the estimation performance of the network on the unknown test data set. Simulation results show that compared with subspace class, sparse reconstruction class and supervised DOA estimation algorithm, the proposed WEUDA algorithm has better estimation performance under the same simulation conditions.

Key words: Direction of Arrival, Deep Learning, Unsupervised Learning Strategy, Weighted Eigenvector, WEUDA