The invention discloses a direction of arrival estimation method based on grid partial refinement. Through learning and fission processes, in the fission process, new grid points are generated to refine the grid. The learning process is continuously close to the direction of arrival, grids near the real direction of arrival are finely divided, and grid areas far away from the direction of arrivalare coarsely divided, so that partial refinement of the grids is realized, the estimation precision is ensured, and compared with a previous off-grid DOA estimation algorithm, the number of grid points is greatly reduced, and the calculation amount is reduced accordingly. According to the method, the information source number does not need to serve as a priori, grid division is as sparse as possible, the grid number is reduced, therefore, the calculation complexity is reduced, and the algorithm consumes less time. And under the condition of very sparse initial lattice point division, the estimation precision of the algorithm is ensured through minimum interval threshold self-definition.