We propose a simple and easy-to-apply unsupervised training method for multi-channel deep separation models used in sound source separation. Such models require a large amount of training data, i.e., source signals and their mixtures. A previous method uses pseudo-target source signals, which can be obtained as the outputs of blind source separation (BSS) based on statistical models in place of ground-truth source signals. However, the model performance of the previous method is degraded by some pseudo-targets that are inadequately separated by BSS. To exploit the reliable part of BSS, we select and remix well-separated signals included in the BSS result. In the selection step, we choose well-separated signals using the direction of arrival (DOA). As a criterion that addresses the quality of the separated signals, we adopted the minimum angular difference of DOA between source signals. In the remixing step, we introduce resampling of the DOA, which generates mixtures composed of source signals with both wide and narrow angular differences. These mixtures are not simply given by BSS and allow the deep separation model to learn both spectral and spatial information. In our experiment, our method's model performance was improved for mixture signals composed of the sources from various angles.