In the φ-OTDR sensing system, separating overlapping multi-source events is a significant challenge, particularly in scenarios with limited training samples. Limited training data can lead to overfitting of the mask estimator, making it difficult to isolate the desired components from the composite signal. Unlike conventional methods that rely on data augmentation to compensate for data scarcity, we propose an adaptive separation strategy. The proposed approach leverages existing datasets and mask estimators to iteratively refine poorly separated signals, thereby enhancing the quality of the separated results. Theoretical analysis and experimental results provide comprehensive evidence demonstrating the feasibility and portability of the adaptive separation strategy across various scenarios. Thus, this method proposes a novel solution for multi-source separation in φ-OTDR systems under data-limited conditions, demonstrating promising potential for practical applications.