Abstract: Water inrush incidents constitute a component of coal mine disasters, and the accurate identification of roof water inrush signs is crucial for preventing accidents. However, the underground environment of coal mine is complex and changeable, and the signs of water inrush are mainly found by manual judgment at present. Due to the influence of harsh underground environment and subjective factors, it is difficult to find anomalies in time, which makes it difficult to effectively identify the signs of water inrush in the roof. In order to efficiently monitor and accurately identify water inrush signs, a self-supervised water inrush sign recognition method based on SAM-XMem is proposed based on image recognition and processing technology. This method employs pixel change rates to construct the water inrush warning system. It autonomously annotates images collected from water inrush-prone areas within coal mines and segments water inrush sign regions using the SAM model. By integrating the XMem long video segmentation framework, real-time dynamic tracking of these regions is achieved. The results demonstrate that compared with the OTSU algorithm, the self-supervised recognition algorithm based on SAM-XMem exhibits superior performance in terms of intersection over union (IoU), precision (P), recall (R), and F₁ score, with evaluation metrics exceeding 90%, which represents 20% improvement over the recognition rate of conventional algorithms. The problem of data set shortage in coal mine is overcome by zero sample segmentation technology. Compared with the traditional method, the feature extraction ability is stronger, and the effect is better under complex conditions such as large background noise and small gray value difference. It is suitable for underground coal mine environment, and has higher recognition accuracy and stronger generalization ability.