Abstract: The distributed fiber optic sensor can sensitively monitor the local temperature changes caused by the leakage at any position of the pipeline. The mass monitoring data has obvious spatial-time nonstationary properties, so it is difficult to diagnose the leakage directly from the monitoring data. In this paper, a spatial-time big data analysis method based on moving-window outlier analysis was developed in the framework of statistical pattern recognition. By virtue of this method, the intelligent identification of pipeline leakage is realized only based on the intrinsic characteristic of the distributed temperature sensing data. Then, the method to determine the lengths of the moving window and the abnormal state diagnosis window was proposed. Finally, the physical simulation was performed on the leakage detection of prototype pre-insulated steel pipe. It is indicated that the method developed in this paper can never make the false alarm when the pipeline is in the intact state, and it can quickly identify the leakage event and accurately locate the leakage once a leakage happens. In conclusion, this method is an unsupervised artificial intelligence approach of big data analysis, and its application prospect in the leakage detection of buried pipelines is promising.