Currently， applying machine learning techniques to anomaly detection of drainage pipe defects， e.g.， pipe leakage， has become the focus of urban intelligent management. However， monitoring data of drainage pipe network collected from the real-world scenarios contains much noise， particularly sudden water level fluctuations caused by rainfall， significantly reduce the accuracy and reliability of detect results. To address the above problems， a robust unsupervised multi-task anomaly detection method was proposed for drainage network defect diagnosis. Firstly， by integrating the spatial-temporal information from multiple physical monitoring stations， a deep multi-task Support Vector Data Description （SVDD） model was established， individual hypersphere-based one-class classifiers were established for each station to extract anomaly detection rules， thus constructing a rule adaptation mechanism to obtain the common feature representation of multiple stations. Secondly， based on the obtained feature representations， sliding windows were introduced into each station’s SVDD model to continuously identify abnormal fluctuations in the pipeline monitoring data， thereby determining noise points caused by common interference factors in the monitoring data sequences. These noise points were corrected by polynomial interpolation to exclude irregular noise interference caused by rainfall. Finally， the modified monitoring sequences were employed to detect pipe network leakage based on AutoEncoder （AE） reconstruction errors. Experimental results on the real-world monitoring data collected from 2017 to 2018 by Qingtan Water Management System in Changzhou City demonstrate that the proposed method is consistent with the hand-operated maintenance records， moreover， the proposed method has higher detection accuracy and lower false alarm rate compared with statistical methods and traditional machine learning approaches. Taking the Qingtan East area as an example， the false detection rate of the method proposed in this paper when dealing with rainfall interference is 5.47 percentage points lower than that of the suboptimal method USAD （UnSupervised Anomaly Detection）， significantly improving the robustness of the model in strong noise scenarios and further verifying the accuracy and practicability of the proposed method.