The key technique for neural network based pipeline leak detection is features extraction. In this paper，piezoelectric dynamic pressure based leak signal transient was chosen as research object，the differences between leak signals of upstream and downstream under different operational conditions were analyzed，and leak signal enhancement with wavelet decomposition was introduced. Positive and negative interval divisions of dynamic pressure signal was proposed. The differences of weighted signal sums，signal mean values，signal peaks of every two successive intervals were selected as the features of leak signal，and their calculations and relative scaling transformations were presented. The feasibility criteria for lengthwise and breadthwise evaluation of leak signal features were presented，and evaluation was done with features extracted from field data and the feasibility was verified. At last，a neural network based leak diagnose model with both features from upstream and downstream and its training，testing results were given. Long term and real time monitoring of pipeline leak showed that the features extraction methodology proposed here had reasonable operational conditions adaptability，and provided an encouraging technical support for robust diagnosis of pipeline leakage.