Leak detection is critical for the integrity management of oil and gas pipelines. The pipeline leak can cause a major accident, especially when transporting dangerous substances. The impact to the environment and human life is paramount and thus it is essential to detect the pipeline leak in time. Usually, a leak signal from the acoustic online monitoring sensor is characterized and identified by its waveforms, absolute amplitudes, and the frequency-domain energy distribution. However, these features are not steadily available due to the propagation attenuation under varied pipeline transportation conditions. In addition, sample leak signals are needed for most existing feature extraction and modeling methods, but the actual leak signals are seldom available. Although artificially simulated leaks can be adopted alternatively, it is not possible to fully duplicate the actual leak signals with complete features. To solve these problems, this paper proposes a pipeline leak detection approach by using time-domain statistical features from acoustic sensors. These features are extracted and vectorized from normal (no leak) sample signals, which are selected by an automated method. The size of the extracted feature vector is further reduced with principal component analysis method. A support vector data description model is built with the processed vectors as the input. The proposed method has been implemented in a field leak detection system. The experimental results from the field tests demonstrate the effectiveness of the proposed method.