The denoising effect of the negative pressure wave signal and the extraction of the feature vector are the key factors affecting the accuracy of the oil pipeline leakage detection. Aiming at the false negatives and false positives in pipeline leak detection, this paper proposed an improved fully integrated empirical mode decomposition algorithm (improved CEEMDAN) with adaptive white noise to preprocess the negative pressure wave signal. The CEEMDAN decomposition is performed on the negative pressure wave signal measured by the upstream and downstream pressure sensors of the pipeline to obtain a plurality of intrinsic mode functions (IMF). And the effective IMF component is selected according to the correlation coefficient principle of the dual channel sensor. An entropy⁃based eigenvector is proposed, and the energy entropy, kurtosis entropy and permutation entropy of the effective IMF component are input to support vector machine (SVM) to distinguish different working conditions. Through field data verification, the improved CEEMDAN combined with the entropy⁃based feature vector can effectively improve the accuracy of oil pipeline leakage condition identification, and has certain field application value.