This work presents a data driven approach for pipe leaks classification, validated on a steel carbon pipe section conditioned with leaks of different sizes and locations in order to emulate abnormal conditions. The tested structure was instrumented with piezoelectric devices attached at different locations over the surface, in order to induce guided waves and to record its behaviour along the structure. For each experiment, one piezo device is excited by means of a high frequency burst type signal and the other ones are used as sensors. A blind bridle is connected to one of the extremes and an air source is coupled to the other extreme to emulate operational conditions. Statistical indices of correlated piezoelectric signals are obtained by using principal component analysis to distinguish different leak scenarios. Next, a selforganizing map is used to classify them. The experimental results show an improvement of the classification-learning rate when correlated signals are used instead of uncorrelated ones