Pipeline leakage is a demand from governmental and environmental associations that petroleum companies need to comply. Recent accidents with Petrobras pipelines increase local demand for leakage detection system. Due the high accuracy on detecting leakage required from that system is necessary to set a procedure that once applied will achieve the best performance considering the quality of the installed instrumentation. This paper describes a procedure to set such system in order to accomplish with the legal requirement keeping high reliability during normal and failure operations conditions. Nuisance alarms are kept at low value while minimum leakage detection is too small. To do that the described system uses a set of models acting as specialists each one observing and diagnosing pipeline leakage. This system also validates the operations according to the business rules. System uses a set of tools, fuzzy logic, neural network, genetic algorithm and statistic analysis, to execute its function. The usage of an optimization tool, genetic algorithm in this case, helps the designer to set a function alarm that uses a statistical approach to assure a reliable performance when detecting the leakage and keeping the nuisance alarm closes to zero. Both qualities make the system highly reliable since once it generates one alarm there is a likelihood of almost a 100% that the event is true. Instead of using the common two parameters alarm, threshold and timing, this system uses pattern recognition to verify the fault or leak condition. The detectable leakage value is function of the difference between the flow measurement at the inlet and the outlet of the pipeline. The minimum leakage detectable is constant and equal to 1.4 times the standard deviation of the error between this two meters for 0.2% of nuisance alarm. In the application it is able to alarm when a leakage of 2% of the total flow happens in a time bellow 5 minutes. If allowed a level of 5% of nuisance alarms the system is able to detect a leakage of one standard deviation of the error. That represents the mentioned amount of 1.4 times the standard deviation of the error. The system is in operation supervising pipeline in a Brazilian installation.