The risk of leakage poses a grave threat to natural gas pipeline safety. The high compressibility of gases combined with unsteady boundary conditions makes detecting leaks in pipelines a challenging endeavor. To date, in the literature, only a limited number of studies have focused on leak detection and diagnostics in gas mixture pipelines. The present study provides a system for detecting, locating, and estimating the size of small gas leaks from a compressible and dynamic natural gas flow in pipelines with improved accuracy. As a case study, a long natural gas pipeline of 80 km is simulated with leak sizes of 0, 2, and 5. The safety system is developed using mass flow rate, temperature, and pressure measurements. Six classes for faulty cases and one class for no fault case were considered for the study. A shallow neural network classifier (SNNC) is trained to identify a specific fault class. The SNNC is based on a two-layered network with 20 and 7 neurons. An input vector of 15 variables is provided to the system, and the output is one of the seven possible classes. Leakage as low as 2 at various locations are correctly diagnosed with more than 99 correct classification rate. © 2022 American Institute of Chemical Engineers.