Scientists from India’s University of Petroleum and Energy Studies (UPES) have developed a machine-learning system that can not only detect methane leaks occurring in gas pipelines and at gas distribution stations but also automatically assess their scale. During tests using simulated data, the algorithm correctly identified the pipeline condition in 93.67% of cases, with a false alarm rate of 2.47%.


Today, gas pipeline leaks are identified via manual inspections, pressure analysis, acoustic sensors, fiber-optic systems and cameras that can detect methane emissions. However, these methods often require expensive equipment and physical involvement. Some are also ineffective at detecting small and slowly developing leaks. In addition, conventional systems generally cannot determine the severity of the problem automatically.

The Indian researchers have developed a system named MethaneGuard AI, which continuously analyzes four parameters: methane concentration, pipeline pressure, temperature and gas flow. Based on these data, it classifies the gas facility’s condition using four categories: normal operation, minor leak, moderate leak or major leak. This allows the operator to obtain information not only about the occurrence of a problem but also about its possible scale.


Since real data on gas pipeline accidents are hard to access, the scientists trained the system by creating a set of 1,000 artificially generated records simulating industrial sensor readings. Of these, 500 records corresponded to normal operation, 250 to minor leaks, 150 to moderate leaks, and 100 to major leaks. As the leak scale increased during the simulation, methane concentration rose as well, while the pressure and gas flow rate decreased. For instance, average methane concentration during normal operation was 48 ppm and the pressure was 102 psi, while during a severe leak, these indicators stood at 78 ppm and 84 psi, respectively.


With these data, the researchers tested several machine-learning algorithms. The best results were demonstrated by random forests, an algorithm that uses multiple decision trees and determines the final outcome based on their combined predictions. A total of 70% of the dataset was used for training, and the remaining 30% was used to validate the model.


The most important factor for leak detection was methane concentration, which accounted for 58.7% of the model’s feature significance. Other factors included pressure at 24.5% and gas flow rate at 13.9%, while temperature contributed only about 3%. This means that methane concentration and pressure jointly provided over 83% of the information the algorithm relied on for classification.


The system was most effective at recognizing major leaks: out of 30 cases in the test set, it correctly identified 29, with no other pipeline condition mistaken for a major leak. Detecting minor leaks proved most challenging: out of 75 cases, the system correctly identified 67, classifying four as normal operation and another four as moderate leaks. It took the algorithm mere milliseconds to process the readings, whereas conventional monitoring methods can take from several hours to days.


A web dashboard was also created for the system to display in real time all sensor readings, their changes and leak levels identified by the algorithm. If any deviations are detected, the system automatically alerts the operator. Later on, it will be possible to connect real sensors on gas pipelines and gas distribution stations to the system, as well as to integrate it with the industrial Internet of Things and digital twins.


As of now, the development remains a software prototype: the model was trained and tested only on artificially generated data, while real gas pipeline readings have more complex relationships and time-dependent variations. This is why testing with physical sensors and real operational data should come next.


Source: Global Energy