Abstract
In this paper, we present results of field trials using a novel machine learning (ML) pipeline leak detection algorithm based on the distributed temperature sensing (DTS) technology. The installation is based on a buried, 9km-long pipeline in Germany. The transported medium is wet oil with a temperature varying in the range of ~20°C–35°C. A multimode fiber optical cable (FOC) is laid at ~10 cm distance from the pipeline and is connected to a Raman-based DTS system. For verification, the leak test setup is composed of a water tank, heater, pump, and an injection pipe introduced on top of the pipeline at an unknown location for the DTS system, thus performing blind testing. The water was heated to 33°C (TEST1) and 19°C (TEST2) to simulate the temperature range of the transported medium and injected at a release rate of 36 L/h (TEST1) and 20 L/h (TEST2). The DTS system was adjusted to automatically alarm when the evaluated temperature deviation reaches 3°C with 3 confirmations. The ML alarm evaluation method was able to detect the anomaly at the test position after only 0.8 barrel (TEST1) and 1.4 barrel (TEST2) of total released fluid volume. In comparison, classical gradient-based approaches would require up to 5 times higher alarm threshold, which may not be reached due to the low temperature difference of the oil to the ground.
