![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
2024 TA PP ADAM PUTRA PRATAMA ZAINURI 1-ABSTRAK
Terbatas  Suharsiyah
» Gedung UPT Perpustakaan
Terbatas  Suharsiyah
» Gedung UPT Perpustakaan
Electrical Submersible Pump (ESP) is one of the commonly used artificial lift methods in the petroleum industry. In its process, ESP requires routine monitoring. This needs to be done to prevent any undesirable things that could happen during the lifting process by ESP. However, just like all other methods of artificial lift, ESP will inevitably fail during its lifetime. However, before failing, ESP will enter a condition called trip. Because of this, a process called troubleshooting is needed. The focus of troubleshooting itself is to identify and solve the failure which causes ESP trip. This identification and solving process requires constant monitoring of the ESP and sensors interpretation which was installed to the ESP. If troubleshooting fails, then ESP will be considered to have Downhole Problem (DHP). To replace the old design of ESP, the root cause of the ESP failure must be identified first. A process called Dismantle Inspection and Failure Analysis (DIFA) is conducted to resolve this problem. DIFA comprise of disassembly of ESP installation and ESP part testing. Compared to the troubleshooting process, DIFA is a much more tedious process. DIFA requires extensive time and resources. As an alternative, the use of Machine Learning (ML) could be an effective and efficient approach in predicting various failure information of an ESP installation by reading and interpreting real-time sensor data. This can assist in the troubleshooting process and potentially reduce the amount of DIFA needed.
The ML model which will be used in this method is a classification model which can read real-time sensor data and predict various failure information of an ESP installation. The model will classify 4 outcomes of an ESP installation using a hierarchy system, starting from the broader category, which is failure type, followed by more specific category, which are failure item, specific failure item, and detailed failure description. The model development process consists of data preprocessing, data splitting, data normalization, dimensionality reduction, model selection, and hyperparameter tuning.
Using a systematic workflow, an ML pipeline is developed, which could predict various failure information regarding an ESP installation by using 1 week sensor data, taken before the failure happens. This method proves to be able to predict failure information with acceptable accuracy. This method could be considered as a preventive method which is potentially more effective compared to the existing conventional methods.