Although mud pumps are critical rig equipment, their health monitoring currently still relies on human observation. This approach often fails to detect pump damage at an early stage, resulting in non-productive time (NPT) and increased well construction cost when pumps go down unexpectedly and catastrophically. Automated approaches to condition-based maintenance (CBM) of mud pumps to date have failed due to the lack of a generalized solution applicable to any pump type and/or operating conditions. This paper presents a field-validated generally applicable solution to mud pump CBM. Field tests were conducted during drilling operations in West Texas and Japan, to verify the feasibility of the developed pump CBM solution. An accelerometer and acoustic emission (AE) sensor were attached to pump modules, and data was collected during drilling operations. Anomaly detection deep-learning (DL) models were trained during run-time to pinpoint any abnormal behavior by the pump and its elements. The models were trained only with normal state data, and a damage score characterizing the extent of damage to the mud pump was calculated to identify the earliest signs of damage. The system correctly identifies the degradation of the pump and produces alerts to notify the rig crew of the damage level of key mud pump components. During the field tests, different hyper-parameters and features were compared to identify the most effective ones for identifying damage while at the same time delivering low false positive rates (i.e., false alarms during normal state pump operation). The developed CBM system thus provides a generalized solution for pump monitoring, capable of working for different pumps and different operating conditions, and only requires several hours of normal state data with no prior pump data information. This system eliminates the environmental, health and safety (EHS) concerns that can occur during human-based observations of mud pump health, and avoids unnecessary NPT associated with catastrophic pump failures. The final version of this system is expected to be a fully self-contained magnetically attachable box containing sensors and processor, generating simple indicators for recommending pro-active pump maintenance tasks when needed. This is the first successful attempt to validate a universally applicable DL-based CBM system for mud pumps in the field. The system allows more reliable continuous and automated pump monitoring by detecting damage in real-time, thereby enabling timely and pro-active mud pump maintenance and NPT avoidance.

Unexpected failure of mud pumps during drilling operations can result in non-productive time (NPT) and increase well construction cost. Several prior studies and implementations of condition-based maintenance (CBM) systems for mud pumps have failed to provide a generalized solution for the variety of pump types encountered in the field, in particular by failing to detect damage early enough to mitigate NPT. Our research is aimed at improving upon this situation by developing a practical, generally-applicable CBM system for mud pumps. In the study reported here, a laboratory test bed with a triplex mud pump was used to collect data to test a new approach to mud pump CBM. Artificial damage was introduced to the two most frequently replaced parts of the pump, i.e., the valve and piston. An accelerometer and an acoustic emission (AE) sensor were used to collect experimental data. Based on this data, an anomaly detection algorithm was constructed using a one-class support vector machine (OC-SVM) to pin-point the early onset of mud pump failure. The CBM methodology thus developed does not require prior knowledge (data) of the mud pump itself or of the failures of its components. This is key to it being more widely deployable. The trained machine-learning algorithm in the test setup provided an accuracy greater than 90% in detecting the damaged state of the valve and piston. Only the characterization of the normal (i.e., non-damaged) state data was required to train the model. This is a very important result, because it implies that the sensors can be deployed directly onto mud pumps in the field – and additionally, that the first few hours of operation are sufficient to benchmark normal operating conditions. Also, it was observed that a multi-sensor approach improved the accuracy of detection of both the valve and piston damage. The system is able to detect early-stage damage by combining the cumulative sum control chart (CUSUM) with the damage index developed in this project. This work is the first attempt at applying semi-supervised learning for CBM of mud pumps. The approach is applicable for field use with very little or no prior damage data, and in various working conditions. Additionally, the system can be universally deployed on any triplex pump and efficiently uses the data collected in the first few hours of operation as a baseline. Consequently, the practicality and scalability of the system are high. It is expected to enable the timely maintenance of critical rig equipment before catastrophic damage, failure and associated downtime occurs. The system has been deemed promising enough to be field-trialed, and is currently being trialed on rigs in North America.