Alexandra Gunderson is a data scientist at Arundo, a software company enabling advanced analytics in heavy industry. She holds a Master of Science in computational methods and in biomedical engineering. She may be reached at firstname.lastname@example.org. Kristin Hornæs is a data scientist at Arundo Analytics. She holds a Master of Science in computational and applied mathematics. She may be reached at email@example.com.
In the global oil and gas sector, failure of pumps, compressors, heat exchangers and other critical equipment systems can have significant consequences for field operations, safety and revenue. Today, most equipment monitoring systems raise static alarms based on high/low thresholds for individual sensor values. These sensors are often inadequate for capturing the wide range of normal behavior of complex oil field equipment systems and unable to separate true potential failures from false positives.
Given the substantial impact of pump and other equipment failure, there is a need to complement existing systems using advanced analytics.
The convergence of several long-term technology creates new opportunities for failure prediction in critical equipment:
- Data storage and compute capacity are increasingly inexpensive and effectively unlimited.
- Sensors continue to decline in cost and physical footprint.
- Sensor, device and asset-level connectivity continue to improve in quality and cost.
- Machine learning tools and techniques are increasingly accessible and easy to use, even with limited training and resources.
The result is the emergence of sophisticated end-to-end equipment condition and performance monitoring and anomaly detection systems for pumps and other industrial equipment.
In oil and gas operations, machine learning models may provide earlier detection of failures and assist engineers and other field personnel in the diagnostic process for anomalous equipment behavior. Critical equipment could contain anywhere from five to 100 sensors.
Typical threshold-based alerts look at individual sensors in isolation; these are inflexible and could lead to false positives. The goal of advanced anomaly detection is to create threshold alerts based on a more complex view of equipment behavior that distinguishes among numerous types of normal operating modes. This significantly improves the fidelity of any resulting alerts.
For instance, one particular method of anomaly detection can learn behavior based on historical sensor data. With sufficient data, the algorithm can differentiate among typical operating modes (such as ramp-up, ramp-down, or various normal operating behaviors) by creating groups, or clusters, of sensor values over time. Modern software can consider multidimensional clusters that are typical of today’s equipment, accommodating dozens of inputs as frequently as once per second or faster.
Machine learning models such as this provide an early warning of abnormal equipment behavior with an indication of where the behavior is outside normal envelopes, allowing an engineer or maintenance personnel to troubleshoot and diagnose failures quickly and efficiently.
In one recent instance, an anomaly detection model for critical oil field equipment was tested against status quo maintenance practice to determine predictive value based on known historical failures. The model raised an alert two weeks in advance of status quo maintenance practices for an actual failure event. It also identified the most anomalous sensors as the pressure and flow sensors of the system, indicating the most likely sources of failure. In this particular event, advanced anomaly detection would have resulted in nine days of additional production, reducing the total lost production by 40 to 50 percent, a potential impact of $12 million to $16 million for this field.
End-to-End Monitoring for Pumps
A typical pump could have 10 or more individual sensors. These could measure individual values such as vibration (drive-end and non drive-end), flow rate, pressure (differential and suction), temperature (pump housing, pool, drive-end bearing, non drive-end bearing), speed and motor-absorbed power. In addition, advanced software can locally calculate values such as pump efficiency and pump head as a function of sensor values, and report those calculated values into an analytic engine at a similar sampling rate as streaming sensor data. Advanced software can also determine more complex outputs, such as runtime, number of starts or time in high vibration. All of these values (and others, depending on the pump application) may vary widely during normal operating behaviors over a wide range of field conditions and operating modes. With advanced monitoring software systems, users can visualize the individual sensors and calculated or modeled values in intuitive dashboards. Anomaly detection systems can also be included in such applications.
With advanced software and robust anomaly detection models, pump systems vendors, owners and operators are able to provide early warning, diagnose failures more quickly and take the appropriate remedial actions. The value at stake is enormous for companies in the oil and gas industry, but the combination of advanced sensors, cheaper data storage, cheaper compute capacity and more easily accessible machine learning techniques enable field equipment to communicate with other systems and human operators in ways that were unimaginable when such equipment was originally designed. Advanced analytics offer significant benefits for adoption. In the rapidly evolving world of the Industrial Internet of Things, those slowest to adapt to this new reality will be disadvantaged.