Why did we build a usage-based fatigue model using GenAI technology?

At Prescient, we wanted to go beyond sensor-based Condition-based Maintenance (CBM). While sensor-based CBM can be effective in detecting potential failures, they do not understand how equipment performs. For example, a vibration sensor could tell us that something is vibrating abnormally, but it doesn’t tell us why. It catches the symptom, not the cause.

To understand the cause, we needed to build usage-based fatigue models. These models track the cumulative fatigue experienced by each equipment or equipment component and predict the Remaining Useful Life (RUL). For example, the model could tell us that the abnormal vibration is due to a valve reaching its end of life.

More importantly, these models allow us to understand how equipment performs across different variables like operating conditions, locations, operators, component types, makes, and material. This vastly increases our understanding and enables effective maintenance planning and optimization. For example, these models help us to:

• Enable planned interventions by predicting failures months in advance across an entire fleet

• Reduce maintenance costs by extending component life by quantifying component health

• Understand why some rigs spend much more on maintenance than others

• Choose best performing or most cost-effective components by removing variability from operating conditions and locations

There are several ways to build fatigue models. The traditional way is physics based, where methods such as multi-variate analysis and finite element analysis are used to model fatigue. This can be very complex and time consuming, and some physics are simply too difficult to model. Today, machine learning is increasingly used to build models. Machine learning is data driven, so as long as there are enough data to represent fatigue, the training process will figure out the model. However, machine learning has its own shortcomings, such as noise, bias, over/under-fitting (i.e. prediction accuracy), and care must be taken to build an accurate and robust machine learning model.

The power of machine learning depends on the model used, of which there are many. For us, this was a journey. We started with a Random Forest model, in which we tried to correlate the RUL of mud pump fluid-end components with their cumulative operational statistics such as cumulative pressure, speed, mud volume, mud quality, and etc. This worked well. We achieved better than 90% accuracy, which meant that 9 out of 10 times our predicted RUL fell within +/-5% of the actual RUL, measured on 120+ rigs. This was quite an accomplishment and an industry first.

However, this model ran into trouble when we tried to apply it to mud pump power-ends and top drives. These last much longer than the mud pump fluid-end components, so it would take years to accumulate enough training data (i.e. failures) to build the models. What could we do?

This was when we turned to Generative AI (GenAI) models. Remember we mentioned that the Random Forest model worked on the cumulative operational statistics? GenAI models worked directly on the real-time operational data. This allowed us to capture much more details in the data, and one benefit of this was that we needed fewer failures to train the model. To be clear, the total amount of data absorbed by the GenAI model was vastly increased, because now we are ingesting 1-second data, but the number of failures needed for model training was significantly reduced.

The GenAI model was also superior at exploring the relationship within the data and performing various analyses. For example, it could let us “translate” operations from one rig in one basin to another rig in another basin. We call this performance normalization. If a component lasted 1000 hours on a rig in Appalachia, how many hours would it have lasted on a rig in the Permian? The ability to normalize performance across rigs and basins allowed us to objectively compare component performance from different vendors. “This is the missing piece,” said the Vice President of Global Supply Chain for a major drilling contractor.

The ability of the GenAI model to ingest high-speed operational data is no longer a dream. This allows us to fully capture detailed signatures in high-speed time-series data over long periods of time, and it enables the industry to maximize the value of large-scale investments in data collection and storage. From surface drilling equipment to downhole tools to ESPs, GenAI is enabling us to build usage-based fatigue models for equipment that were too complex to model in the past.

If you are interested to learn more about our GenAI model, you can attend our presentation, “Transformer vs. Classical Machine Learning: Harnessing Normalization for Enhanced Asset Life Prediction” at this year’s SPE ATCE conference.