Our client is a company that specializes in providing comprehensive solutions for the oil and gas industry. With a wide range of products and services, this client aims to optimize the production and operations of its clients in this sector. They offer expertise in areas such as production chemistry, artificial lift systems, drilling fluids, and well stimulation.
KUNGFU.AI, an AI partner with Google, was brought in to investigate opportunities to ultimately improve profitability under the Unbridled Electrical Submersible Pump (ESP) program for the client. Over the course of roughly 6 weeks, KUNGFU.AI performed three separate phases of work:
- Alignment and advisory: the team held a number of subject matter expert (SME) interviews with a variety of technical and non-technical stakeholders at Our client with the intention to refine and justify the project from a business perspective
- Exploratory data analysis (EDA): once the requisite data access was granted, the team performed unstructured data exploration to build familiarity with the domain as well as uncover data engineering requirements
- Proof of concept (POC) modeling: experimental models were trained on the objective uncovered during the alignment and advisory phase, with the intention of selecting the most performant architecture
Because this client covers costs associated with short-run (sub-90 day) electrical submersible pump (ESP) failures, the ability to predict and subsequently mitigate these failures would bring substantial value to the program. Doing so could at minimum limit potential damage, and in the best case prevent it altogether. KUNGFU.AI spoke with many representatives from this client close to Unbridled ESP, and ultimately determined this initiative as the most likely to succeed.
Assuming such measures prevented failures to 50% of ESPs, the client would save roughly $3M annually (based on 2022 losses) and customers could simultaneously improve well throughput by minimizing downtime. The engineering work associated with this project was focused on providing a proof of concept for the failure prediction component.
The KUNGFU.AI team learned of three unique data types available for performing predictive tasks.
- ESP metadata: Structured data about ESP parts and equipment, e.g. installation specs, constituent part types, failure dates, serial numbers, etc
- Time series data: Various time-stamped data related to well activity at approximately 5 minute intervals (intake pressure, casing pressure, intake temperature, etc.)
- Document metadata: Digital reports on well failures including structured data, unstructured data, as well as images
In order to maximize experimentation time, the first two were the main focus of the engagement.
Exploratory Data Analysis
Using the ESP metadata, we analyzed time to failure by cause. While most pumps fail due to reservoirs or fluids, the majority of these failures occur after 90 days. On the other hand, installation and facilities failures account for 11.4% of sub-90 day failures, and 4.5% of total failures.
Plotting out the time series allowed us to find trends between ESP measurements and failures. For example, we found that intake pressure tends to increase before solid failure, while tubing pressure often spikes inversely in these cases. In general, we found that there was a significant difference between ESP measurements before a failure and during normal operation.
We trained a decision tree model to predict ESP failure within different sized windows. The input data were 24-hour averages and other metrics of various pump features such as vibration and temperature. Using a 5-day window we were able to predict failure within 24 hours with an average F1 score of 0.74, which means that we were able to correctly predict failure the majority of the time.
While this result was worse than our deep learning based approach, we showed that even with 24-hour averages there was enough signal in the data to predict failure. Additionally, we were able to gain insight into what features carried the most signal of failure.
Convolutional Neural Networks
We trained a Convolutional neural network to predict ESP failure within 24 hours. The input data were random samples consisting of raw values of pressure, temperature, and vibration measurements. These samples were time series values with a length spanning 48 hours, in other words, given the data for the past 48 hours, predict the probability of failure within the next 24 hours. We achieved an F1 score of 0.87 on the validation dataset, which consisted of ESP pumps that were not included in the training dataset.
The overall dataset size was small, around 100 unique samples, therefore the F1 score achieved is impressive. We would expect the score to improve given more data, as well as more time to address data quality issues that were ignored in this phase.
At the conclusion of the project, we presented our findings to the data and product executives of the client, with promising results based on the limited data we were provided, alongside future work opportunities we can engage in with this client. More specifically, the lower bound for performance was determined to be an F1 score of 0.87 for predictions made 24 hours prior to ESP failure. Said another way, for every 100 detected failures, 15 false positives and 15 false negatives would be present. The team included steps to both improve and employ this approach in the final report.