Reducing Unplanned Maintenance Shutdown Using Predictive Analytics

Reducing Unplanned Maintenance Shutdown Using Predictive Analytics
  • August 11th, 2021


Carbon-di-oxide (CO2), one of the primary impurities in Natural gas is commonly removed with the use of Acid Gas Removal Unit (AGRU). Breakthrough in Acid Gas Removal Unit (AGRU) is one the most frequent problems faced while removing impurities from feed gas in the Oil and Gas industry. This breakthrough can sometimes cause an additional maintenance shutdown to remove CO2 residual – stuck in the plant machinery, hence deferring production.

Predictive analytics is being used by energy companies to anticipate plant failure before shutdown occurs. This method ensures the right measures are taken to avoid unplanned shutdowns and reduce maintenance costs.

Mu Sigma worked with the Asset Development team of a leading Oil and Gas company to explore ways to reduce the rate of Lean amine used to remove CO2.

The Problem

One of the largest Oil and Gas companies approached Mu Sigma to explore methods that would help them predict the AGRU – CO2 breakthrough. The objective of the project was to automate the process of monitoring and adjusting the gas flow which when done manually in the past – resulted in significant production downtime. The client wanted to provide their production engineers with a forward notification of 10 mins or more, allowing them to take preventive measures well in advance and reduce the lean amine used to remove CO2.

The Mu Sigma Approach

We took on a systematic approach of problem solving to gather a comprehensive understanding of the business problem.

Stated problem representation

muPDNA – one of Mu Sigma’s design thinking frameworks helped us in making the decision of using derived values in cases where data was absent which in turn enhanced the predictive model

The Solution

1. Dataset Creation & Hypotheses testing

•   Since the data captured by the sensors within the plants was of very high volume, we needed to resample the data multiple times to arrive at an optimal frequency.
•   After extensive hypothesis testing, we narrowed down the ideal interval for alerts to be between 10-30 minutes

2. Overcoming Hurdles

2.1 Unavailability of data for the class that needs to be predicted

•   One of the main problems while building a predictive modeling framework was that the ratio of the event class to the non-event class was much smaller which led to very little data available to predict the event accurately.

•   To overcome this barrier, Mu Sigma team came up with a solution framework that trained the model using additionally generated synthetic data for the event class which in turn enhanced the overall accuracy of the prediction.

2.2 High False alarm counts

•   Another major challenge was that the number of false alarms which were initially quite large. This in turn would have reduced the reliability of the predictive model. To overcome this, the Mu Sigma team trained the predictive model leveraging a probability optimization algorithm due to which we were able to reduce the false alarm count significantly.

•   In the end, the team was successfully able to build a Neural-Net-based predictive model which could predict the occurrence of a CO2 Breakthrough event at least 10 minutes in advance with very few false alarms and a true positive rate of 86%.

2.3 Redefine Solution Deployment

•   Once the predictive model was developed, we encountered a change in requirement where the solution needed to be deployed in an Advanced Process Control System (APC). To accommodate this change, the model needed to be converted into the native GCC compliable language of that system.

•   The Mu Sigma team developed an extensive conversion pipeline which enabled us to efficiently covert and deploy our model into the Advance Process Control System’s native compliable language.

3. A Model Monitoring UI tool

•   The addition of an intuitive UI layer on top of this solution empowered the production engineers to analyze various key metrics contributing to the events and helped them identify the key drivers behind them.

The Impact

The advance predictive modeling algorithm enabled the production engineers to get a forward notification for the CO2 Breakthrough events at least 10 mins in advance in real-time and minimize the maintenance shutdowns. This led to the saving of deferral costs upwards of $30 MM per year.