Blog
The future of maintenance and control effectiveness
  • linkedin icon
  • twitter icon
  • facebook icon
  • youtube icon
  • instagram icon
Prabu Parthasarathy
Vice President of Applied Intelligence

The continuous advancement of technology has enabled intelligent solutions for decision-making and data collection mechanisms. Leveraging Artificial Intelligence (AI) and Machine (ML) solutions, organizations can maximize the utilization of their operating equipment and enhance productivity.

We will experience more technological progress in the next decade than we did in the last 100 years combined. Businesses that take advantage of advances in technology will benefit the most, in terms of return on those investments. While some industries may find that a formidable task, Industry 4.0 technologies are already playing a decisive role in the future of digital operations. The future is closer and smarter than you might think.

The ability to understand the root causes of asset failures before they occur shouldn’t have to be after-the-fact guesswork. The use of AI is already making an impact, by predicting needed maintenance or repairs to functional equipment before a failure can occur. For example, we can send an alert to a client that their assets have been running at hazardous operating levels for a long time, increasing the risk they may fail well in advance of failure. The ability to obtain information, like this – in advance of potential breakdowns – enables operators to make needed changes to reduce any risks to their assets.

Even as the cost of new technologies is decreasing, recent advances in machine learning and deep learning techniques are enabling us to build smart systems that leverage knowledge from operating data and simulation models at a rapid pace. This enables edge solutions that deliver higher productivity and increased throughput.

Machine learning and predictive maintenance

Let me share some cases where we enabled successful machine-learning applications:

  • Offshore asset failure detection: On this project, some wellheads were failing, while others were not. In a bid to understand the root cause of these failures, we examined manufacturing data, installation information, and combined this with other critical operating data such as how many days each well was in service, the operating pressure, temperature ranges, etc. The goal was to find a pattern through machine learning that helped build an advance warning system.
  • Pipeline corrosion: Here, our client had a large network of about 1,200 pipelines, each of which had operated for varying number of years. The client needed to assess the risk of leak in their pipelines in order to prioritize intelligent pigging for which leak information from about 200 pipelines that had failed previously. All available data was mapped together to build a machine learning model, which examined any data patterns and what caused the failure. Once the model was built, it was deployed to the other thousand pipelines. This enabled our client to prioritize the maintenance activities required.
  • Dew point breach: A gas plant was tripping due to a dew point breach. Initially, we tried to model the production fluid that was coming into the plant by using traditional process models to predict when the breach was occurring and if we could identify a cause. No matter what we did or how much we adjusted the physics-based model, we were not able to get a good match. So, we deployed machine learning algorithms which were able to understand what was causing the failure. With these models we were also able to interrogate the model to understand which part/variable in the data had the biggest impact in the prediction, and this is where we could combine our domain knowledge with machine learning models to get to the root cause and extract more value.
  • Gas-Lift Optimization: A production system involves multiple wells and pipelines that feed into the production facility. Gas lift is a technique used to life wells when they don’t have sufficient pressure to lift the weight of the fluid column. Deciding the amount of gas required at each well pad to maximize production with available gas is a challenging problem to solve in real-time. While existing simulation models do a reasonable job of providing real-time advisory system, the amount of computation required at the edge to provide the most optimized solution is large and not feasible. Combining deep reinforcement learning with simulation model allows us to build an artificial intelligence solution  that can provide the advisory / closed-loop control with minimal edge computation required.

Recent advances in machine learning and deep learning techniques allow us to build smart systems that leverage knowledge from operating data and simulation models at a rapid pace. These solutions help deliver higher productivity and smarter maintenance solutions. If you want to hear more from me on this subject and how machine teaching and process simulation can help solve industry problems, click here.

Be in the know
Get in touch
United by our common purpose to unlock solutions to the world’s most critical challenges, we are future ready, now.