Exploring the case for digitalization and planning predictive modeling of your equipment.

In recent years, predictive analytics for major rotating equipment maintenance has been marketed by major original equipment manufacturers (OEMs) and digitalization pioneers as the future of maintenance planning. This is especially key in the Upstream industry, where solution providers are claiming that invested companies are saving billions of dollars.

Is Predictive Analytics Really a Large and Profitable Opportunity for Upstream Operators?

The answer is yes. Predictive analytics will generate advantages and possibly be profitable, but the “billion-dollar” business case is largely exaggerated. Some of the business cases propose savings that are higher than actual maintenance costs. You cannot turn your maintenance cost into a revenue-generating post for your financials.

Predictive maintenance models assume that you can train a digital model of your equipment to forecast possible breakdowns, well before they happen. The equipment failure calculation is based on variations in sensor data from the equipment and its surroundings. For instance, by monitoring vibrations from a compressor, you should be able to identify issues on bearings through a change in vibration patterns. Combining multiple sensor data allows you to build a model that accurately predicts failure and breakdowns before the breakdowns occur. This should allow you to remove your schedule-based maintenance programs and avoid any potentially excessive cost spending on unnecessary maintenance items. With this strategy, only necessary maintenance is performed to keep the integrity and reliability in place.

This is how it is in theory. And in all fairness, it has proven quite successful in the aviation industry for a long time where planes have been operated on with combined conditioning monitoring and predictive maintenance models for major rotating equipment (turbines).

Are the Return on Investment (ROI) Models Accurate Even If They are Exaggerated?

As a promotional tactic, you can build your case based on the worst-case scenario. For example, this is where an operator has the worst reliability of the production equipment, the absolute most inefficient time to repair, and the absolute maximum of lost oil in connection with the breakdown.

Add to that, the repetition of this incident on multiple pieces of equipment will cause the costs of not implementing predictive maintenance analytics to explode. The fact is, most oil and gas operators have a relatively reliable operation, with production efficiency around 92 to 94 percent when including planned maintenance. If we exclude planned maintenance accounting for approximately 3 percent of losses, then we are talking about 95 to 97 percent efficiency. Of this maximum, 40 percent is directly connected to major rotating equipment, and in most cases, it is the surrounding equipment, not the actual generator or turbine that caused the loss.

So yes, the ROI models are extremely exaggerated. When running real ROI comparisons, the costs of establishing the digital twin, building the model, achieving a reliable history and predictive model, and license fees to the OEM, could often eat up the profitability of a project.

How to Build a Valid Predictive Model in The World of Artificial Intelligence (AI) Learning

To build a valid predictive model, you need first to build a digital twin. This is the digital copy of your actual equipment, which is constantly replicating the operative state, based on as much data as possible. The digital twin is then used to build the predictive model, where historical breakdowns and failure data are used for building the predictive scenarios.

There is much talk about the ability of artificial intelligence to make accurate predictions, but this is a stretch. Machines are not intelligent beings that can learn by themselves, but electronic mechanisms that must be taught. What they can do, though, is analyze massive amounts of data, and find patterns and connections. This is where we can train the machine to analyze data and scenarios and come to conclusions.

Related Articles & Content

Upstream Benchmarking — Your Path to Performance Excellence

Read Article

Upstream Turnaround Strategy: How Do You Compare?

Read Article

Production Reliability or Operating Costs: Where Should You Place Your Focus?

Read Article

Solomon's FPSO Global Database

View Infographic


We build machine learning where all learning is based on data, cases, and inputs that we supply. Through machine learning today, we can analyze and combine massive amounts of IIoT1 data from our equipment and find correlation and patterns, though this is not artificial intelligence. The model can only predict what we teach it.

This also means that to build a functional predictive model, you need lots of historical data including breakdowns and failure. This is also true for other industries. For example, the aviation industry benefits from having collected and shared this data for years and has succeeded in building reliable predictive models for turbines. For example, a turbine used on an Air France Boeing 747 is showing operational patterns similar to a turbine used on a British Airways Boeing 747.

Successful Predictive Maintenance in Aviation vs the Upstream Industry

It is also important to note that although predictive maintenance modeling is successful in aviation, it is because it is highly supported by intense condition monitoring. Any deviation is monitored, and it is not the model’s responsibility to issue warnings.

Key conditions for Upstream:

          1. Upstream oil and gas installations differ highly in design, capacities, structure, and process. This means that you will not have a massive global dataset that can be used across all equipment. You will only have limited benefits from repetitive generation of digital twins and predictive analytics.
          2. Oil and gas companies are quite protective with IIoT process data. In contrast, the aviation industry has shared operational data across companies, which is then used by OEMs to generate predictive models. Upstream would need to adjust its willingness to share data within the industry to take full advantage of predictive modeling.
          3. Oil and gas companies have traditionally followed rigid schedule-based maintenance programs. Any breakdown is expensive, due to oil lost in connection with the breakdown and the investment in scheduled maintenance is therefore commonly accepted. This means that operators have relatively few historical data points on breakdowns and therefore, a reliable predictive model cannot be established. Any attempt to start a predictive regime, such as removing scheduled maintenance, will result in unforeseen breakdowns at some point, which is a costly occurrence for the operator.
          4. An oil and gas operation has a limited lifespan, typically 15 to 20 years, which is why payback on an investment in predictive maintenance must be relatively fast. Given the differences in assets and the limited ability to reuse developments, a project might not prove feasible.
          5. Overall, a project might prove unfeasible if the cost of establishing the model, license fees of application, and the cost of the learning period result in an ROI that is too low to make the project feasible.


If you are looking to enter the great frontier of digitalization and planning predictive modeling of your equipment, it is wise to carefully take these considerations into your business case.

 

[1] IIoT - Industrial Internet of Things