Jul-2017

Optimising the health of rotating equipment

Applying predictive analytics to rotating equipment enables a preventative maintenance programme for improved savings

MARTIN TURK
Schneider Electric

Article Summary

In the midst of tough economic times, oil and gas businesses are facing an urgent need to operate at the highest levels of reliability and efficiency while also increasing productivity and controlling costs. They want to limit downtime and minimise risks to safety and the environment. While there are a number of challenges facing the industry, including digital disruptions and financial uncertainty, applications enabled by the Industrial Internet of Things (IIoT) can provide significant performance and reliability improvements.

From an asset performance management (APM) perspective, organisations are leveraging industrial data and advanced analytics to keep equipment running safely and reliably for as long as possible. This is made possible through data collection and analysis for predictive maintenance execution, consequently empowering personnel to act before equipment failure occurs.

Smarter maintenance
Petroleum refineries require a diverse set of complex assets to produce fuels and other by-products from crude oil. Reducing the incidents of slowdowns or unplanned shutdowns due to equipment failures is potentially worth millions of dollars a year in terms of avoiding production losses, damage to equipment and unscheduled maintenance. Real-time health and performance insights can be used to influence decisions and actions that drive efficiencies and improve competitive advantage. This asset health data is already being created and can be used in maintenance and asset management programmes to mitigate risks and ensure that critical equipment is operating as expected.

A smart, comprehensive maintenance programme includes several approaches that are appropriate for various types of equipment with the goal of obtaining the greatest return on each asset. Figure 1 provides a graphical view of the maintenance hierarchy.

The most basic approach, reactive maintenance, involves letting an asset run until it fails. It is only suitable for non-critical assets that have little to no immediate impact on safety and have minimal repair or replacement costs so that they do not warrant an investment in advanced technology.

On the other hand, preventative maintenance (PM) approaches are implemented in hopes that an asset will not reach the point of failure. The PM strategy prescribes maintenance work to be conducted on a fixed time schedule or based on operational statistics and manufacturer/industry recommendations of good practice. PM can be managed in the enterprise asset management (EAM) or computerised maintenance management system (CMMS).

A more proactive approach, condition based maintenance (CBM), focuses on the physical condition of equipment and how it is operating. CBM is ideal when measurable parameters are good indicators of impending problems. The condition is typically defined using rule based logic, where the rule does not change depending on loading, ambient or operational conditions; the rules drive automated work order generation based on a specific situation.

For more complex and critical assets, a predictive strategy is appropriate. Predictive maintenance (PdM) relies on the continuous monitoring of asset performance through sensor data and prediction engines to provide advanced warning of equipment problems and failures. PdM typically uses advanced pattern recognition (APR) and machine learning, and requires a predictive analytics solution for obtaining real-time insights of equipment health.

Risk based maintenance (RBM) is a comprehensive prognostic strategy that allows plant operations and maintenance personnel to make decisions using PdM, CBM and PM outcomes. As a result, the planning for the maintenance and the operation of equipment helps to ensure reliable and safe plant performance.

Predictive analytics
Predictive analytics uses a mathematical engine to transform raw data into actionable information. There are various approaches available as discussed below. Some are easier to use than others and have applications for which they are best suited:
• Neural networks: parametric technology that has been around for years. Used when the modes and networks are well known and there are not many deviations from them. The challenge is the lengthy training process required to model the dynamics of the process.
• Clustering: non-parametric techniques, meaning it takes an unlabelled data set and forms subsets of groupings/clusters of data and reorganises the data into smaller clusters, for instance turbine cycles from start-up to shutdown.
• Decision tree learning maps observations to outcomes. This is simple to understand because there is typically a fixed set of outcomes for this prescriptive approach, meaning you limit the analysis to the possible outcomes, which may or may not all be known.
• Fuzzy logic: the opposite of Boolean, which is based on 1 or 0. Fuzzy is somewhere between that which says something is more true than not. This approach is okay if you do not need the engineering accuracy or detection of subtleties of behaviour.
• Deep learning: used for ‘unsupervised’ data mostly where you generate more complex relationships from simpler ones. You need a lot of domain specific data and a data scientist to understand how to configure the system, but it is very powerful for extracting information on systems, technologies and trends. 
• Machine learning: a way to automate the model building and learning process in both supervised and unsupervised modes. The idea is for the programme to learn more as time goes on and become more intelligent. For very dynamic processes this can be a challenging approach but it is being applied by the industry in the space during normal operating cycles, not during start-ups/shutdowns.
• Pattern recognition: a branch of machine learning that focuses on the recognition of patterns and regularities in data, although it is in some cases considered to be nearly synonymous with machine learning.

So predictive analytics can be as simple as using rate of change to predict a value at some time in the future or much more complex, for instance use of pattern recognition and machine learning.

Using a predictive asset analytics solution in support of a PdM strategy can lead to the identification of issues that may not have been found otherwise. According to research by ARC Advisory Group, only 18% of asset failures increased with use or age. This means that PM alone is not enough to avoid the other 82% of asset failures, and a more advanced approach is required. Predictive analytics software uses historical operational signatures for each asset and compares it to real-time operating data to detect subtle changes in equipment behaviour. The software is able to identify changes in system behaviour well before traditional operational alarms, creating more time for analysis and corrective action.

Figure 2 shows various approaches to deployment of a predictive analytics solution for continuous monitoring of equipment health.