CAN BIG DATA REALLY IMPROVE O&M?

Published on: 

Not only Facebook, but Google and many others work on the ability to squeeze useful information from vast quantities of data. This surge in available computing power for turbomachinery control and supervisory systems has led to discussions about the capability to gather, analyze and use large quantities of data to improve machine availability, reliability, and reduce maintenance cost and unplanned downtime.

Sensors are being added to gas turbines and other turbomachinery systems that collect vast amounts of data which is processed by high-powered computers that can enable analysis of the data in real time and drive better decisions for O&Ms and perhaps even design. Applications are being developed for jet engines, and there could be turbomachinery uses coming soon.

As a result of these breakthroughs, there is discussion concerning the integration of large amounts of computing power, monitoring and analytics into engines to monitor them more in real time, and allow back-office personnel to optimize operations, fine-tune maintenance needs and obtain far greater reliability, efficiency and profitability. Some people are talking about this not only in terms of selecting operating conditions to optimize engine life or fuel economy, but also, to be able to keep a closer eye on components, knowing when to intervene, and when to schedule an overhaul or replace parts.

This sounds great. However, related to turbomachinery, one might want to step back and ask: who will use it and what is it good for?

Current conventional trend monitoring and diagnostics rely on monitoring vibration, temperatures, and pressures vs. established baseline data. Empirical knowledge of how the machine should behave is compared with measured data, and provides indications of problems.

More state of the art systems can also cross reference observations. Over the last ten years we have also seen diagnostic tools that continually compare the performance of the machine against what is theoretically predicted from a cycle deck program. The cycle deck program performs a 1-D thermodynamic analysis that predicts the performance of the machine to the detail of component levels based on actual operating conditions.

This is compared to what is actually measured at these operating conditions, and any discrepancy indicates the potential for a problem. These cycle deck programs are fairly accurate as they are the same programs the manufacturer uses for gas turbine performance guarantees. However, since they are 1-D programs they do not require a lot of computing power. Even 20 years ago one could run a cycle deck program in real time with a gas turbine using a simple PC.

The need for large computing power comes about for either comparing massive quantities of measured data or to run a more advanced physical model of the gas turbine in parallel (such as 3-D CFD, heat and structural dynamics, finite element analysis, coupled rotordynamics, combustion chemistry). In either case, the data have to come from more, high-speed transient sensors for comparisons with the physical model or empirical trend sets.

Most gas turbine packages currently have already somewhere between 200-300 sensors. The sensors that provide better modeling often do not gracefully fail, and are a leading cause of spurious engine shutdowns. Adding more sensors may provide many more opportunities for sensor failures and associated shutdowns and maintenance requirements.

Most importantly, the accuracy of these sensors has to be high to be useful: what is the total uncertainty of the measurement, and how often do the sensors have to be calibrated to maintain a reasonable uncertainty?

Measuring uncertainty

As an example: a complex 3-D physicsbased flow and thermodynamic simulation of the gas turbine predicts baseline performance. The uncertainty of this prediction is probably within 1% for bulk performance values (power, efficiency, pressures, temperature) but for the local 3-D predictions (local combustor temperatures, blade temperatures, wall pressures) even a 3% uncertainty may not be attainable.

Advertisement

Pressure, temperature, flow sensor accuracy (or uncertainty) in the hostile environments of a gas turbine will seldom be better than 2-3%, especially after they have been installed in the gas turbine for a couple of years. Just like the sensors, the models will need to be updated and maintained; they may be less accurate as the uncertainties stack from years of equipment use.

Any performance or maintenance decision needs to account for these compounding uncertainties a factor that back-office personnel may neglect to consider.

The other area we see for increased computing power is in the attempt to use artificial intelligence (for example, artificial neural networks), for fault detection, predictions and diagnosis by training them with sets of known data or allowing unsupervised learning from free-form data. The issues here are: it takes significant time and effort to train these systems; the factors affect analysis accuracy; and the results lack interpretability.

Many of the promises of big data only apply after a significant data set has accumulated. This means that a system bought today would not help until loaded with baseline data that could take months or years to accumulate. A common way to accelerate this process is to provide a baseline from a significant fleet of similar turbomachinery, often accomplished by working with the OEM provider.

The other risk of these systems is related to actions taken based on system notification. Without experienced turbomachinery personnel to perform a validation of the data, a turbine may be taken offline for maintenance as a result of invalid data (a false positive).

Generally, systems can be trained to some degree to detect and account for sensor faults if operating conditions fall outside of the coverage of the training data, and error rates increase. Finally, the largest risk of the current generation of algorithms lies in explaining what changed and where. The black-box model often used today can only tell that a multi-parameter trend has changed, leaving it up to the user to determine if that change is spurious or valid.

However, for either system, when, after years of operation sensors fail, sensor drift will occur, and performance predictions will reduce in accuracy as they cannot properly account for degradation of the system. Then the combined inaccuracy will even be higher and less than useful for any decision making, risk assessment, or maintenance planning.

Current approaches, using a combination of physics-based models augmented by empirical data and applied by knowledgeable engineers, rather than office staff crunching massive amounts of arbitrary data, seem, for the time being, a more reasonable approach to condition monitoring. It simply does not make sense to counter the lack of accurate data by merely using a larger amount of not particularly accurate data.

Authors

Klaus Brun is the Machinery Program Director at Southwest Research Institute in San Antonio, Texas. He is also the past Chair of the Board of Directors of the ASME International Gas Turbine Institute and the IGTI Oil & Gas applications committee.

Rainer Kurz is the manager of systems analysis for Solar Turbines Incorporated in San Diego, CA. He is an ASME Fellow since 2003 and past chair of the IGTI Oil & Gas applications committee.