Blog

Sira-Kvina Hydro Power – Tracking Down Why the Generator Failed

Peter Bruce, Ramon Perez, Mark Smith

August 7, 2020

BLOG_Sira-Kvina Hydro Power – Tracking Down Why the Generator Failed-1

In early 2020, Sira-Kvina Kraftselskap, a large producer of hydroelectric power in Norway, suffered a breakdown of one of its major generators. Company technicians went through established diagnostics to identify the cause, but they were unable to pinpoint the trouble. Efforts to restart the generator continued to fail and the shutdown dragged on for months.

Sira-Kvina turned to Elder Research which, coincidentally, was already working with the company on a project to build machine learning models to better predict failures. Essentially, they said “Our standard checklists of component performance haven’t found the problem. How about taking a look at the data to see if you can find the cause?”

There was, indeed, plenty of data -- far too much to review manually. There were hundreds of sensors that produced minute-by-minute readings on the bearings, rotors, turbines, subcomponents, and more. Elder Research investigated the data leading up to the failure with a view to finding any unusual statistical changes that might indicate the cause. The raw minute-level data was very noisy, so data was grouped into power cycles (the time from power-on to power-off) and summarized by features -- e.g. mean temperature in cycle -- to capture the sensor behavior.

At first glance, nothing stood out. The noise in the data was expected, given the level of granularity at which they were working. It was not until the analysts began to look at changes in variability across time that a picture began to emerge. Variability of a sensor in a cycle was summarized by the “lumpiness” feature, defined as the variance of the variance across time windows. Inspecting by eye how lumpiness changed leading up to the failure (see Figure 1), they found an interesting pattern in the temperature sensor for the upper guide bearing of the shaft that connects the generator to the turbine: a flat period through most of the power cycles, then an ascending series of spikes, culminating in a high peak at the time of failure (the vertical dotted line towards the right). Failures tend to develop over time and so naturally the question arose as to whether the spikes capture symptoms leading up to the failure?

Figure 1a

Figure 1: “Lumpiness” (y-axis) plotted sequentially over time, by power cycle

While not directly informative, this plot was a key clue. It pointed the team in the direction of temperature as an issue, and suggested another review tactic: zoom into the cycle where the large pre-failure spike occurred, and look at the raw minute-by-minute temperature readings (see Figure 2). Here they found an apparent anomaly for the bearing sensor (cycle ID 295 below).

Figure 2-Jul-28-2020-01-44-24-65-PM

Figure 2: Bearing sensor temperature; x-axis is hours elapsed in the power cycle

Consider the plot on the lower right (Cycle 295). The starting temperature (just over 25 ºC) for this sensor on this power cycle was considerably lower than the starting temperature on the others (31-34). (Note that the plot shapes differ in other ways, but that is due to the different scales on the x-axes.)

The data science team zoomed in a step further and was able to confirm that this temperature jump of 13 degrees from startup to regular operating temperature was well outside the normal range of temperature jumps (median cycle temp minus start-up temp). Figure 3 shows the distribution of the temperature jump over all power-cycles, allowing a comparison of the typical jump to the pre-failure jump for this sensor. The cycle of interest is the outlier (dotted line) on the far right, which shows this observation to be significantly high.

Figure 3. Distribution of bearing sensor temperature jump over all power-cycles

Other sensors showed a similar pattern. Elder Research brought these statistical findings to the experts at Sira-Kvina, who found the following;

  • Vibration level was 30% higher in the case where the cold start was present. Vibrations cause stress on the components.
  • Generator heaters were suspected to have not been working properly, causing the cold start.
  • This cold start was proven to be the culprit in causing this massive, extremely expensive generator failure

Understanding the cause was very valuable to Sira-Kvina, not just in identifying this particular failure, but also in providing the methodology that will help prevent such costly and time-consuming faults like this in the future.

Summary

The standard industry logs and checklists did not provide the flexibility and granularity to quickly locate the generator fault. Only after the data was “mined” and analyzed, in a combination of automated and human expert review, was the fault pinpointed. Sira-Kvina was sufficiently impressed that they committed even more deeply to the data-driven approach to preventive maintenance that Elder Research was already developing for them.


Need Help Developing a Data Pipeline in the Cloud?

Cloud-2Our team of certified cloud developers build data pipelines and machine learning solutions that scale to meet changing business requirements.

  • Certified AWS Solutions Architects
  • Certified Google Cloud Professional Data Engineers
  • We can develop within a client’s cloud environment or host a client instance in our own cloud environment

learn-more-button-300x98


Related

Predictive Maintenance Optimizes Gas Well Production

Playing with Fire: The Sensor Analytics Explosion

Do We Have the Right Data?

What is Data Wrangling and Why Does it Take So Long


About the Author

Ramon-Perez

Ramon Perez focuses on financial modeling and market risk detection. A decorated war veteran, Major Perez served as an Air Force Intelligence Officer specializing in Signals Intelligence. As a mission director at the National Security Agency, Ramon led teams analyzing enormous data sets to produce actionable information in support of worldwide contingency operations--including deployments to Iraq, Afghanistan, and South America in support of Special Operations. Upon leaving the military, Ramon returned to graduate school with an interest in understanding economic models. Prior to joining Elder Research, he analyzed long-term defense budget projections for the Congressional Budget Office. Ramon holds a Bachelors degree in Systems Engineering from Georgia Tech as well as Masters degrees from Harvard and Georgetown in Finance and Economics, respectively. He and his wife are avid travelers, having lived in Asia and South America, and having visited over 30 countries.


About the Author

smith, markMark earned a first-class BSc in Mathematics from King's College London where he was awarded the Drew second prize. During his time at Kings he completed a research project in online sparse Gaussian process regression.

He further earned an MSc with distinction in Statistical Sciences from the University of Oxford. His dissertation focused on the application of Machine learning to detecting leading & lagging relationships in financial time series data. Mark has gained experience building and applying Machine learning & Statistical models to various problems. He is particularly interested in Bayesian Statistics and Deep learning.


About the Author

Peter Bruce, Ramon Perez, Mark Smith Peter Bruce is Founder and President of The Institute for Statistics Education at Statistics.com. Previously he taught statistics at the University of Maryland, and served in the U.S. Foreign Service. He is a co-author of Data Mining for Business Analytics, with Galit Shmueli and Nitin R. Patel (Wiley, 3rd ed. 2016; also JMP version 2017, R version 2018, Python version 2019; plus translations into Korean and Chinese), Introductory Statistics and Analytics (Wiley, 2015), and Practical Statistics for Data Scientists, with Andrew Bruce, (O'Reilly 2016). His blogs on statistics are featured regularly in Scientific American online.