HomeNewsWind turbine monitoring: Spotting the difference

Wind turbine monitoring: Spotting the difference

Similarity-based modelling software can detect departures from normal wind turbine operation that would otherwise be missed, allowing operators to anticipate problems and take preventive action before the performance and reliability of their turbines are compromised.

Donald S Doan, SmartSignal Corporation, USA

Wind is unpredictable in predictable ways. The chaotic behaviour of wind makes monitoring and analyzing modern wind turbines a challenge à‚— not only for the owner operators, but also for the original equipment manufacturers (OEMs).

The variability of the data from wind turbines makes monitoring for abnormal behaviour difficult for technicians and engineers. New predictive-analytic technology for monitoring and analyzing the performance and mechanical condition of wind generation systems can detect degradation in advance of the OEM monitoring systems and help mitigate potential failures. These new predictive-analytic systems focus attention on abnormal equipment conditions, resulting in improved system performance, reliability and availability.

SmartSignal’s predictive technology uses OSI data infrastructure and network/internet systems. It applies patented similarity-based modelling (SBM) software algorithms to real-time data to identify equipment abnormalities. This technology provides alerts by exception, as opposed to manual review of all the component sensor data trends, to highlight abnormal conditions. This exception-based alerting approach allows the modern wind farm analyst to reduce review time by focusing only on deviations from normal system operations.

Improved performance

The amount of data being retrieved, recorded, trended and viewed in today’s wind farms is growing exponentially. Given the variability of the data and the quantity of data being stored, the job of the analyst to make sense of the raw data is next to impossible.

The need to detect abnormal behaviour before it affects the operation of a turbine is fundamental to improving availability and reliability. Typically, the alarm system set by OEMs is designed to protect people and equipment from ancillary damage. For example, the high-temperature alarm on a gearbox or generator bearing is designed to reduce the damage to the asset. However, the damage to the component (in this case, the bearing) may have already occurred.

Until now, the monitoring of the nacelle and support equipment is done by embedded OEM systems, and by deploying technicians to the turbine and taking local data to support condition-based maintenance practices. The idea behind condition-based maintenance is to monitor all the turbines’ critical equipment to detect the onset of a failure, so that engineering, maintenance and operations can determine the condition of the equipment and the potential impact in terms of safety, reliability and availability. The data collected is used to assist maintenance and operations in the planning and scheduling of work proactively, as opposed to reactively.

When a system is considered to be running in an optimal condition, the easiest way to monitor performance is to compare current behaviour with the historical operating condition and note deviations from the normal condition. With the large number of sensors now monitoring a turbine, operators and technicians detect early drift from normal behaviour in a number of ways.

One method is to trend the data using the data historian and to ‘look’ for changes. Other methods include exporting the data to a spreadsheet and writing custom calculations to determine the change from normal, and comparing the sensor data to first-principle curves developed for the component (turbine curves, for example).

All are applicable methodologies, but have one common weakness: the time required to look at, download, export and manipulate the data. Only after the data has been manipulated or someone has observed a trend is the data useful in determining the condition of the asset. Determining how the performance compares against known operation typically requires significant knowledge.

In a wind farm, this can be especially difficult, due to the variability in wind speed, direction, shear, turbulence and so on. The easy detection of a developing fault is complicated by the motive force behind the electrical generation à‚— wind.

Similarity-based modelling

Similarity-based modelling (SBM) is not new in the power industry. It was initially developed for nuclear power plant instrument monitoring. These systems use modelling techniques and algorithms to incorporate data from sensor suites around an asset and determine changes in behaviour from historical operation data for that asset.

This predictive technology is very accurate in modelling the behaviour of an asset with minimal sensor input. If a sensor deviates from historical operating conditions (ambient conditions, input energy and load), the system will notify the end user. This assists the monitoring of a mass of sensor data. It also accounts for deviations due to the random behaviour of wind and highlights abnormal behaviour that may be indicative of equipment problems.

Figure 1 shows SmartSignal SBM of turbine gear oil and bearing temperature from 10 to 800 KW. The blue graph overlaying the red is the actual measured value, and the red in the same chart is the modelled estimate. The all-red graph is the residual (the difference between actual and modelled estimate). The residual graph (red graph) demonstrates how SBM models the changes due to wind variability within 5 à‚°F of actual from 10 to 800 KW.

Figure 1. SmartSignal SBM of turbine gear oil and bearing temperature from 10 to 800 KW
Click here to enlarge image

Any temperature that persistently drifts 10 à‚°F above normal would post an incident to the monitoring system, SmartSignal’s WatchList.

This temperature drift is well below the OEM-recommended alert level and is identified as an abnormal condition. The alert gives monitoring personnel time to diagnose the problem and plan proactively. More lead time and improved fidelity lead to increased reliability and availability.

Within the WatchList, the user is able to ‘drill down’ to the asset and compare the sensor to the modelled suite of sensors. This focuses the user on abnormal deviations in an asset as opposed to the proverbial needle-in-the-haystack trending tasks that they usually face.

Figure 2. Sample WatchList
Click here to enlarge image

Although wind technicians and operators are highly effective at trending and monitoring their systems, the ability to focus on abnormal behaviour reduces the burden on them by focusing on abnormal changes as opposed to picking abnormal behaviour out of the expected values. The biggest challenge in wind turbine operation is detecting anomalous behaviour early enough to mitigate a failure.

Wind farm monitoring

The wind generation industry is rapidly expanding into remote and diverse areas: oceans, deserts, mountains and plains. The new wind farms are deploying more modern turbines that are monitored using more comprehensive sensor suites than are available to owner operators.

With the growth in wind farm assets and the deployment of newer turbines has come a growth in the use of digital data. This increase means more data needs to be stored, reviewed and analyzed. In process-control computers, the software uses digital data to control the process parameters. This data is what personnel trend. SBM cuts through the sensor noise to help find the ‘gem on the ocean floor’.

Predictive analytics software analyzes in real time all the data that is collected by a programmable logic controller in the nacelle. The data collected in predefined time intervals is sent to a hosting system that applies the predictive analytics model to the sensors around an asset and returns information on the behaviour of the sensors in the model.

If a parameter deviates from the normal historical process conditions, the parameter alerts the WatchList analyst of the deviation and focuses analysis on the off-normal behaviour. The deviations can be as simple as a sensor starting to deteriorate or efficiency losses in a turbine. This information facilitates proactive maintenance of the asset and maintenance forecasting.

These tools are critical to analysts’ daily activities. With the increase in data density around a turbine, the importance of determining what is normal is paramount. If an analyst is to perform their job effectively, software tools are imperative in mining the data for abnormal trends. Armed with this abnormality detection, the analyst can focus on determining the cause and formulating a response.

These systems do not take the human element out of the analysis. They help the analyst look at large amounts of data and make sense of them. If the data is reviewed without a filter of some sort, it appears chaotic.

In Figure 3, there are only 14 sensors, yet it challenges the human mind to make sense of the data. Traditional alert limits will assist personnel in protecting equipment and personnel, but the availability and reliability of equipment are still at risk with these alerting systems.

Figure 3. Sensor data chaos
Click here to enlarge image

With predictive analytics systems, analysts do not need to be involved in direct data trending. The data is monitored by the software, and analysts only review the data when an exception is posted, giving analysts ample time to respond to the changing condition.

By highlighting only signals deviating from a pattern, monitoring is more efficient. The analyst no longer has to search all signals looking for visible problems.

Efficiency in this context means that either the scope of the analyst’s monitoring increases or the time required is reduced. By performing this analysis routinely and continuously (every five to ten minutes for a typical wind application), the analysis is near real time.

By using algorithms to identify pattern changes earlier than the human eye can typically detect them, the analysis is highly accurate. By using specific historical data from the equipment or system to ‘train’ the models (reducing the need for prior knowledge), the analysis is ‘smart’ to a degree.

SBM models a group of related signals by analyzing historical data from those signals to characterize normal behaviour. The patterns identified are then put online and used to analyze each sample of data collected. From this pattern, an estimate of each signal’s behaviour is generated and compared with the actual value.

The difference between the actual and estimate, the residual, is continuously compared with empirical thresholds. The results of the comparison cause rules to fire that then draw the attention of the analyst. SBM results in accuracies that are typically more than adequate for any application.

One significant advantage of SBM is that signals can be modelled together when they are physically linked in behaviour. Modelling a generator is one example of this. The generator amps, voltage, wattage, speed, bearing temperatures, bearing vibrations, wind speed and so on are all modelled together. No regression or other parametric analyses are required. The parameters all move relatively together and identifiable patterns of behaviour will be present. The analyst can bring together monitoring disciplines that are often conducted separately.

One step ahead

In the wind farm context, the unique SmartSignal technology combines with analyst staff and tracking systems to help wind farm owners and operators increase the reliability and availability of individual turbines.

With, say, 300 1.5 MW wind turbines operating in a wind farm, SmartSignal models each asset against its own individual operating behaviour. Taking into account ambient temperature, humidity, air flow fluctuations and so on, the turbine is modelled to detect any sensor deviation from normal operating behaviour.

For example, when a generator bearing high-temperature or vibration value deviates from historical behaviour by as little as 10 à‚°F or 0.05 ips, the technology confirms that it is a sustained deviation and posts an incident on the WatchList. The analyst monitoring the WatchList is alerted to the deviation.

The analyst can then contact the wind farm owner or operator and inform them that there has been a deviation on the generator bearing. The owner or operator can then investigate the problem next time a technician is at the farm. That technician may find that the oil filter for the bearing is clogged and be able to mitigate a bearing failure.

Abnormality detection maximizes the effectiveness of wind farm personnel and increases the reliability and availability of wind assets. SBM software focuses the analyst on departures from normal operating behaviour that are nearly impossible to detect otherwise, especially when the analyst is monitoring hundreds to thousands of sensors.

It is easy to rely on actual-value hard alerts that are typically set by OEMs and refined by operations personnel to protect ancillary turbine equipment.

However, these alerts are generally too late to protect equipment from damage, whereas SBM systems allow the analyst to take preventive action, averting damage to the turbine and associated equipment. With steel and equipment prices currently at historically high levels, risking damage to wind turbines could prove a costly error.