Fig. 2 shows how vibration monitoring is carried out. Data about key components within the generator, gearbox, main shaft and rotor bearings are gathered from acceleration sensors that detect vibrations. This data is collated by a monitoring unit such as the SKF IMx that sends automatic alarms to the remote data centre operated by SKF. Each turbine, depending on configuration and location, may end up having customized alarm levels that are determined during the set-up phase to ensure that the monitoring is the most efficient for that turbine.
This gives patterns of machine behaviour and enables SKF to gather reference cases in relation to different types of failure. Analysis of this data has given SKF the opportunity to move CMS to a new level of performance and efficiency through the development of new statistical techniques.
The volume of data collected from a large wind farm is staggering. As a rule of thumb there are approximately eight sensors on a geared wind turbine with roughly three measurements for each of them, 24 indicators in total; one indicator is one spectrum and one overall value. This information is collected by the condition monitoring hardware and is sent over the Internet, either wired or wireless, to a CMS server, which can be located anywhere on the earth. In one year, with on average one download a day, there are 9,000 spectra to analyse per wind turbine. For a wind park, which may hold hundreds of turbines, it is completely impossible to analyse this data in a meaningful way without using statistical modelling.
Statistical modelling compares the turbines with one another as much as that is possible, given differences in location and models. First, SKF compares what is comparable and then uses historical data amassed over the 10 years of monitoring wind farms of different models. Based on the history of this data, and dependent on turbine type, it provides the background information for any new machine that SKF starts to monitor.
A wind turbine presents unique challenges in terms of CMS that are not experienced in other industries. Wind turbines are complicated machines with a great many variables. It is not possible, for example, to apply one model of alarm level to all machines, so individual alarm models have to be developed to enable rapid comparison of those machines that can be compared. However, to attempt to do this without being guided by the statistics, filter and selection would be impossible in a reasonable amount of time.
The growing pool of historical data related to wind turbine performance is extremely useful, especially if it contains the entire life cycle of a unit starting from installation. Unfortunately, this is not always the case. Even though an increasing number of turbines are factory-fitted with monitoring technology, much of the existing fleet requires retrofitting. This traditionally occurs close to the end of the warranty period or when the operator or the service provider wants to renew a service contract.
It is, however, vital to select the right kinematic data to analyse in order to increase the statistical process accuracy. The system has some features that enable it to scan for potential theoretical defaults. Automatic scanning relies on the actual information about the type of components in the system. Each gearing has its own theoretical frequency, so without having a certain degree of certainty about the kinematics inside the turbine, assumptions made by the analyst need to be added. Given SKF’s extensive historical database, there is a good understanding and good knowledge about the components inside the gearbox and inside the generator.