Detecting Spikes in Wind Speed from an Acoustic Anemometer


An acoustic anemometer like the R M Young Model 85000 Ultrasonic Anemometer sometimes will give spikes, especially during rain. An example is shown below from Port Nelson during a rainy day, in which the wind speed recorded by a traditional anemometer at Nelson Port is compared with the wind speed recorded by the R M Young 85000 ultrasonic anemometer at the Fairway Beacon offshore (see locations here). We assume these spikes occur as a result of spurious acoustic reflections off raindrops.
The problem is: they cause the plotting program to go offscale and upset the pilots, so what can we do about them?

Figure 1. A. Rainfall at Nelson Airport and B. Comparison of wind speed at Nelson Port and Fairway Beacon.

Tukey 53H

John Tukey (co-inventor of the fast Fourier transform, or FFT) developed an algorithm for detecting what he called "wild" points.
Here it is:

For this implementation, I have used the median of the absolute deviations or MAD instead of the standard deviation because it is more robust in a spiky environment.
There are a couple of parameters that must be determined:


The figure below shows a sequence of several days when there were spikes in the acoustic anemometer record and how the Tukey 53H algorithm performed in detecting them. In this case, k = 1.5 and n = 1440, corresponding to a day of data at 1 minute intervals. The Tukey 53H algorithm was applied twice. On the first pass, 45 spikes were detected, and on the second pass 200 more spikes were detected. Many of the latter were points adjacent to the spikes that were detected in the first pass. For a record containing only isolated spikes, one pass would be sufficient, but in this case many of the spikes overlapped. The resulting record in the lower plot has had almost all the spikes removed, though there are a few points remaining in the first day of record that could be spikes. Notice that no spikes were detected on the second day. Thus, the algorithm has successfully ignored good data and refrained from removing data that should have been retained.

Figure 2. Illustration of the application of the Tukey 53H algorithm.


A practical solution to the problem of spikes in an acoustic anemometer record is to detect the spikes using the Tukey 53H algorithm and eliminate them. Ideally, of course, there would be no spikes because they would be eliminated at an earlier stage in the measurement process. Ultimately, this is the best solution, but in the meantime the methodology described here can be used to avoid unsightly spikes in the display used by the pilots at Nelson.

Derek Goring
Mulgor Consulting Ltd
Ph: +64 3 942 5452