Drones over power plants: This is how a Danish radar manufacturer distinguishes them from birds
Danish radar manufacturer Weibel has for decades used the so-called Doppler radar to help NASA track spacecraft, or to help the military of various countries to spot air strikes. But now the Danish tech company is expanding into a civilian market, which is interested in something much smaller than fighter jets and ballistic missiles—drones.
For power plants and airports, drones are a safety risk that they would prefer to be able to keep a close eye on. But when the objects one is looking for in radar data are small, the task of identifying them is on a whole different level, says José Navarro, engineering manager at Weibel Scientific, based in Oslo.
“The smaller the objects you want to detect, the more junk you’ll find as well,” he says.
“If you’re only looking for aircraft, then you can set a high threshold for detection and you won’t find any birds because birds aren’t interesting. But if you want to detect small drones, you need to have a very low threshold, and then you get to see everything. Not only birds but also bumble bees.”
For the same reason, Weibel has trained an ML model to answer the question: Bird or drone?
The classification takes place in several steps. The first thing that happens is that the rotating radar gets a hit and spots an object. Then it takes three seconds before the radar performs a full turn again, and sees the object again.
If the object has moved in the meantime, the system can start calculating a trajectory.
The strength of the signal—what Weibel calls the radar cross section—can give indications about the size of the object. However, since birds and drones largely overlap in size, it is not in itself a good parameter for classification.
The next step is therefore to assess whether the object has propellers, José Navarro explains.
“If you, for example, have a bird flying 10 meters in your direction, you’ll get what we call a Doppler signature. That signature will be very similar to the signature that we can measure from a drone’s body,” he says.
“But because the drone’s propeller moves differently than the body, it will look as if a part of the drone is moving at 200 meters per second in one direction and another part of the drone is doing the opposite. It also indicates that an object is a drone and not a bird, and that’s extremely valuable.”
The micro-Doppler signal, as the signature of the propellers is called, still contains a lot of noise. And a single measurement is not enough. Small birds can flutter their wings quickly, or a flock of birds can create a micro-Doppler signature that looks like a drone.
“We could do a simple approximation that determines that all objects with a micro-Doppler signature above a certain threshold are drones, and the rest are birds. That would be okay, but not very robust,” José Navarro says.
This is where the ML model comes into the picture. To collect training data, José Navarro and his colleagues have flown drones around their radar, producing around 200 datasets.
“From the drone’s GPS, we have the ground truth that we can combine with radar data during flight. That way, we can use supervised machine learning to teach a model to recognize the drone.”
In addition to the crucial micro-Doppler signature, the model is trained on variables such as the radar cross section and the object’s trajectory. But not all information goes into the model.
“We try to be very careful about what variables we use. For example, if we provide the information about how high the drone is flying, then we risk the model learning that drones never fly higher than 120 meters, as our drones do. But that may not be the case when the model is put into use,” José Navarro says.
“So we remove height from training data, but that also means that we have less information to train on. So we try to strike a meaningful balance.”
Supported by rules
The result is a random forest model that provides a better estimate of what the radar has intercepted.
The model is supported by a set of rules to help avoid mistakes. If a drone e.g. flies far enough from the radar, the micro-Doppler measurement can become so weak that it ends up being classified as a bird.
“We have tried a rule that says that as soon as an object is classified as a drone, it continues to be a drone. It prevents drones from turning into birds. But the downside is that a misclassified bird continues to be a drone. So this is something we’re working hard to solve,” José Navarro explains.
Ultimately, it will be up to the customer to decide on the balance between false positives and false negatives.
Robust regardless of the location
In general, Weibel aims to make the most robust general model that works no matter where it is used.
“If you optimize a model for a specific location, you get better performance. But we would rather spend our energy on a model with which all customers can achieve good results,” José Navarro says.
“When we first trained the model, we flew drones at the Hans Christian Andersen Airport in Odense, where we were able to fly them without a lot of complications caused by cars and so on. But if we use the same model that was trained in Odense here in Oslo, then it will think that there are a lot of drones hovering over the rooftops, because there are a lot of rooftop fans.”
Instead of training a model for Odense and a model for Oslo, Weibel is taking all the data and training a model that is used in both places. This may mean that stationary drones are not captured, or that fans are classified as drones, but it will generally mean there are fewer errors, José Navarro says.
Just as rooftop fans can imitate drones, wind turbines in radar data can be easily confused with a helicopter hovering in the air. Therefore, Weibel is generally also trying to get customers to accept a classification with several classes—which, for example, says that an object is either a helicopter or a wind turbine.
“Reducing the number of possible classes is an interesting prospect in itself,” José Navarro says.
“Instead of taking a shortcut and making an uncertain decision, we would rather pass that information on to the user, who can then, for example, check a database of wind turbines to make the final decision in the case.”