Sensor(s) (singular or an array of microphones) detect acoustic signals (sound waves), parse the signals in real time, and send them to the server. The information the server receives involves the sensor’s identification, its geographic coordinates, and a time stamp of when the sensor detected the sound waves. The server can instantaneously process the messages received from the sensor. Server processing involves transmitting and deciphering each sensor’s messages, calculating coordinates of drones, and informing allies or anti-drone weapon systems. It is important to note that the interactions between the server and sensor assume preexisting cell communication infrastructure [1].
Differentiating Background Noise From Drone Noise
The main challenge that arises for acoustic drone detection systems is from the sounds in the surrounding environment. How does the system differentiate between a noise from a drone and another noise in the environment, such as a bird or a plane? Most significantly, how can this system differentiate an enemy drone from a friendly one? Acoustic signal processing and machine learning work together in acoustic drone detection systems to account for these circumstances.
Acoustic Signal Processing
Acoustic signal processing is a method of studying the sound detected by the sensor to verify the presence of a drone and estimate its altitude and geographic location. Each drone has a specific acoustic signature produced by its motor and blades. Advanced programs, machine learning, and data-labelling methods are often used to differentiate between the drone’s specific sound waves and other signatures being tracked as a byproduct of microphone sensitivity [1, 2].
Relevant Machine-Learning Algorithms
The three most common types of machine-learning models used to train data on drones are (1) convolutional neural networks (CNNs), (2) recurrent neural networks (RNNs), and (3) long short‑term memory networks (LSTMs). They are used independently or in various combinations with one other to increase reliability and accuracy.
-
CNNs: Utilize drone images and recordings to train themselves using a variety of characteristics such as color, shape, size, texture, and motion. A very large number of images must be acquired and verified for size and format consistency and to remove any noise or unwanted features. Eventually, the images collected are formatted as input and binary labels to produce an output as drone or not drone [3].
-
RNNs: Use a drone’s past trajectories to predict future flight-path patterns and movements. CNNs can be combined with RNNs to produce a system for drone detection that provides both visual and movement accuracy [3].
-
LSTMs: Similar to RNNs but can be trained to detect drone patterns using acoustic data and radar [3].