Applying Deep Learning to Aircraft Data to Improve Airspace Management


2/26/202310 min read

In addition to our core aims of building a market leading UAS UTM, UAS charging & communication infrastructure and global drone detection network, DeFli are quietly becoming leaders in the field of applying deep learning to the data we acquire from our host network to provide various stakeholders with applications that will both enhance and change their operations and policies. What follows is a list of ways in which we apply deep learning to our acquired data and the output products this offers to our ever expanding client base.

Rf Signal Classification

We take the ACARS and ADS-B data provided by our hosts to train and verify the output of a single basic deep neural network in identifying and classifying the received radio signals. Using deep learning to classify received radio signals has significant improvements over the currently used methodology known as “feature classification”. Using a convolutional neural network (CNN) to classify received radio signals enables a mass reduction in the receiving hardware required for operators who parse signals from multiple channels whilst reducing the false classification rate dramatically. For application to the real-world in respect of airspace management, the above described model allows for easier to deploy, more reliable and edge based classification systems which will be fundamental to safer skies as UAS grows in deployment numbers. The incorporation of a dedicated band (C-Band) for UAS communications further outlines the need for a robust classification system.

Detecting ADS-B Spoofing Attacks

The Automatic Dependent Surveillance-Broadcast (ADS-B) system is a key component of the Next Generation Air Transportation System (NextGen) that manages the increasingly congested airspace. It provides accurate aircraft localization and efficient air traffic management and also improves the safety of billions of current and future passengers. While the benefits of ADS-B are well known, the lack of basic security measures like encryption and authentication introduces various exploitable security vulnerabilities. One practical threat is the ADS-B spoofing attack that targets the ADS-B ground station, in which the ground-based or aircraft-based attacker manipulates the International Civil Aviation Organization (ICAO) address (a unique identifier for each aircraft) in the ADS-B messages to fake the appearance of non-existent aircraft or masquerade as a trusted aircraft. As a result, this attack can confuse the pilots or the air traffic control personnel and cause dangerous manoeuvrers. DeFli currently train and deploy a two-stage Deep Neural Network (DNN)-based spoofing detector for ADS-B that consists of a message classifier and an aircraft classifier. It allows a ground station to examine each incoming message based on the PHY-layer features (e.g., IQ samples and phases) and flag suspicious messages.

We use a message classifier and an aircraft classifier systems. The message
classifier decides whether the message is malicious or not. If a message is considered non-malicious, it means that it is not transmitted by the SDR-based spoofer, but it may come from a malicious aircraft instead of the legitimate transmitter as indicated by the ICAO address in the message. Hence, the aircraft classifier aims to further determine the transmitting aircraft of the message and compare the output ICAO address against the
claimed ICAO address to detect the aircraft spoofing attack. Since both message and aircraft classifiers are based on DNN, they need to be trained on a large labeled dataset the likes of which DeFli collate through or host networks. At runtime, the system takes each incoming ADS-B message (IQ
data) as input and decides whether this message is suspicious or not.

Message classification is considered as a binary classification problem. In our model, the raw IQ samples are used as features. Since each message lasts 120 μs, a sampling rate of R MHz will produce 240R interleaving IQ samples.

Aircraft classification predicts the source ICAO address of the received ADS-B message and compares it against the claimed ICAO address. Since there are many aircraft that transmit ADS-B messages, it then becomes a multi-class classification problem. It is worth noting that through aircraft
classification, the system can not only detect the aircraft spoofing attack, but also identify the masquerading aircraft. Unlike message classification, the aircraft classifier does not use IQ samples or their magnitudes that encode the (possibly spoofed) ICAO address as features in order to avoid
being deceived. Instead, it uses the phases that are independent of the claimed ICAO address as features.

Predicting Go-Around Manoeuvres

The predictability of go-arounds is quite an interesting problem to solve; its solution can help mitigate the effects produced after it occurs. From a machine learning point-of-view, the main inconvenience of this use case is the data imbalance problem. Since this event happens in about 0.3% of operations, if we try a binary classification problem predicting 0/1 if a go-around occurs, the model will need equality between the classes to learn more of go-arounds. Another challenge will be the labelling of the data, since we won’t know at glance if a certain flight experienced a go-around just by looking at raw ADS-B columns.

Most go-around maneuvers share some similarities in aircraft behavior:

  • Full or maximum allowable takeoff power must be applied until flying speed and controllability are restored.

  • The nose should be kept from pitching up too soon, with an altitude that permits a buildup of airspeed well beyond the stall point.

  • The flaps and gear will be retracted.

  • The rate of descent must be positive.

From these patterns, even without all the information, it’s feasible to look over flight trajectories in ADS-B that meet these requirements (speed, altitude, heading), and label them as potential go-arounds. Afterwards, we could set a prediction point a few minutes before the touchdown point and see if we can predict (0/1) if a go-around is about to happen. Of course, weather features, like METAR or wind profiles, measured at runway surface will highly improve the model accuracy.

Loss of Separation Hotspots

We understand a Loss of Separation (LoS) to be when two airborne aircrafts breach the separation minima established by the controlled airspace. When surveillance systems are used, for instance ADS-B, the minimum horizontal separation between two adjacent planes should be at least 5 nm, and vertical separation for IFR flights as 1000ft below FL290, and 2000ft above FL290. These rules might change a bit depending on the airspace. Similar to the previous use case, it’s possible to label pairs of trajectories to detect losses of separation between aircrafts. We use machine learning to identify geographical regions more prone to induce or become associated with LoS, probably by using unsupervised learning techniques to infer patterns between observations and set clusters. These results are plotted in our UTM through hotspots indicating when and where LoS happen. In this case, machine learning aids to better understand the precursors of this kind of safety event to prevent similar situations in the future.

Sector Occupancy

The ADS-B radar is composed of multiple time series from different flights flying across different sectors. Sectors are nothing more than a visual segmentation of airspace, distributing the control workload into different ANSPs. The trade-off between sector capacity and demand is always present in the ATC world. The opening/closing of sectors should precisely monitor the demand in order to achieve efficient use of all available resources. It’s important to know the expected sector demand in advance so that airlines can dispatch their flights properly. We use machine learning to train a regressor able to forecast the demand for a specific sector at a given time.

Air Routes Profiling

In aviation, there are multiple ways to travel a route in terms of origin to destination (OD) airports. These routes are normally pre-established, and depending on external factors, aircraft might change the followed path. We use ADS-B data to analyze all the extant ways of reaching the destination for a given OD-pair, profiling routes for any airport combination. This improves the waypoints selection and provide insights into how route selection works and the best way to optimize this process.

Machine learning helps select the optimal trajectory during pre-tactical phase, considering the context of the flight (en-route weather events, airspace regulations), and also provide better insight into the precursors of each flight path, detecting common procedures repeated for certain OD-pairs, such as fly-by-waypoints or directs.

Identifying Engine Types

We use multivariate long short-term memory–fully convolutional network (MLSTM–FCN) to classify time series data and predict engine type using ADS-B kinematic data. Using the MALSTM–FCN algorithm as a basis for the model, the DSDEC algorithm employs a unique 2-stage approach. When using the model to make predictions, the first stage predicts if the aircraft is a jet or not a jet. Then, the ‘not jet’ predictions are fed to the second stage. The second stage predicts if the aircraft has a piston or turboprop engine. The results from both the first and second stage are combined to provide an engine prediction for each observation. For DeFli we take this a step further and apply the mechanisms to the specific structures of UAS’s and their blade types. The specific components of ADS-B message data that we use are:

  1. Altitude and Ground Altitude — Since jet, turboprop, and piston engine aircraft tend to fly best at different altitudes, these features are important in distinguishing between them.

2. Airspeed — Jets fly faster than piston or turboprop engines. Turboprop engines can reach greater speeds easier at higher altitudes than piston engines.

3. Barometric Pressure — This feature is another way to measuring altitude. For every thousand feet of elevation, the pressure drops by 1 inHg.

4. Vertical Speed — The importance of this feature is similar in nature to air/ground speed.

5. Time — This feature would allow the comparison of different chronological points of the flight

6. Track — With location and speed, track can be used to learn specific aircraft patterns.

7. Lat/Long — Aircraft may exhibit different behaviors depending on the geography. For example during the first ten minutes, an aircraft would takeoff differently from a mountainous region than an open field.

8. Location (X,Y,Z) — Lat/Long are normalized to better represent a 3-dimensional space. This feature was not in the raw data, but was instead generated to help better represent the data.

The extension of our work to incorporate UAS enables us to include specific features such as the propellor blade size and type a well as potential payload sizes (weight) this can help to provide an overview classification of the types of UAS operating in a given sector.

Handling Overlapping

As the amount of air traffic increases, especially in lower altitudes ADS-
B signal is easily overlapped due to the large number of aircraft and aircraft location information cannot be efficiently obtained. It is therefore vital to separate the ADS-B signals to ensure ongoing safety in lower, densely populated altitudes.

Traditional features-based ADS-B signal separation can be divided into three categories. The first category is to utilize power difference between overlapping signals. However, this category can only separate two overlapping signals and are easily affected by the power difference. The second category is to use carrier frequency difference information to separate overlapping signals, but this category needs large carrier frequency difference. The third category uses other features of ADS-B signal to separate signals utilizing the sparse characteristics of ADS-B signal to separate signals based on compressed sensing methods.

Our ADS-B signal separation network consists of three modules including the encoder, separation network and decoder. Firstly, ADS-B overlapping signals are fed in encoder to generate an adaptive 2-D representation. Then segmentation operation is used to segment 2- D representation and generate 3-D representation. The separation network is utilized to generate effective masks from 3-D representation. In order to get the same size of 2-D representation, the overlap-Add is used to transform 3-D features into 2-D representation in the separation network. Finally, the decoder is to separate the overlapped signals from output of encoder and separation masks.

The separation network is where our deep learning occurs, it consists of six Dual-path independently convolution Gated Recurrent Unit (Ind-CGRU) modules, one overlap-Add module and one 1D Conv. Compared with traditional RNN, Dual-path Ind-CGRU can explore more spatial-temporal information to generate effective feature masks. The Ind-CGRU effectively takes the advantages of GRU and convolution models in exploiting temporal information and mining spatial information respectively to improve the separation performance. The efficacy of proposed method has been verified on our ADSB datasets.

GNSS Interference Detection

Global Navigation Satellite System (GNSS) interference events that occur near airports can cause severe safety issues by denying GNSS based approaches and landings. Current solutions for interference detection and localization (IDL) such as using radio direction finding are generally costly and time-consuming. The approach described in this paper uses Automatic Dependent Surveillance — Broadcast (ADS-B) reports for IDL and applies machine learning algorithms to this data. We utilized a standard neural
network (NN) and a convolutional neural network (CNN) to detect GNSS interference event in a given airspace. These models take airplane’s ADS-B reports as inputs and output a classification of whether this airplane has experienced jamming.

In our model Neural Networks are used to determine the size and shape of the jammer impact region. By achieving this purpose, we are able to identify the complicated environmental factors from local airspace such as the signal blockage caused by the mountains, which are commonly difficult to identify or represent using mathematical models. Second, our Convolutional Neural Network is used for identifying the most likely location of the interference source which only requires ADS-B data collected within a few hours’ time window from target airspace. The key significant step to this approach is that it is mimicking the way of how humans identify the location of interference source which is by looking at the overall picture of the airspace. In addition, it also learns the complicated reasons of how interference source could cause impact on the overall picture of ADS-B data in current airspace.

When it comes to UAS, GNSS jamming is a major issue. GNSS is currently used to provide inputs in to the control loop of a drone or other autonomous vehicle, allowing it to maintain position, return to home or follow a series of preset waypoints. If this signal is conflicted or jammed then a UAS operator faces catastrophic consequences.

Aircraft Classification Using VHF Radar Signatures

Automatic target recognition has generated great interest in aeroplane traffic control and surveillance. The increasing number of radar opportunity illuminators, for example, radio transmitters, digital audio/video broadcast transmitters, and mobile-phone base stations, can be used as natural sources of radio waves in passive radar systems. Notably, broadcast commercial transmitters of FM radio (88–108 MHz) and VHF omnidirectional range (VOR, 108–118 MHz) prove attractive due to their high power transmission and wide coverage.

The radar cross section (RCS) in VHF could be a useful fingerprint to identify the class of aeroplane or UAS By exploiting the automatic dependent surveillance-broadcast (ADS-B) data broadcasted (L band) by the aeroplanes, their position and heading are then known and their BiStatic-RCS (BS-RCS), that is, RCS for separated transmitter and receiver, can be effectively estimated.

By mapping and comparing real ADS-B data from our ground station network to the BC-RCS output we are able to train our NN and CNN models to classify aircraft based on their BC-RCS which is also obtained from the VHF element within our DeFli Devices. This modelling allows us to granularly identify the size, type, manufacturer and potential payload of aircraft and crucially UAS including those that are not broadcasting on ADSB or Rf.

Contact us

Whether you have a request, a query, or want to work with us, use the form below to get in touch with our team.