US considers the Electromagnetic domain that is the electromagnetic spectrum portion of the information environment as a domain of Warfare. Assured access to the RF portion of the electromagnetic spectrum is critical to communications, radar sensing, command and control, time transfer, and geo-location and therefore for conducting military operations. The electromagnetic spectrum now includes complete range of frequencies of electromagnetic radiation from zero to infinity . Electromagnetic weapons have now become common in all three domains Air, Land and Water.
The US Electromagnetic Spectrum Strategy 2013, calls for ensuring the access to the congested and contested electromagnetic environment of the future, by adopting new agile and opportunistic spectrum operations, and through systems which are more efficient, flexible and adaptable and adopting new technologies capable of more efficient use of the spectrum and reduced risk of interference.
Artificial intelligence (AI) and machine learning (ML) as an application of AI, has today become an inevitable part of major industries such as healthcare, financial trending, and transportation. For many years machine learning approaches have been successfully applied to numerous detection and classification tasks from image processing to voice separation and text recognition. However, it is only recently that similar techniques have been applied to the processing of radio frequency (RF) signals and the electromagnetic environment (EME).
One such promising technology is the deep neural network (DNN). DNN classification algorithms have shown significant capability to process audio signals of similar structure with high accuracy for applications as varied as music recognition, speaker identification, earthquake detection, and gunfire localization. While DNNs are just beginning to be applied to radio frequency signals, the signal detection and classification is a prime candidate for this technology.
Future urgent need to intelligently utilize wireless resources to meet the need of ever-increasing diversity in services and user behavior, has actuated the wireless communication industry to deploy AI and ML techniques.
Spectrum usage has greatly increased with the explosion of IoT devices and LTE/5G enabled cell phones. In turn, the spectrum has become more and more congested, which can degrade network performance and reliability. The EME is becoming more congested, contested and complex. This can be seen by:
- Machine learning enabled RF smart sensor systems being deployed in domains as diverse as medical diagnostics, driverless vehicles, satellites, defence and agriculture.
- An increase in the deployment of Internet of Things (IoT) devices.
- An increase in more complex communications in the form of MIMO rollout of 5G and development of WiFi6.
- The number of commercial and civilian satellite launches with synthetic aperture radar (SAR) capability set to pass 60 this year. (Rosen, J., 2021).
Historically, the spectrum has been managed by forcing each communication system to operate in a fixed frequency range allowing spectrum management to be simple, but requiring a-priori knowledge of user requirements. As technology changes, so do the needs of users, which leads to scenarios where certain bands are underutilized, while others are congested. A more advanced approach is to allow for dynamic spectrum allocation to maximize utilization and prioritize usage. This approach is typically referred to as spectrum sharing.
It is therefore becoming increasingly difficult for a human in the loop to handle the flood of information. This is resulting in the adoption of deep learning approaches for the detection, classification, identification and transmission of signals.
ML for RF covers a wide range of scales in terms of distances, frequencies, and applications. Small scale passive systems are used for monitoring health and in a COVID world and beyond wireless IoT technologies dominate our day-to-day home lives. Signals intelligence, electronic warfare and communications are increasingly seeing the need to develop new approaches to automate the detection, classification, and identification of signals, from urban scale analytics to larger scale signals intercept on airborne platforms for situational awareness.
At an earth observation scale Interferometric Synthetic Aperture Radar (InSAR) is being used to automatically extract features in the difference in phase between satellites. This is being used to detect earthquakes, monitor subsidence, and track ice flows to monitor the effects of climate change. All of these processes cover a range of frequencies from oscillations on the scale of an atom to the size of a football pitch. Our ability to successfully deploy ML algorithms at such a wide range of scales depends on our ability to successfully adapt solutions to domain specific applications.
Application of Machine Learning in Cognitive Radio
Intelligent wireless communication aims at maximizing the efficiency of resource allocation by enabling the system to first recognize the available resources, then perceive and learn the wireless environment, and finally reconfigure its operating mode to adapt to the perceived wireless environment. The cognition capability and reconfigurability are the essential features of intelligent radio (also called cognitive radio-CR) in which machine learning techniques mark immense potential in system adaptation.
The concept of CR was first proposed by Joseph Mitola in order for mitigating the scarcity in limited radio spectrum by improving spectrum allocation efficiency by allowing unlicensed users (cognitive radio users) to identify and transmit over the frequency bands which are already assigned to the licensed users (primary users), but idle over specific time/space (spectrum holes). Spectrum hole identification is usually achieved by sensing the radio frequency (RF) environment through a process called spectrum sensing. It should be noted that the title of CR is not limited only to unlicensed wireless users and implied to any wireless user that adds the cognition capability along with reconfigurability to its system function.
The current applications of machine learning in CR can be listed as follows.
- Spectrum sensing and management: One application of machine learning techniques in realizing intelligent wireless communication is to track the occupancy statistics of the primary user and estimate the detection performance of the CR users. This process will lead to detecting the subbands, which are sensed and accessed consistently as spectrum holes. Sensing the potential subbands instead of the entire frequency band will minimize the number of sensors and, accordingly, improves energy efficiency.
- Power allocation: The strict limit for the aggregated interference caused by CR users on the primary network brings the role of efficient power allocation for CR users into the picture. The main goal here is to allocate power to the CR users such that not only meet the primary user interference constraint but maximize the quality of CR user’s provided service. Various learning algorithms have been proposed to accomplish this task among which reinforcement learning is the most popular one.
- Radio access technology: Given the numerous wireless technologies over the same frequency band, automatic network recognition is an important task that necessitates the application of machine learning. The main task here is to classify technologies and interference entities operating over the same band of interest.
- Signal classification: CR users often require a capability to recognize used waveforms either for communications or detection purposes. Multi-class signal classification based on automatic modulation recognition is the famous task in this regard and shines another application of machine learning in intelligent wireless communications.
- Medium access control (MAC) protocol identification: Sensing and identifying the MAC protocol types of any existing transmissions will be used by CR users to adaptively change their transmission parameters to not only improve spectrum utilization but facilitate the communications among heterogeneous CR networks. The identification capability here necessitates the utilization of machine learning techniques.
- Attack detection: Machine learning has been widely proposed to be used in jamming detection at the base stations.
Machine learning (ML) for military and commercial radars
Many of the characteristics of RF signals that are exploited to enable long range imaging, transmission and communication without direct line of sight, create a new set of challenges and opportunities for ML algorithms intended to learn and monitor activity. Examples of this include RF propagation effects from multipath in urban environments and diffraction from high water vapour content in the atmosphere. Similarly, the development of covert capability such as passive radar and low probability of intercept waveforms mean ML algorithms need to be resilient to a wide range of dynamic ranges, interference and low signal to noise ratios.
Both military and commercial radars are exhibiting ever increasing levels of agility across multiple parameters and over short timescales. Such signals provide a challenge for electronic surveillance receivers attempting to detect, cluster, separate and identify radars in a contested and congested EME. Processing techniques relying on a-priori knowledge of expected signals in the environment will be limited in their performance, and as such this provides an opportunity for the application of novel ML approaches to the aforementioned processes.
The modern agility of radars provides both a challenge for detection but an opportunity for the application of novel approaches for spectrum sharing and waveform distribution and design. Traditionally the spectrum was managed by operating comms systems within a fixed bandwidth. ML approaches, e.g., in cognitive radios and radars, are now being used to adaptively change transmission parameters to improve spectrum utilization, optimize channel conditions and enable adaptive routing between multiple nodes and networks (Deepwave, 2021). To meet the demand for automatic network recognition and to build resilience in hostile environments, we need to be able to detect and classify overlapping RF signals from multiple sources operating over ever-increasing frequency bandwidths.
Machine learning techniques are increasingly being explored for protection against jamming and deception. It provides the means to see anomalies and unusual patterns. Being able to counter jamming requires the ability to detect the signal and automatically adapt to it. This could be by adapting your waveform or moving to another part of the EME. ML for jamming and deception detection requires an understanding and improved awareness of the operational EME. Spoof detection requires algorithms capable of identifying and distinguishing features often based on higher-order statistics and thus lends itself to ML. As radar systems gradually move towards using ML techniques themselves, waveform structure, timing and agility may all be used to concurrently optimise probability of detection while avoiding interception by an adversary. As such ML may be the only feasible concept for exploiting such signals.
Machine learning resilience in contested environments necessitates strong verification and validation of algorithms that requires drawing from a large community of experts. Developing universal test sets of data is crucial for benchmarking codes and explainability of methodologies. It is critical for user confidence and wider adoption that we move away from using ML algorithms as a black box, explore new methods for explainability of network performance, and start to encode uncertainties in our decision making and predictions. In training ML algorithms, the importance of pre-processing and choice of features and embeddings can often be overlooked compared to the choice of ML architectures and hyperparameter fine-tuning. It is important when testing algorithms to identify which parts of a new algorithm contribute to better performance as well as having a universal set of metrics to use for testing.
To tackle the scarcity of tagged real datasets synthetic dataset creation is in many cases being used to augment datasets. This is of particular relevance in defence, where complete databases of signals may not be available. As such development of, validation and verification of sufficiently large, variable and realistic datasets consisting of both real and synthetic data is of particular interest. Generating realistic RF datasets that incorporate the interactions between multiple sensors and consider interference is a big challenge. To auto-generate datasets that are representative of different types of real data we also need automatic methods for feature extraction which reflect aspects such as characteristic parameter ranges, and skews of distributions. We then need to find ways to map these features onto RF functional IDs and to understand how we can use features to identify and explain phenomena causing signal interactions with the environment.
Multi-source signal fusion and distribution
Multi-sensor distributed systems measure parameters independently then use signal processing techniques to combine observations. Distributing signals across multiple sensors can make operations more covert, increase platform agility, allow rapid switching between modalities and help to solve trade-offs between platform performance and Size Weight and Power (SWAP). Being able to integrate observations from multiple sensors can improve accuracy, reliability, and detectability, reduce ambiguity, increase spatial-temporal ranges, enhance resolution, increase the dimension of target observations, and help to resolve multipath, and improve SNR (Kong et al, 2020).
Enhanced integration of multi-platform systems operating in an agile and real-time way requires novel multi-source signal fusion and distribution techniques. ML techniques are being explored for rapid, efficient, automatic allocation, transmission and reception of signals across multiple platforms. Distributed systems need very accurate position and timing information. Whilst GPS and atomic clocks can help to maintain good coherence, signal processing is still currently used post acquisition to make a number of corrections. This is an example of the kind of operations which may begin to be replaced by ML algorithms to improve coherence, and perform timing and positioning corrections and adjustments in real time. How we acquire and integrate data from multi-user distributed sensors and use them to cross validate each other has many solutions in the realm of embedded hardware and software.
RF embedded hardware and software
A drive towards real-time distributed processing at the edge with reduced human in the loop is pushing solutions towards embedded hardware and software approaches. Hybrid computing architectures, and software defined radios for ML applications are rapidly advancing areas of technology from embedded control, to autonomy and Artificial Intelligence (AI). The strong coupling between hardware and software in the RF domain and the use of purpose-built deep learning accelerators will need to be exploited to meet future requirements for data retrieval and transmission as well as considerations of SWAP. For signal detection it will be desirable to adjust the amount of power investment to make it proportional to the level of interest in a particular signal, and we will be looking to determine whether a signal is interesting as early as possible (Mullins, R., 2021). Multi-purpose RF sensors with ML capability using embedded hardware and software will be used to detect RF signals including Wi-Fi, Bluetooth and cellular to exploit the order of magnitude mark up in speed compared to conventional techniques.
In the case of multiple sensors, we will be looking to control and adapt the power consumption, parameters and precision of each sensor to optimise our use of the available power. There will be opportunities to co-design sensors, pre-processing and neural networks (Mullins. R, 2021). We are beginning to see frameworks designed to generate efficient neural network accelerators perform automatic transferral of machine learning architectures to FPGAs (Mullins, R., 2020). This has multiple applications notably for improved situational awareness. Strategies for early exit from inference at different stages in network architectures are beginning to be explored (Laskaridis, S. et al, 2020). In-network computing is being used to offload standard applications to network devices to increase throughput by processing data as it traverses the network (Zilberman, N., 2020). In-network data processing on wireless sensor nodes can be used to collect data at multiple distributed sources and aggregate it on the way to its final destination (Leung.K, 2020). There is great potential for the use of ML for data aggregation and resource optimisation and allocation. Dynamic hardware adaptation is already enabling in-orbit satellite updates and partial reconfigurations. Autonomous, unmanned vehicles will require automatic algorithm updates to embedded hardware to meet changes in the environment, cross platform modifications and advances in technology often on legacy hardware.
Developing efficient ML solutions on smaller platforms requires the reduction of models, dynamic compression, compact representations and knowledge distillation using techniques such as pruning of networks, improving performance in lower precision modes, dimensionality reduction, and sparse layer representations. We need to have a good understanding of when COTs solutions are fit for purpose and situations where we require custom specialised hardware. There are a number of choices to be made about what processing should be done in hardware, what to do in software, where to perform computations at the edge and when to push back to the cloud. The answers to some of these questions are in many cases strongly linked to requirements for data security and anonymization.
Machine Learning for Radio Frequency (RF) Signatures Detection and Classification System
In 2019 , The US Army Communication-Electronics Research Development & Engineering Center (CERDEC) published SBIR on Machine Learning for Radio Frequency (RF) Signatures Detection and Classification System.
CERDEC is interested in experimenting with signals analysis tools which can assist Army operators with detecting and identifying radio frequency emissions. Recent advances in machine learning (ML) may be applicable to this problem space. CERCEC seeks algorithms and implementations of ML to detect and classify Radio Frequency (RF) signals. The desired implementation will be capable of identifying classes of signals, and/or emitters. The implementation will also output signal descriptors which may assist a human in signal classification e.g. modulation type, and bandwidth. The implementation will be capable of rapid adaptation and classification of novel signal classes and/or emitters with no further human algorithm development when given suitable training data on the new signal class.
PHASE I: Identify/generate necessary training data sets for detection and classification of signatures, the approach may include use of simulation to train a machine learning algorithm. The Army has invested in development of some training data sets for development of ML based signal classifiers. This task aims to explore the strengths and weaknesses of existing data sets and prepare a validated training set to be used in Phase II. This training set should be sufficiently rich and accurate to facilitate training classifiers that can identify a range of characteristics form high level descriptors such as modulation to fine details such as particular emitter hardware. This data set should be representative of congested environments where many different emitter types are simultaneously present.
PHASE II: Produce signatures detection and classification system. Acquire, and modify as required, a COTS hardware and software. Demonstrate such a system. Demonstrate ability to detect and classify signatures. Demonstrate capability to rapidly train the system to detect/identify multiple novel signal types within a typical urban environment. Deliver a prototype system to CERDEC for further testing.
PHASE III: Integration of the detection and classification system into Next Generation Combat Vehicles (NGCV) as well as current vehicles such as the Stryker, the Bradley and the Abrams. Understanding of the signal that the Active Protection System (APS) in these vehicles produces and if that signal might interfere with other vehicle software or provide its own signature that could be picked up by the enemy sensors. Integration of the system into commercial autonomous vehicles. Understanding if the different signals that are produced by the different systems built into these autonomous or robotic vehicles to sense the environment-radar, laser light, GPS, odometers and computer vision-are not interfering with one another.
References and Resources also include: