Data Augmentation for Blind Signal Classification

An RFML Ecosystem: Considerations for the Application of Deep Learning to Spectrum Situational Awareness

“…RFML-based approaches aim to replace the human intelligence and domain expertise required to identify and characterize these features using deep neural networks and advanced architectures, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), to both blindly and automatically identify separating features and classify signals of interest, with minimal pre-processing and less a priori knowledge [18], [23]- [30]. Given the significant research in RFML-based modulation classification, it can be argued that AMC is one of the most mature fields in RFML, and has been deployed in real-world products [52].…”

“…Such datasets are useful in testing detection and classification performance of signals in a congested or interference-heavy environment with real-world transmitted signals. An additional augmentation technique often used includes adding synthetic noise to real world captures, which decreases the SNR without performing additional signal captures, thereby increasing the range of test SNRs [30], [63].…”

“…Perhaps the most important of these open questions is how to balance the amount of real, synthetic, and augmented data used in training datasets to avoid network bias. Work in [30] and [57] both examined data augmentation in the context of AMC, with results in [57] showing that for a fixed quantity of captured data, augmentations which consider the distribution of receiver degradations (SNR, Frequency Offset, Sample Rate Mismatch) can improve the performance of a model over captured data alone. While both works showed that augmentation of the original data result in increased performance, particularly when the degradation distribution is considered, no conclusions were drawn as to how performance changes in response to varying levels of augmentation.…”

Training data augmentation for deep learning radio frequency systems

“…For a given problem within the RFML space, once a reliable training routine and a network of sufficient size have been identified, how well a trained network is able to solve the problem often comes down to the quantity and quality of the data available. 7 Effectively, there are three sources of data that can be used to train networks within the RFML space: simulated/synthetic, 5, captured/collected, 4,6,11,12,18,23,35,37,48,54,[68][69][70][71][72][73][74][75][76][77][78][79][80][81][82][83][84][85] and augmented, 11,47,48,64,71,78 which is a combination of the first two using domain knowledge (focus of this work), or using generative adversarial networks (GAN) as performed in Davaslioglu et al 47 Due to the nature of the RFML data space, simulated data are inexpensive thanks to opensource tool-kits like GNU Radio, where observations can be generated uniquely in parallel, with the only bottleneck being the available computer resources. 30 Comparatively, performing an over-the-air (OTA) collection costs many orders of magnitude greater in terms of time and money due to procurement and configuration of the hardware transceivers and having to generate data in real time rather than in parallel as is done in simulation, yet all the work done in order to simulate the data is still needed when not examining commercial off-the-shelf (COTS) equipment...…”

“…In Wang et al, using synthetic permutations of an additive white Gaussian noise (AWGN) class, and combining that with other classes, was enough to increase the performance of their network in the Army Rapid Capabilities Office's (RCO) Blind Signal Classification Competition with an augmentation factor of seven, i.e., adding seven augmentations per observation to the original dataset. 64 In order to understand the impact that augmentation brings to RFML, an application space is needed without loss of generalization. Due to the widely studied signal classification problem of automatic modulation classification (AMC) within RFML, the AMC problem space provides a good way to test the promise of augmentation in RFML data without having to perform a full exploratory study determining the network and training routines needed in order to perform well.…”