Feature-based Research Articles

Blind Modulation Classification Under Uncertain Noise Conditions: A Multitask Learning Approach

Blind modulation classification is widely used in various military and civilian applications. Characterized by its exemption from likelihood calculation and the ability of using well-trained models, feature-based (FB) classifiers are popular with their high computational efficiency. In recent years, Deep Neural Networks based methods have shown significant improvement in the classification accuracy. However, the performance of FB methods decreases rapidly under low signal-to-noise ratios. To overcome this shortcoming, this letter proposes a joint denoising and modulation classification method based on Multitask Learning, where a denoising network and a classification network are simultaneoulsy trained in an end-to-end manner. In addition, the focal loss function is adopted to highlight the importance of hard-to-classify samples during training. Our numerical experiments show that the proposed method can effectively improve the classification accuracy, and outperform state-of-the-art methods. For example, under the 0dB SNR condition, the performance of our proposed Multitask CNN method is 20% higher than that of the traditional CNN method.

Deep Learning-Based Automatic Modulation Classification Over MIMO Keyhole Channels

Automatic modulation classification (AMC) is a significant part of cognitive communication systems. In early researches, likelihood-based (LB) and feature-based (FB) solutions were proposed for the AMC problem. With the developments in the data-driven approaches, a third method based on deep learning (DL) has recently gained prominence among AMC researchers. It is shown that convolutional neural network based classifiers are very efficient in the AMC for both single input single output (SISO) and multiple-input multiple-output (MIMO) systems. However, for most of the works in MIMO-AMC, the channel considered is full rank. This work addresses the problem of AMC over rank deficient channels such as a keyhole channel using a DL-based classifier. The proposed DL-classifier utilizes a CNN, which does not employ pooling layers or dropouts in the convolutional layers. To further improve the classification accuracy, decision cooperation as well as feature fusion is employed. In addition to the keyhole effect, this work investigates the effect of antenna correlation on DL-based AMC. A comparative study of the proposed method and the existing FB AMC method for the MIMO keyhole channel is also presented.

A survey of traditional and advanced automatic modulation classification techniques, challenges, and some novel trends

SummaryAutomatic modulation classification (AMC) is an important stage in intelligent wireless communication receivers. It is a necessary process after signal detection, and before demodulation. It plays a vital role in various applications. Blind modulation classification is a very difficult task without information about the transmitted signal and the receiver parameters like carrier frequency, signal power, timing information, phase offset, existence of frequency‐selective multipath fading, and time‐varying channels in real‐world applications. The AMC methods are divided into traditional and advanced methods. Traditional methods include likelihood‐based (LB) and feature‐based (FB) methods. The advanced methods include deep learning (DL) methods. In addition, the AMC methods are used to classify different modulation schemes such as ASK, PSK, FSK, PAM, and QAM with different orders and different signal‐to‐noise ratios (SNRs). This paper focuses on summarizing the AMC methoods, comparing between them, presenting the commercial software packages for AMC, and finally considering the new challenges in the implementation of AMC.

AUTOMATIC MODULATION CLASSIFIER: REVIEW

The automatic modulation classification (AMC) is highly important to develop intelligent receivers in different military and civilian applications including signal intelligence, spectrum management, surveillance, signal confirmation, monitoring, interference identification, as well as counter channel jamming. Clearly, without knowing much information related to transmitted data and various indefinite parameters at receiver, like timing information, carrier frequency, signal power, phase offsets, and so on, the modulation’s blind identification has been a hard task in the real world situations with multi-path fading, frequency-selective in addition to the time-varying channels. There are 2 methods could be utilized to decide the classification signal technique: Feature-based (FB) approach and the Maximum likelihood functions (LB) method. With regard to the FB (referred to as pattern-recognition) classification method used in the study. In the presented work, thorough study is provided to find easy method to identify and classify the digital modulation signals at low SNRs. Spectral-based features, high-order statistic features, wavelet-based features, also cyclic features on the basis of cyclostationary typically utilized to determine and discriminate modulation types have been examined. The number of the classifiers which have been utilized in the process of discrimination have been studied thoroughly and compared for helping researchers in determining and finding the drawbacks with pattern-recognition according to past works. The presented study serving as guide with regard to studies of AMC for determining adequate algorithms and features.

A Novel Deep Learning and Polar Transformation Framework for an Adaptive Automatic Modulation Classification

Automatic Modulation Classification (AMC) is an approach to identify an observed signal's most likely modulation scheme without any a priori knowledge of the intercepted signal. In this research, the authors present a new direction for both stages of feature-based (FB) approach. In the feature extraction stage, the authors design a new architecture that 1) removes the bias issue for the estimator of fourth-order cumulants, and 2) extracts polar-transformed information of the received $IQ$ symbols, and finally 3) forms a unique dataset to be used in the labeling stage. Furthermore, the authors contribute to increasing the classification accuracy in low signal-to-noise ratio (SNR) conditions by employing the deep belief network (DBN) platform in addition to the spiking neural network (SNN) platform to overcome execution latency concerns associated with deep learning architectures. For this research, the authors first study each individual FB AMC classifier to derive their respective upper and lower performance bounds and then propose an adaptive framework that is built and developed with these findings. This framework aims to efficiently classify the modulation scheme by intelligently switching between these different FB classifiers to achieve an optimal balance between accuracy and execution latency for any observed channel conditions derived from the main receiver's equalizer. Subsequently, a performance analysis is conducted using the standard RadioML dataset to achieve a realistic evaluation. Numerical results indicate a notably higher classification accuracy, by 16.02% on average, when DBN is employed, whereas SNN requires significantly lower execution latency to label the modulation scheme when compared against two other modified FB classifiers that are built upon convolutional and recurrent neural networks, shown to be reduced by 34.31%.

Performance of Feature-Based Techniques for Automatic Digital Modulation Recognition and Classification—A Review

The demand for bandwidth-critical applications has stimulated the research community not only to develop new ways of communication, but also to use the existing spectrum efficiently. Networks have become dynamic and heterogeneous. Receivers have received various signals that can be modulated differently. Automatic modulation classification (AMC) is a key procedure for present and next-generation communication networks, and facilitates the demodulation process at the receiver side. Under the presence of noise from the channel, the transmitter and receiver with its unknown parameters, such as carrier frequency, phase offset, signal power, and timing information, have become cumbersome because detecting the modulation scheme of the received signal is a complicated procedure. Two main methods, namely maximum likelihood functions and the signal statistical feature-based (FB) approach, are used for the automatic classification of modulated signals. In this study, a comprehensive survey of various modulation techniques based on FB approach is conducted. In this research, a number of basic features that are usually used in determining and discriminating modulation types were investigated. The classifier that was used in the discrimination process is studied in detail and compared to other types of classifiers to help the reader determine the limitations associated with the FB approach. Both classifiers and basic features were compared, and their advantages and disadvantages were investigated based on previous researches to determine the best type of classifier and the set of features in relation to each discrimination environment. This work serves as a guide for researchers of AMC to determine the suitable features and algorithms.

Automatic Wireless Signal Classification: A Neural-Induced Support Vector Machine-Based Approach

Automatic Classification of Wireless Signals (ACWS), which is an intermediate step between signal detection and demodulation, is investigated in this paper. ACWS plays a crucial role in several military and non-military applications, by identifying interference sources and adversary attacks, to achieve efficient radio spectrum management. The performance of traditional feature-based (FB) classification approaches is limited due to their specific input feature set, which in turn results in poor generalization under unknown conditions. Therefore, in this paper, a novel feature-based classifier Neural-Induced Support Vector Machine (NSVM) is proposed, in which the features are learned automatically from raw input signals using Convolutional Neural Networks (CNN). The output of NSVM is given by a Gaussian Support Vector Machine (SVM), which takes the features learned by CNN as its input. The proposed scheme NSVM is trained as a single architecture, and in this way, it learns to minimize a margin-based loss instead of cross-entropy loss. The proposed scheme NSVM outperforms the traditional softmax-based CNN modulation classifier by managing faster convergence of accuracy and loss curves during training. Furthermore, the robustness of the NSVM classifier is verified by extensive simulation experiments under the presence of several non-ideal real-world channel impairments over a range of signal-to-noise ratio (SNR) values. The performance of NSVM is remarkable in classifying wireless signals, such as at low signal-to-noise ratio (SNR), the overall averaged classification accuracy is > 97% at SNR = −2 dB and at higher SNR it achieves overall classification accuracy at > 99%, when SNR = 10 dB. In addition to that, the analytical comparison with other studies shows the performance of NSVM is superior over a range of settings.

A robust modulation classification method using convolutional neural networks

Automatic modulation classification (AMC) is a core technique in noncooperative communication systems. In particular, feature-based (FB) AMC algorithms have been widely studied. Current FB AMC methods are commonly designed for a limited set of modulation and lack of generalization ability; to tackle this challenge, a robust AMC method using convolutional neural networks (CNN) is proposed in this paper. In total, 15 different modulation types are considered. The proposed method can classify the received signal directly without feature extracion, and it can automatically learn features from the received signals. The features learned by the CNN are presented and analyzed. The robust features of the received signals in a specific SNR range are studied. The accuracy of classification using CNN is shown to be remarkable, particularly for low SNRs. The generalization ability of robust features is also proven to be excellent using the support vector machine (SVM). Finally, to help us better understand the process of feature learning, some outputs of intermediate layers of the CNN are visualized.

Automatic Modulation Recognition of Radar Signals Based on Manhattan Distance-Based Features

Feature-based (FB) algorithms for automatic modulation recognition of radar signals have received much attention since they are usually simple to realize. However, existing FB approaches usually focus on several specific modulations and fail when applied to various modulations. To overcome this issue, we propose a simple and effective FB algorithm based on Manhattan distance-based features (MDBFs) in this paper. MDBFs are new features for radar signals that can be applied for recognition of different modulations. The main contributions of this paper are as follows. First, radar signals are represented as wavelet ridges, which includes important information that can distinguish different modulations, and the piecewise aggregate approximation algorithm is introduced to reduce signal dimensions. Then, the dynamic time warping averaging is employed instead of the traditional k-means algorithm to extract realistic centroids for each class. Finally, the Manhattan distances between each data sample and each centroid are used to construct MDBFs, and decisions are made using the k-nearest neighbor. In addition, we prove that MDBFs have better class separability power than the Euclidean-based features. MDBFs contain information about the correlations between different classes, which means that these features suitable for discriminating various modulations when their class distributions do not overlap badly in representation space. The extensive experiments on a synthetic dataset demonstrate the outstanding performance of our proposed method and are hardly affected by the pulse width of the signal. Thus, the proposed method with the effectiveness and robustness could be a promising modulation recognition method of the radar signal.

Estimation of Left Ventricular Motion from Cardiac Gated Tagged MRI Using an Image-Matching Deformable Mesh Model

In this paper, we propose a deformable mesh model (DMM) method for left-ventricular (LV) motion-estimation from cardiac-gated tagged magnetic resonance imaging (MRI). DMM starts by partitioning short-axis (SA) and long-axis (LA) images at end-diastole (ED) in a collection of non-overlapping mesh elements (triangles). Next, mesh elements are deformed attempting to track temporal changes in pixel intensities by minimizing an image-matching cost function from frame to frame. Finally, dense motion fields can be obtained at any frame from the displacement of mesh nodes, without any additional user interaction. For validation, 23 human tagged MRI studies were used, allocated as 7 normal, 9 hypertensive (HTN), and 7 with mitral regurgitation (MR). DMM end-systolic (ES) motion and strains were quantitatively compared to a “gold-standard” feature-based (FB) motion analysis method that includes manual tagline editing. DMM motion achieves subpixel average displacement error of SA landmarks with respect to FB. 3-D ES strains obtained from DMM correlate well with FB method (ρ ranging from 0.79 to 0.97). Finally, we show temporal curves for several DMM strain components, grouped according to the condition of patients (normal and diseased). These curves are qualitatively compared to a strain from unwrapped harmonic phase motion estimation algorithm. DMM produces robust temporal curves, which capture strain fluctuations reported in medical literature for HTN and MR conditions. This makes it a viable alternative to analyze motion from images at phases suffering from low contrast-to-noise ratio that may convey important information about the presence of certain cardiac pathologies.

Robust Automatic Modulation Classification Under Varying Noise Conditions

Automatic modulation classification (AMC) plays a key role in non-cooperative communication systems. Feature-based (FB) methods have been widely studied in particular. Most existing FB methods are deployed at a fixed SNR level, and the pre-trained classifiers may no longer be effective when the SNR level changes. The classifiers may also need to be re-trained to be suitable for the varying channel environment. To address these problems, a robust AMC method under varying noise conditions is proposed in this paper. The method attempts to select noise-insensitive features from a large feature set to ensure that the trained classifiers will be robust to SNR variations. First, a feature set consisting of 25 types of features is extracted, and 4 features that are insensitive to noise are chosen through a feature selection method based on rough set theory. The generalizability of an SVM classifier trained on the 4 chosen features is evaluated based on numerical results. The classification accuracy remains reasonable when the SNR varies between 5 and 20 dB, indicating that the proposed method can be deployed under varying noise conditions.

Signal identification for emerging intelligent radios: classical problems and new challenges

Signal identification, which initially found applications in electronic warfare and spectrum monitoring and surveillance, has been recently considered for commercial communications in the context of software defined and cognitive radios. In this article, I present a snapshot of the status of signal identification algorithms, starting from a general description of maximum likelihood (ML) and feature based (FB) approaches to a more detailed discussion of a practical methodology using cyclostationarity-based features. I discuss the cyclostationarity-based features of various signals and the criteria of decision for their identification, while considering classical problems of identifying single carrier linearly digitally (SCLD) modulated signals, as well as new challenges posed by the identification of orthogonal frequency division multiplexing (OFDM), SC frequency domain equalization (SC-FDE), and multiple-transmit antenna signals. I conclude the article with remarks on practical solutions to signal identification and open research issues.

3D+time left ventricular strain by unwrapping harmonic phase with graph cuts.

In previous work, a three-dimensional left ventricular strain throughout the cardiac cycle was reconstructed using a prolate spheroidal B-spline (PSB) method with displacement measurements obtained from unwrapped tagged MRI (tMRI) harmonic phase images. Manually placed branch cuts were required for each harmonic phase image to resolve phase inconsistencies and to guide the phase unwrapping (mSUP), which is both labor intensive and time consuming and therefore not proper for clinic application. In this paper, we present an automated graph cuts based phase unwrapping method for myocardium displacement measurement (caSUP) which can be used to compute 3D+time cardiac strain. A set of 8 human studies were used to optimize parameters of the energy function and another set of 32 human studies were used to validate the proposed method by comparing resulted strains with those from mSUP and a feature-based (FB) method using the same PSB strain reconstruction. The automated caSUP strains were close to the manual strains and only required 6 minutes after myocardium segmentation versus - 2 hours for the manual method.

Three‐dimensional plus time biventricular strain from tagged MR images by phase‐unwrapped harmonic phase

To validate a method called bi-ventricular strain unwrapped phase (BiSUP) for reconstructing three-dimensional plus time (3D+t) biventricular strain maps from phase-unwrapped harmonic phase (HARP) images derived from tagged cardiac magnetic resonance imaging (MRI). A set of 30 human subjects were imaged with tagged MRI. In each study, HARP phase was computed and unwrapped in each short-axis and long-axis image. Inconsistencies in unwrapped phase were resolved using branch cuts manually placed with a graphical user interface. The 3D strain maps were computed independently in each imaged time frame through systole and mid diastole in each study. The BiSUP strain and displacements were compared with those estimated by a 3D feature-based (FB) technique and a 2D+t HARP technique. The standard deviation of the difference between strains measured by the FB and the BiSUP methods was less than 4% of the average of the strains from the two methods. The correlation between peak minimum principal strain measured using the BiSUP and HARP techniques was over 83%. The BiSUP technique can reconstruct full 3D+t strain maps from tagged MR images through the cardiac cycle in a reasonable amount of time and user interaction compared with other 3D analysis methods.

3D left ventricular strain from unwrapped harmonic phase measurements

To validate a method for measuring 3D left ventricular (LV) strain from phase-unwrapped harmonic phase (HARP) images derived from tagged cardiac magnetic resonance imaging (MRI). A set of 40 human subjects were imaged with tagged MRI. In each study the HARP phase was computed and unwrapped in each short-axis and long-axis image. Inconsistencies in unwrapped phase were resolved using branch cuts manually placed with a graphical user interface. 3D strain maps were computed for all imaged timeframes in each study. The strain from unwrapped phase (SUP) and displacements were compared to those estimated by a feature-based (FB) technique and a HARP technique. 3D strain was computed in each timeframe through systole and mid-diastole in approximately 30 minutes per study. The standard deviation of the difference between strains measured by the FB and the SUP methods was less than 5% of the average of the strains from the two methods. The correlation between peak circumferential strain measured using the SUP and HARP techniques was over 83%. The SUP technique can reconstruct full 3D strain maps from tagged MR images through the cardiac cycle in a reasonable amount of time and user interaction compared to other 3D analysis methods.

Improving the Ocean Initialization of Coupled Hurricane–Ocean Models Using Feature-Based Data Assimilation

Abstract Coupled hurricane–ocean forecast models require proper initialization of the ocean thermal structure. Here, a new feature-based (F-B) ocean initialization procedure in the GFDL/University of Rhode Island (URI) coupled hurricane prediction system is presented to account for spatial and temporal variability of mesoscale oceanic features in the Gulf of Mexico, including the Loop Current (LC), Loop Current eddies [i.e., warm-core rings (WCRs)], and cold-core rings (CCRs). Using only near-real-time satellite altimetry for the “SHA-assimilated” case, the LC, a single WCR, and a single CCR are assimilated into NAVOCEANO’s Global Digitized Environmental Model (GDEM) ocean temperature and salinity climatology along with satellite-derived daily sea surface temperature (SST) data from 15 September 2005 to produce a more realistic three-dimensional temperature field valid on the model initialization date (15 September 2005). For the “fully assimilated” case, both near-real-time altimetry and real-time in situ airborne XBT (AXBT) temperature profiles are assimilated into GDEM along with SST to produce the three-dimensional temperature field. Vertical profiles from the resulting SHA-assimilated and fully assimilated temperature fields are compared to 18 real-time AXBT temperature profiles, the ocean climatology (GDEM), and an alternative data-assimilated product [the daily North and Equatorial Atlantic Ocean Prediction System Best Estimate (RSMAS HYCOM), which uses an Optimal Interpolation (OI) based assimilation technique] to determine the relative accuracy of the F-B initialization procedure presented here. Also, the tropical cyclone heat potential (TCHP) from each of these profiles is calculated by integrating the oceanic heat content from the surface to the depth of the 26°C isotherm. Assuming the AXBT profiles are truth, the TCHP rms error for the F-B SHA-assimilated case, the F-B fully assimilated case, the GDEM ocean climatology, and the RSMAS HYCOM product is 12, 10, 45, and 26 kJ cm−2, respectively.