Search

The concept of neuromuscular activity recognition using instantaneous high-density surface electromyography (HD-sEMG) images opens up new avenues for the development of more fluid and natural muscle-computer interfaces. However, the existing approaches employed a very large deep convolutional neural network (ConvNet) architecture and complex training schemes for HD-sEMG image recognition, which requires the network architecture to be pre-trained on a very large-scale labeled training dataset, as a result, it makes computationally very expensive. To overcome this problem, we propose S-ConvNet and All-ConvNet models, a simple yet efficient framework for learning instantaneous HD-sEMG images from scratch for neuromuscular activity recognition. Without using any pre-trained models, our proposed S-ConvNet and All-ConvNet demonstrate very competitive recognition accuracy to the more complex state of the art for neuromuscular activity recognition based on instantaneous HD-sEMG images, while using a ~ 12 x smaller dataset and reducing learning parameters to a large extent. The experimental results proved that the S-ConvNet and All-ConvNet are highly effective for learning discriminative features for instantaneous HD-sEMG image recognition especially in the data and high-end resource constrained scenarios.

Indoor location information is increasing in importance in contemporary communication services and applications. In this paper, we discuss the long short-term memory (LSTM) performance for indoor localization in non-line-of-sight (NLoS) conditions using the received signal strength (RSS) and channel state information (CSI) obtained from Wi-Fi signals. As such, we describe the CSI and RSS acquisition system that is used to build a rich dataset to experiment with classical machine learning and deep learning models. The distance range error matrix is combined with the confusion matrix to obtain the distance range error probability where we have demonstrated that the LSTM model exhibits a maximum range error of less than 5 m with 4% probability.

Low-power wide-area networks (LPWANs) increasingly attract attention in the IoT community. The provision of long communication ranges with low energy consumption is the main reason behind LPWAN’s growing popularity. Energy efficiency is crucial for LPWAN devices that are mostly battery-powered and required to function in a crowded environment. To reduce energy consumption over these networks, minimizing the collision rate in the packet transmission process is one of the possible solutions. However, existing radio resource allocation management algorithms do not fulfill the energy efficiency required by IoT devices. We propose Energy Efficient Exponentially Weighted Algorithm Based Resource Allocation, which considers each packet’s energy consumption level and transmission time in learning the best set of resources to be allocated to each end-device. We achieve 30% lower energy consumption per packet transmission than the baseline methods, which is noticeable when considering the whole network packet transmission, at the expense of losing 2% of the successful transmission rate.

The unmatched threat of Android malware has tremendously increased the need for analyzing prominent malware samples. There are remarkable efforts in static and dynamic malware analysis using static features and API calls respectively. Nonetheless, there is a void to classify Android malware by analyzing its behavior using multiple dynamic characteristics. This paper proposes EntropLyzer, an entropy-based behavioral analysis technique for classifying the behavior of 12 eminent Android malware categories and 147 malware families taken from CCCS-CIC-AndMal2020 dataset. This work uses six classes of dynamic characteristics including memory, API, network, logcat, battery, and process to classify and characterize Android malware. Results reveal that the entropy-based analysis successfully determines the behavior of all malware categories and most of the malware families before and after rebooting the emulator.

This paper proposes a method to generate multisines signals with reduced peak derivative. This is done by adjusting the phases in the Fourier domain with Genetic Algorithm such that in the time domain the maximum absolute signal's derivative is minimized. The performance evaluation is made by comparing the method to signals with the same spectrum but with phase adjusted randomly. It will be shown for an audio test signal case study of a multisines signal comprising 2500 sines that a reduction of 42% with respect to the mean maximum peak derivative and 31% with respect to the minimum maximum peak derivative over 1 million random trials. The proposed method is contrasted with a low crest factor signal method. It is found that there is a positive correlation between peak amplitude and peak derivative when signals are generated with random phases. Moreover, the signals obtained by minimizing either the peak amplitude of the peak derivative had also good performances with respect to the non-optimized criterion.

This paper shows how adding a second step of windowing after each phase randomization can reduce the False Rejection Rate in Fourier based Surrogate Analysis. Windowing techniques improve the resolution of the Power Spectrum estimation by reducing the sampling gap caused by the periodic extension of the Fourier Series. However, it adds a time domain non-stationarity which affects the Surrogate Analysis. This effect is particularly problematic for short lowpass signals. Applying the same window to the surrogate data allows having the same non-stationarity. The method is tested on order 1 autoregressive process null hypothesis by Monte Carlo simulations. Previous methods were not able to yield good performances for left-sided and right-sided tests at the same time, even less with bilateral tests. It is shown that the new method is conservative for unilateral tests as well as bilateral tests.

Over the last few years, the application of Digital Signal Processing (DSP) techniques for genomic sequence analysis has received great interest. Indeed, among its applications in genomic analysis, it has been demonstrated that DSP can be used to detect protein coding regions (exons) among non-coding regions in a DNA sequence. The period-3 behavior exhibited by exons is one of its features that has been exploited in several developed algorithms for exon prediction. Identification of this periodicity in genomic sequences can be done by using different methods such as the well-known Fast Fourier Transform (FFT) and the Goertzel algorithm for complexity reduction in which the reduction of computational time is a great challenge in genomic analysis. Therefore, this paper presents a novel one frequency analysis by using half of the arithmetic complexity of the Goertzel algorithm for gene prediction. Compared to the Intel®’s FFT (MKL) optimized function, the Goertzel’s (IPP) and the dedicated Goertzel compiled function with ICC on Xeon CPU (24 cores), the proposed method conserves the same accuracy provided by the referenced methods which will manifest a speedup of 3000, 10 and 2 compared to MKL FFT, IPP Goertzel and the dedicated Goertzel with ICC, respectively.

Neuromuscular activity recognition using low- resolution instantaneous high-density surface electromyography (HD-sEMG) images present a great challenge. The recent result shows the high potentiality and hence opens up new avenues for the development of more fluid and natural muscle-computer interfaces. However, the existing approaches employed a very large deep ConvNet, which requires learning >5.63 million training parameters only during fine-tuning and pre-trained on a very large-scale labeled HD-sEMG training datasets, as a result, it makes high-end resource bounded and computationally expensive. To overcome this problem, we propose a lightweight All-ConvNet model that consists solely of convolutional layers, a simple yet efficient framework for learning instantaneous HD-sEMG images from scratch through random initialization. Without using any pre-trained models, our proposed lightweight All-ConvNet demonstrate very competitive or even state of the art performance on a current benchmarks HD-sEMG dataset, while requires learning only ~460k training parameters and using ~12xsmaller dataset. The experimental results proved that the proposed lightweight All-ConvNet is highly effective for learning discriminative features for low-resolution instantaneous HD-sEMG image recognition and low-latency processing especially in the data and high-end resource constrained scenarios.

This paper presents two novel substrate integrated waveguide (SIW) sensing-elements for microfluidic biosensing and environmental detection. The first proposed structure is a SIW interferometer for biological liquids characterization. The operation principle of the latter is based on the variation of the effective dielectric constant of a sensitive branch due to the introduction of aqueous solutions into a sensitive region of a total volume of ~1 µL. To achieve a destructive interference, the signal division and combination have been carried out using microstrip Wilkinson power dividers. The sensitive characteristics of the device have been tested by measuring two buffer solutions; Phosphate-buffered saline (PBS) and Roswell Park Memorial Institute medium (RPMI). In regards to the second proposed device, it consists of an H-plane SIW-based horn antenna demonstrated for relative humidity (RH) sensing. The sensitive characteristics of the proposed antenna without the use of a sensitive layer were tested in the range of 25%–75% RH. The combination of antenna and sensor functions in a single substrate integrated device offers multiple advantages and enables the development of simple, compact and cost-effective sensors. Passive RFID sensing is also possible with this technique without resorting to the use of additional sensors.

This paper shows how a relaxation of the high frequency requirements can help reducing the latency in linear phase interpolation filter, with an audio production system perspective. The reduced need for attenuation is justified when the interpolation filter is followed by a noise-shaping Delta-Sigma loop and an analog filtering stage. This is done by using a non-constant error weight of the stop-band. In order to use the Parks-McClellan method for finite impulse response filter design from Matlab, the stop-band is divided and weighted logarithmically. Quantitative results are shown for different example filter design, limited to situations where the Parks-McClellan converges well. It has been found that the shorter the filter length needed to respect a given filter template, the more relative group delay reduction can be achieved by relaxing the high frequency requirement. For filter size of the order of 100, reduction of group delay of 30% can be expected. For sake of simplicity, the Delta-Sigma loop is discussed but not analysed here. The idea is demonstrated in the context of Digital-to-Analog converters (DAC) but by duality could be applicable also to Analog-to-Digital converters (ADC). The main performance metric used is a relative reduction of the impulse response group delay. The results are also presented as impulse responses and power spectrum examples. The presented approach may be generalised to complex and non-linear phase filters and does not prevent the use of polyphase structures.

Fast Fourier transform (FFT) is one of the fundamental processing block used in many signal processing applications (i.e. for orthogonal frequency division multiplexing in wireless telecommunication). Therefore, every proposal to reduced latency, resources or accuracy errors of FFT implementation counts. This paper proposes the implementation of the butterfly processing elements (BPE) where the concept of the radix-r butterfly computation has been formulated as the combination of α radix-2 butterflies implemented in parallel. An efficient FFT implementation is feasible using our proposed multiplexed and pipelined BPE. Compared to a state-of-the-art reference based on pipelined and parallel structure FFTs, and FPGA based implementation reveals that the maximum throughput is improved by a factor of 1.3 for a 256-point FFT and reach a throughput of 2680 MSps on Virtex-7. The analysis extends to touch on key performance measurements metrics such as throughput, latency and resource utilization.

For the problem of acoustic echo cancellation (AEC) with nonlinear distortions, we propose to use a linear adaptive filter that maximizes the Correntropy similarity measure instead of the conventional minimization of the mean squared error (MSE) criterion. The maximum Correntropy criterion (MCC) offers robustness to outliers and impulsive noise, which is interesting for the case of speech signal coupled with nonlinearities. To assess the performance of the algorithm, we consider a hard-clipping memoryless saturation nonlinearity. Our simulation results show very interesting performance of the normalized MCC-based linear adaptive filter for the echo return loss enhancement (ERLE) and misalignment measures compared to the MSE-based normalized least mean squares (NLMS) approach. Furthermore, the NMCC adaptive filter has a similar computational complexity as the NLMS algorithm, which makes it very attractive in practical implementations

In this paper, we propose an efficient precoding solution for a correlated Multi-Input Multi-Output (MIMO) broadcast channels. Indeed, we apply the principal component analysis (PCA) to uncorrelate (whiten) the channel prior to codeword selection. As such, the channel state information at the transmitter (CSIT), picked up through finite-rate feedback, corresponds to the uncorrelated channel version. Thus, the optimality of the approach. The simulation results compared with conventional MIMO precoders for various levels of spatial correlation as well as different receive antenna settings, show that the proposed scheme provides greater system performance enhancement in terms of sum rate.

In large scale multiple-input multiple-output (MIMO), high spectral and energy efficiencies comes at the expense of a high computational complexity baseband processing. Many contributions have been proposed to reduce such complexity using matrix inversion approximation techniques for instance. On the other hand, to reduce the constraint on the interconnects’ bandwidth, fewer decentralized processing techniques have emerged. Here, we propose a computationally efficient technique based on embedding one single Gauss-Seidel iteration within every ADMM based detection iteration. The simulations are performed using an LTE-like TDD-OFDM frame structure and waveform, under perfect and non-perfect channel state information (CSI). Early results reveal that the proposed ADMM-GS algorithm can outperform the centralised GS based technique processing in a high SNR region and high load regime. In addition ADMM-GS’ performance exhibits relatively less sensitivity to channel estimation error; a characteristic inherited from the centralised GS technique.

Solving Complex Problem that is coupled with intensive workloads; necessities the access to a massively parallel computational power. Up to date, Graphic Processing Units (GPUs) are the only architecture that could handle the most complex computationally intensive workloads. In the light of this rapid-growing advancement in computational technologies, this paper will propose a high-performance parallel radix-23 FFT suitable for such GPU and CPU systems. The proposed algorithm could reduce the computational complexity by a factor that tends to reach p r if implemented in parallel (pr is the number of cores/threads) plus the combination phase to complete the required FFT.

The Surrogate Analysis (SA) is known to detect nonlinear signals, non-stationary signals and ARMA systems driven by non-Gaussian processes. This paper adds to address the detection of non-random phase signal, of which the linear phase signal is the best-known example. This is a new interpretation of the SA. In order to highlights the benefits of the interpretation, a new theoretical signals is constructed. The signal has a perfect Gaussian distribution and is not affected by periodic extension and is a linear phase signal. The SA will be shown able to detect this signal in a noise with exactly the same power spectrum. It will be clear that the SA is able to detect phase linearity even when the data is normally distributed. An application of the detection by SA is given regarding very noisy and short time electrocardiogram (ECG) signal and compared to higher order statistics and normality tests for this purpose.

In this paper we consider an efficient method to resolve the underlying large matrix inversion problem inherent in the antennas' mutual coupling based reciprocity calibration. Such calibration enables the downlink pre-coding using the uplink channel estimates in a time-division-duplex (TDD) massive MIMO systems. Based on least squares estimators, a large matrix inversion is required. Herein, we derive an efficient method based on successively updating a matrix inverse by exploiting the Gram matrix structure. The simulation results reveal that our proposed method performs as well as the direct matrix inversion (based on Cholesky decomposition) whereas the approximation techniques based on Gauss Seidel (GS) and Neumann series expansions (NSE) require a large number of iterations. The proposed method is computationally efficient and lend itself for an efficient parallel and pipelined architecture implementation.

An efficient pattern recognition system based exclusively on forearm surface Electromyographic (sEMG) signals is proposed to provide a more intuitive control of a robotic arm used by some of the disabled. The main contribution of this paper is the use of an original set of features characterizing the muscle activity distribution obtained with high-density sEMG (HD sEMG) sensors. Contrary to simple sEMG, HD sEMG can produce muscle activity images with spatial distributions that differ according to forearm movement. In order to translate this distribution, the proposed set of features includes the center of gravity, the mean amplitude and the percentage of influence computed in each HD sEMG image divided in sub-images. Based on these features, the recognition system locates nine forearm movements with high classification accuracies (99.23%). The results in terms of the number of learning data, the image resolutions (spatial filtering) and the number of sub-images demonstrate the potential of the proposed recognition system and its good performance-complexity trade-off.

In this paper, we propose a direct nonlinear optimization method to solve the Boolean Compressive Sensing (BCS) problem for large signals when sparsity level is unknown. While traditional CS results from linear Algebra, BCS, is given by logical operation in the Boolean workspace. To overcome this inconvenience, we relax the problem in an equivalent formulation in the Real workspace using appropriate modeling as a first step. Thereafter we turn out the problem in an unconstrained form that can be solved directly by nonlinear optimization method called Trust Region methods (TRM). Our solution is based on an Approximate version of (TRM). Numerical results are presented to sustain efficiency of our proposal.

The concept of neuromuscular activity recognition using instantaneous high-density surface electromyography (HD-sEMG) image opens up new avenues for the development of more fluid and natural muscle-computer interfaces. The state-of-the-art methods for instantaneous HD-sEMG image recognition achieve prominent performance using a computationally intensive deep convolutional networks (ConvNet) classifier, while very low performance is reported using the conventional classifiers. However, the conventional classifiers such as Support Vector Machines (SVM) can surpass ConvNet at producing optimal classification if well-behaved feature vectors are provided. This paper studies the question of extracting distinctive feature sets, thus propose to use Histograms of Oriented Gradient (HOG) as unique features for robust neuromuscular activity recognition, adopting pair wise SVMs as the classification scheme. The experimental results proved that the HOG represents unique features inside the instantaneous HD-sEMG image and fine-tuning the hyper- parameter of the pair wise SVMs, the recognition accuracy comparable to the more complex state of the art methods can be achieved.

This paper shows how to use beamforming to remove the power-line interference (PLI) in large surface electromyography (sEMG) sensor array or high-density sEMG. The method exploits the highly correlated nature of the different sources of interference, being part of the same electrical grid, and their narrow frequency bands. The idea is to use a very narrow pass-band filter around 50 or 60 Hz to get signals with high PLI content before applying a spatial filtering by principal component analysis (PCA). This way, beamforming are done on the frequency bands where PLI are presents. Also, it ensures that even if the PLI has a smaller overall power than the desired signal, it will be easily found as the most powerful component of the decomposition. The PLI can then be removed from the signal. With trivial modification, harmonics of the PLI can also be removed. The approach was used in the context of muscle behavior analyses of low back pain patients using a sEMG array of 64 sensors. The performances of the filter are studied by experimental and semi-empirical methods. Compared to the usual notch filter, an improvement of up 10 dB is found.