This article studies a robust expectation maximization (EM) algorithm to solve the problem of hypersphere fitting. This algorithm relies on the introduction of random latent vectors having independent von Mises-Fisher distributions defined on the hypersphere and random latent vectors indicating the presence of potential outliers. This model leads to an inference problem that can be solved with a simple EM algorithm. The performance of the resulting robust hypersphere fitting algorithm is evaluated for circle and sphere fitting with promising results.

Drowsiness detection is still an open issue, especially when detection is based on physiological signals. In this sense, light non-invasive modalities such as electroencephalography (EEG) are usually considered. EEG data provides informations about the physiological brain state, directly linked to the drowsy state. Electrocardigrams (ECG) signals can also be considered to involve informations related to the heart state. In this study, we propose a method for drowsiness detection using joint EEG and ECG data. The proposed method is based on a deep learning architecture involving convolutional neural networks (CNN) and recurrent neural networks (RNN). High efficiency level is obtained with accuracy scores up to 97% on validation set. We also demonstrate that a modification of the proposed architecture by adding autoencoders helps to compensate the performance drop when analysing subjects whose data is not presented during the learning step.

Multifractal analysis is a reference tool for the analysis of data based on local regularity and has proven useful in an increasing number of applications involving univariate data (scalar valued time series or single channel images). Recently the theoretical ground for a multivariate multifractal analysis has been explored, showing its potential for capturing and quantifying transient higher-order dependence beyond correlation among collections of data. Yet, the accurate estimation of the parameters associated with these multivariate multifractal models is challenging. Building on these first formulations of multivariate multifractal analysis, the present work proposes a Bayesian model and studies an estimation framework for the parameters of a quadratic model for the joint multifractal spectrum of bivariate time series. The approach relies on a novel joint Gaussian model for the logarithm of wavelet leaders and leverages on a Whittle approximation and data augmentation for the matrix-valued parameters of interest. Monte Carlo simulations demonstrate the benefits of the method with respect to previous formulations. In particular, we obtain significant performance improvements at only moderately larger computational cost, for large ranges of sample size and multifractal parameter values.

Monitoring agriculture from satellite remote sensing data, such as multispectral images, has become a powerful tool since it has demonstrated a great potential for providing timely and accurate knowledge of crops. Detecting anomalies in time series of multispectral remote sensing images for crop monitoring is generally performed using a large sample of historical data at a pixel level. Conversely, this article presents a framework for anomaly detection (AD), localization, and classification that exploits the temporal information contained in a given season at a parcel level to detect and localize outliers using hidden Markov models (HMMs). Specifically, the AD part is based on the learning of HMM parameters associated with unlabeled normal data that are used in a second step to detect abnormal crop parcels referred to as anomalies. The learned HMM can also be used in time segments to temporally localize the anomalies affecting the crop parcels. The detected and localized anomalies are finally classified using a supervised classifier, e.g., based on support vector machines. The proposed framework is applicable to images partially covered by clouds and can handle a set of crop parcels acquired in the same season bypassing problems due to crop rotations. Numerical experiments are conducted on synthetic and real data, where the real data correspond to vegetation indices extracted from several multitemporal Sentinel-2 images of rapeseed crops. The proposed approach is compared to standard AD methods yielding better detection rates with the advantage of allowing anomalies to be localized and characterized.

The first part of this paper presents some works conducted with Jose Bioucas Dias for fusing high spectral resolution images (such as hyperspectral images) and high spatial resolution images (such as panchromatic or multispectral images) in order to build images with improved spectral and spatial resolutions. These works are related to Bayesian fusion strategies exploiting prior information about the target image to be recovered constructed by dictionary learning. Interestingly, these Bayesian image fusion methods can be adapted with limited changes to motion estimation in pairs or sequences of images. The second part of this paper explains how the work of Jose Bioucas Dias has been a source of inspiration for developing new Bayesian motion estimation methods for ultrasound images.

Machine health condition monitoring is evidently a crucial challenge nowadays. Unscheduled breakdowns increase operating costs due to repairs and production losses. Scheduled maintenance implies taking the risk of replacing fully operational components. Human expertise is a solution for an outstanding expertise but at a high cost and for a limited quantity of data only, the analysis being time-consuming. Industry 4.0 and digital factory offer many alternatives to human monitoring. Time, cost and skills are the real stakes. The key point is how to automate each part of the process knowing that each one is valuable. Leaving aside scheduled maintenance, this paper copes with condition-based preventive maintenance and focuses on one fundamental step : the signal processing. After a brief overview of this specific area in which numerous technologies already exist, this paper argues for an automated signal processing at an expert level. The objective is to monitor a system over days, weeks, or years with as great accuracy as a human expert, and even better in regard to data investigation and analysis efficiency. After a data validation step most often ignored, any multimodal signal (vibration, current, acoustic, ...) is processed over its entire frequency band in view of identifying all harmonic families and their sidebands. Sophisticated processing such as filtering and demodulation creates relevant features describing the fine complex structures of each spectrum. A time-frequency feature tracking constructs trends over time to not only detect a failure but also to characterize and localize it. Such an automated expert-level processing is a way to raise alarms with a reduced false alarm probability.

In recent years, our society is preparing for a paradigm shift toward the hyper-connectivity of urban areas. This highly anticipated rise of connected smart city centers is led by the development of low-cost connected smartphone devices owned by each one of us. In this context, the demand for low-cost, high-precision localization solutions is driven by the development of novel autonomous systems. The creation of a collaborative based network will take advantage of the large number of connected devices in today's city center. This paper validates the positioning performance increase of Android low-cost smartphones device present in a collaborative network. The assessment will be made on both simulated and collected smartphone's GNSS raw data measurements. We propose a collaborative method based on the estimation of distances between network mobile users used in a SMARTphone COOPerative Positioning algorithm (SmartCoop) . Previous analysis made on smartphone data allow us to generate simulated data for experimenting our cooperative engine in nominal conditions. The evaluation and analysis of this innovative method shows a significant increase of accuracy and reliability of smartphones positioning capabilities. Position accuracy improves by more than 3m, in average, for all smartphones within the collaborative network.

Parameter estimation is a problem of interest when designing new remote sensing instruments, and the corresponding lower performance bounds are a key tool to assess the performance of new estimators. In global navigation satellite systems reflectometry (GNSS-R), a noncoherent averaging is applied to reduce speckle and thermal noise, and subsequently the parameters of interest are estimated from the resulting waveform. This approach has been long regarded as suboptimal with respect to the optimal coherent one, which is true in terms of detection capabilities, but no analysis exists on the corresponding parameter estimation performance exploiting GNSS signals. First, we show that for certain signal models, both coherent and noncoherent Cramér-Rao bounds are equivalent, and therefore, any maximum likelihood estimation coherent/noncoherent combination scheme is efficient (optimal) at high signal-to-noise ratios. This is validated for an illustrative GNSS-R estimation problem. In addition, it is shown that considering the joint delay/Doppler/phase estimation problem, the noncoherent performance for the delay is still optimal, which is of practical importance for instance in altimetry applications.

We investigate on binary Protograph Root-LDPC codes that can embed an interesting property, called nested property. This property refers to the ability for a coding scheme to achieve full diversity and equal coding gain for any number of received coded blocks and for any configuration of the received code blocks. One can take advantage of this property for “carousel”-type transmissions broadcasting cyclically coded information. For regular Root-LDPC codes, we show that these codes inherently have both properties over the nonergodic block fading channel and when message passing decoding is used. Then, we show that irregular Root-LDPC structures could not provide equal coding gain except if explicit design rules are enforced to ensure that the nested property is fulfilled when designing irregular Root-LDPC codes. Simulation results show that designed nested Root-LDPC codes achieve good performance and full diversity for coding rates R=1/2, R=1/3 and R=1/4.

This paper presents some research works based on the PhD thesis of B. Pilastre (B. Pilastre, Estimation Parcimonieuse et Apprentissage de Dictionnaires pour la détection d’Anomalies Multivariées dans des Données Mixtes de Télémesure Satellite, PhD Thesis of the university of Toulouse, Nov. 6, 2020.), supported by CNES and Airbus Defence & Space, on a new Anomaly Detection algorithm based on a sparse decomposition into a DICTionary (ADDICT). The proposed method addresses two main challenges related to anomaly detection for satellite telemetry parameters, namely the multivariate processing of these parameters and the mixed continuous and discrete nature of the data. Different variations of the ADDICT algorithm, referred to as C-ADDICT and W-ADDICT, have been investigated differing by the data decomposition term defined using a linear combination of the atoms or its convolutional equivalent. The resulting ADDICT, C-ADDICT and W-ADDICT algorithms have been evaluated on a small representative dataset containing satellite anomalies with an available ground-truth and have shown competitive results with respect to the state-of-the-art. They have also been tested on industrial use-cases, especially regarding online processing (i.e., sequential learning taking into account the feedback of users). The results of these tests are presented in this paper.