The SIGNAL area brings together two main fields of expertise:
- Evolutionary spectral analysis and modeling
- Multi-sensor processing.
Applications vary:
- Analysis and understanding of bioelectrical signals (electrocardiogram ECG, electromyogram EMG and phonocardiogram PCG)
- Radio astronomy
- Signal processing for telecommunications
- Vibration analysis of rotating machines
- Power line studies for power line carrier transmissions.
Research themes:
The main objective is to extract components: indicators or parameters.
- Spectral analysis and modeling
- Time-frequency and time-scale processing
- Cyclostationary analysis
- Multi-sensor processing
- Parametric estimation
- Source separation techniques
- Antenna processing and spatial filtering
- Instrumentation & Measurement
Seminars on Thursdays at 1pm:
Seminar 1: Thursday, February 3
Machine Learning Strategies in Time Series Forecasting
by Zuokun OUYANG (doctoral student CIFRE ATTILLA)
Abstract: With the increasing availability of large amount of historical data and the requirement of precise prediction for future values, robust and efficient methods are demanded to capture the dependency of historical observations and the future values. The time series forecasting (TSF) domain has been dominated for decades by statistical methods such as ARIMA models. Over the last decade, machine learning methods have drawn much prominence for their successful applications in many fields like computer vision and natural language processing. This seminar presents an overview of machine learning techniques for time series forecasting with a focus on 1) transforming a TSF into a supervised learning problem; 2) several machine learning techniques for TSF; 3) strategies for multi-step forecasting.
Seminar 2: Thursday, March 3
Estimation of ground reaction forces from measurements of instrumented footings
by Amal Kammoun (doctoral student CIFRE EMKA)
Abstract: In the world of work or sport, the mechanical constraints to which a worker or practitioner is subjected can lead to functional motor disorders or even immediate bodily injury. Their assessment is therefore a major health and safety issue for the individual. In these environments, assessments are carried out by prevention specialists or health professionals (ergonomists, physiotherapists, doctors), and are still based on visual assessments, in particular rating grids for ergonomists. These do not take into account the biomechanical realities that are not directly accessible to an outside observer, and are therefore underestimated. In order to support the individual in his or her activity, it is necessary to develop appropriate measurement and modeling tools that consider these factors. The present research project focuses on estimating reaction forces to foot-ground contact, according to the three components of the human body frame: anterior-posterior (Fx), medial-lateral (Fy) and vertical (Fz). In order to reconstruct this three-dimensional vector (Fx, Fy, Fz), the processing methods used will fall within the fields of data analysis and artificial intelligence, including dimensionality reduction techniques and supervised and/or deep learning methods.
Seminar 3: Thursday, March 17
Detection of wave front: recent advances and new challenges
by Benjamin Poste (doctoral students CIFRE CEA)
Abstract: Recently, a new algorithm for broadband source detection and estimation based on likelihood has been implemented in order to overcome the errors of the current algorithm: PMCC (based on TDOA, for time difference of arrivals). This study presents the two approaches and compares them in terms of detection and estimation capacity on synthetic signals but also on real signals from the IMS network.
In a second time, the multi-source aspect is also studied through this new algorithm. It is compared with state of the art algorithms such as MUSIC, Fisher or CLEAN. The results are very encouraging in the generalization of the algorithm based on the likelihood for the detection of multiple sources.
Seminar 4: Thursday, April 7
Application of the EEMD-PCA-FastICA-based single-channel blind separation algorithm in electric charge identification
by Gisèle Béatrice Sonfack (post doc)
Abstract: The deployment of smart meters in the residential sector is providing a large amount of data enabling a better understanding of occupants' electricity consumption habits in order to provide energy optimization solutions. However, conventional methods for analyzing domestic electrical load profiles based on time series techniques, such as Fourier transform and wavelet transform, have shown limitations due to non-linearities and non-stationary processes. In this paper, we present an approach to electrical load profile identification based on the detection of oscillation modes in the time-frequency domain, combined with conventional signal processing algorithms. The methodology is based on the Hilbert-Huang transform (HHT), which enables instantaneous frequency to be evaluated using Empirical Mode Decomposition (EMD) and, more precisely, its CEEMDAN variant. Thanks to the CEEMDAN decomposition of the single-channel observation signal of the electric current, the multidimensional components of the IMF are obtained, and principal component analysis (PCA) is performed on the matrix of these IM[5]F. The number of principal components is determined and a new matrix is generated to satisfy the overdetermined blind source separation conditions, the new matrix is then applied as input to the FastICA to perform source signal separation. The study was carried out on two loads from the COOLL database, and the mixtures were made artificially. The advantage of this approach lies in the fact that it bypasses the constraint of having a number of observations greater than or equal to the number of sources. It therefore applies very well to the NILM context, where a single observation is very often available. Furthermore, this method reveals the importance of data sampling frequency for accurate load profile characterization.
Seminar 5: Thursday, April 21
SATSal: A Multi-Level Self-Attention Based Architecture for Visual Saliency Prediction
by Merouane Tliba (doctoral student)
Abstract: Human visual Attention modeling is a persistent interdisciplinary research challenge, gaining new interest in recent years mainly due to the latest developments in deep learning. That is particularly evident in saliency benchmarks. Novel deep learning-based visual saliency models show promising results in capturing high-level (top-down) human visual attention processes. Therefore, they strongly differ from the earlier approaches, mainly characterized by low-level (bottom-up) visual features. These developments account for innate human selectivity mechanisms that are reliant on both high- and low-level factors. Moreover, the two factors interact with each other. Motivated by the importance of these interactions, in this project, we tackle visual saliency modeling holistically, examining if we could consider both high- and low-level features that govern human attention. Specifically, we propose a novel method SAtSal (Self-Attention Saliency). SAtSal leverages both high and low-level features using a multilevel merging of skip connections during the decoding stage. Consequently, we incorporate convolutional self-attention modules on skip connection from the encoder to the decoder network to properly integrate the valuable signals from multilevel spatial features. Thus, the self-attention modules learn to filter out the latent representation of the salient regions from the other irrelevant information in an embedded and joint manner with the main encoder-decoder model backbone. Finally, we evaluate SAtSal against various existing solutions to validate our approach, using the well-known standard saliency benchmark MIT300. To further examine SAtSal’s robustness on other image types, we also evaluate it on the Le-Meur saliency painting benchmark.
Seminar 6: Wednesday, May 11
Deep learning for injunction classification in French oral utterances
by Asma Bougrine (ATER at Polytech)
Abstract: The classification of injunctions in French oral discourse is a difficult task, since no standard linguistic structure has yet been devised for their characterization. An injunction can be an order, an advice or a prohibition, and an initial study showed that prosodic features of speech could be relevant indicators for this task, notably logarithmic energy. Our aim is first to validate, on an initially reduced and then extended basis, the predominance of this feature using other conventional classifiers such as SVM and K-NN. When applied to our first database, the logarithmic energy feature showed the best classification rates (CR) for both classifiers, with $CR=82\%$ for SVM and CR=71.42% for K-NN. On the other hand, energy was not the relevant feature when applied to our second, wildly heterogeneous base. Secondly, in order to improve classification rates, we applied two deep learning networks, LSTM and CNN. When applying the LSTM network, a CR=96.15% was found using the 6 prosodic features when applied to the first database versus a CR=64% with the second database data. CNN, a network capable of automatically extracting the most relevant features, performed better on the second database, with a CR=73%, well ahead of SVM, K-NN and LSTM.
Seminar 7: Thursday, June 2
Assessing the impact of power grid quality on device signatures
by Gisèle Béatrice Sonfack (post doc)
Abstract: The active participation of the residential customer in the smart grid via an energy manager can benefit from the control of a certain amount of data relating both to the behavior of the network and to the appliances connected to it. In fact, knowing the type of appliances in operation and their consumption is essential for forecasting overall consumption in the home, managing operating modes and optimizing energy expenditure. Although several load identification techniques are being developed in scientific laboratories, it is important to note that studies on the impact of power grid quality on the signatures underlying these developments have yet to be elucidated. Given the dependence of load profiles on environmental factors, user habits and other non-deterministic factors, it is important to assess the impact that network quality can have on load signatures, in order to take this into account in the development of identification solutions to be integrated into the smart grid. In this context, the study carried out in this article presents an assessment of the various network faults likely to impact the signatures of connected loads. The study considers the two types of loads encountered in households (linear and non-linear) and analyzes characteristic quantities such as waveforms, effective current and voltage values, and distorting active and reactive power (harmonics). Simulation results from Simulink/Simscape show that some network faults have a significant and permanent impact on signatures, while others have a temporary and sometimes insignificant effect.
Seminar 8: Friday, June 24
Development of Orthopaedical Insoles and Shoes 3D Design Process for Atypical Foot Morphologies
by Maneesh Kumar Mishra (ATER at Polytech)
Abstract: The needs in the sector of orthopaedic insoles are becoming more important. The medical purpose of these plantar prostheses is often to correct the posture or biomechanical imbalance that tends to cause pain in various areas of the body depending on the pathology of the patient. The need for personalization of orthopaedic shoes is also increasing. This is due to the progression of people who are increasingly obese or have problems with diabetes or other diseases. However, the cost of its products can vary from one to two times or more depending on the quality of the product that requires adaptation by a manual process to the morphology of the foot. Whatever the therapeutic product envisaged, the economic model is no longer adapted to the needs of patients and the budgetary imperatives of social security. A new creative process is needed and must incorporate digital tools to reduce these manufacturing costs, while improving the quality of products. Our study was to take stock of the means of measurement used and analyse the profession of podiatrist by highlighting the medical and economic aspects to detect the technological barrier. The implementation of the new design process for these therapeutic foot products required the acquisition of morphological, anthropometric and biomechanical knowledge of the foot. In order to characterize the foot dimensionally and morphologically, a process of detection of anthropometric points and creation of morphological curves has been implemented. At this stage, the process was able to analyse a population of male and female individuals to extract the most representative morphologies of a target population. It was from unsupervised classification methods that has classified the population and detected centroids. The techniques that were then used to detect the anthropometric points of the foot were incorporated into the process of creating customized insoles. By combining this technique that extracts the outline encompassing the footprint, a method that uses this outline to create the 2D shape of the insole and the 3D shape of the standard shoe-last, the 3D shape of the insole was created and adjusted to foot and the standard shoe-last. This process showed the need to create the custom shoe-last of the person. After a detailed technical analysis of the shoe-last, set up of a customized shoe-last 3D creation process has been stablished. This process was directly able to take into account the shape of the patient's foot on which detection of the anthropometric points and creation of morphological contours were done. These contours were to support a 3D space of ease on which was hung a network of curves defining the wire model of the surface of the shoe-last. This spatial ease was parametrizable in order to define a specific comfort to the patients or to control their therapeutic needs. These two customization processes can be connected to each other to improve the wearing comfort of shoes for a patient or for the entire population by applying it to each centroid of the classification.
Seminar 9: Thursday, October 6
Signal and image data stream analytics: from subspace to tensor tracking
by Thanh Trung Le (thesis pre-defense)
Abstract: Stream processing has recently attracted much attention from both academia and industry since massive data streams have been increasingly collected over the years. This thesis focuses on investigating the problem of online low-rank approximation (LRA) of data streams over time. When data samples are one-dimensional, the online LRA problem is referred to as subspace tracking. It turns out to be tensor tracking when streaming data are multi-dimensional. For subspace tracking, we proposed two novel algorithms for tracking the underlying subspace of data streams under two specific scenarios. To deal with sparse outliers and missing data, an effective two-stage subspace tracking algorithm was developed, namely PETRELS-ADMM. The proposed algorithm is based on the alternating direction method of multipliers and recursive least-squares filtering techniques. The second algorithm called OPIT was specifically designed for tracking the sparse principal subspace in high dimensions. Specifically, OPIT introduces a new adaptive variant of power iteration and a new column-based thresholding operator. Both two proposed algorithms belong to the class of provable tracking methods with a convergence guarantee. For tensor tracking, we developed several new algorithms for tracking the online LRA of streaming tensors over time. Under the CP/PARAFAC format, we leverage the alternative minimization and randomized sketching techniques to develop ACP and RACP which are capable of factorizing incomplete tensors and corrupted tensors, respectively. Under the Tucker format, we proposed another online algorithm called ATD. ATD first tracks the underlying low-dimensional subspaces covering the tensor factors, and then estimates the core tensor using {a} stochastic approximation. A unified convergence analysis was presented to justify their performance. In parallel, it is well-known that the tensor-train format offers a memory-saving representation and can break the curse of dimensionality. Accordingly, we designed three new algorithms for streaming tensor-train decomposition. The first algorithm called TT-FOA is capable of tracking the low-rank components of high-order tensors from noisy and high-dimensional data with high accuracy, even when they come from time-dependent observations. The second algorithm called ATT is specifically designed for handling incomplete streaming tensors. Finally, to deal with sparse outliers, we proposed the so-called ROBOT. Technically, ROBOT has the ability to track streaming tensors from imperfect streams (i.e., due to noise, outliers, and missing data) as well as estimate their time variation in dynamic environments. In conclusion, our study provides several novel contributions to big data stream analytics in general and the online LRA problem in particular. They are new analysis tools allowing to effectively track the online LRA of data streams from one-dimensional to multi-dimensional observations in different settings, and thus, they are expected to take a step forward real-world online applications.
Seminar 10: Thursday, October 21
Design and evaluation of an on-board system for assessing mechanical stress during handling and load-carrying activities.
by Nicolas Grandin (doctoral student CIFRE TEA)
Abstract: MSDs (musculoskeletal disorders) are a major problem for French society, as they represent the main category of occupational illness in France, according to INRS. In fact, MSDs account for some 10 million lost days a year, or the equivalent of €1 billion in losses for French society and companies alike. To date, there is no device for automatically assessing handling and load-carrying activities from a mechanical point of view (mass of loads handled and stresses on the body) under industrial conditions (portable, wireless system, etc.). For this reason, the aim of this thesis is to design an on-board system for determining the mechanical stresses experienced during handling and load-carrying activities. The first step is to determine the mass of the load carried by the operator. Then, using assumptions about anthropometric data (sex, age, morphology, etc.) and biomechanical data (mechanical properties of tissues, reactions to ground forces, etc.), it is then possible to determine the mechanical stresses experienced during these tasks, and thus quantify their deleterious effects.
Seminar 11: Thursday, November 24
Mobility aid using localization via Local Magnetic Field Imprint
by Ouyang Guanglie (thesis pre-defense)
Abstract: Due to the ubiquity of the geomagnetic field and the availability of its measurement without the need for expensive infrastructure, magnetic field-based indoor positioning has attracted considerable interest from academics and industry alike. However, indoor positioning applications based on the magnetic field are difficult due to its poor spatial discrimination, the heterogeneity of measurement devices and the interference of ferromagnetic materials. This thesis contributes to demonstrating the feasibility of magnetic field-based indoor positioning by addressing the challenges of magnetic field positioning. Our contributions are as follows: the first contribution is a survey of the state of the art of magnetic field-based indoor positioning, including advantages and challenges, magnetometer models and calibration, coordinate transformation and magnetic field localization methods. The second contribution is an analysis of the challenges and feasibility of magnetic field localization through static, trajectory and rotation tests, calibration tests and classification tests. The third contribution presents a robust method based on ADMM (Alternating Direction Method of Multipliers) for calibrating three-axis magnetometers in the presence of outliers and spurious measurements. The fourth contribution designed a magnetic trajectory classification algorithm based on temporal convolutional networks. This study of the challenges and potential solutions relating to the use of magnetic field measurements for mobile localization makes it a promising step forward for real-world applications.
Seminar: Thursday, December 1st
Prosodic focus in oral corpora
by Fadila El Adlouni (intern Master RTR Diams)
Abstract: Classification of speech audio corpora from the ravioli project into injunctive and non-injunctive classes, where injunction can be defined as a precise, non-discussible order or command that must be compulsorily carried out and is often accompanied by threats and sanctions.
The bidirectional LSTM is used for the classification task, where it is given as input a vector containing the six prosodic features: pitch PI, energy E, speed D and acceleration A (A and D are associated with PI and E) extracted for each 10 ms of the audio, this vector is supposed to represent the original vector of the audio.