HAL Portal UPHF (Université Polytechnique Hauts-de-France)
Not a member yet
28473 research outputs found
Sort by
Outage Probability and Throughput Analysis of Rate-Splitting Multiple Access for Power Line Communications
No full textInternational audiencePower line communication (PLC) suffers from channel impairments, among which impulsive noise is a major challenge impacting communication performance. In this letter, we incorporate rate splitting multiple access (RSMA) into PLC to ensure robustness, leveraging RSMA’s ability to mitigate channel impairments. We derive analytical expressions for the outage probability (OP) and throughput. We optimize the power and target rate allocation to minimize the system OP. The proposed algorithm significantly reduces the OP. An analysis of the OP-throughput trade-off as a function of target rate shows that RSMA-PLC achieves higher throughput while maintaining lower OP
Influence of electric properties on thermal boundary conductance at metal/semiconductor interface
No full textInternational audienc
Diffusion-Driven Hybrid Unknown Input Observer for Vehicle Dynamics Estimation
No full textInternational audienceVehicle sideslip angle (or lateral speed) and steering angle are essential variables for autonomous vehicle control and active safety systems. Existing estimation methods often depend on deterministic models or neural networks for mapping, which restricts their ability to capture vehicle dynamics distributions in complex driving conditions. This study introduces a diffusion-driven hybrid estimation framework to achieve real-time estimation of lateral speed and steering angle. The nonlinear vehicle dynamics model is represented as a linear parameter-varying (LPV) model. A score-based diffusion model is embedded into an LPV unknown input observer (UIO) to capture the multi-modal distribution of dynamics modeling uncertainty under different state-space model (SSM) parameters without retraining. The model uncertainty prediction sequences from the diffusion model are asynchronously generated to ensure real-time performance. Furthermore, we construct an end-to-end-based hybrid observer incorporating a fullyconnected neural network as an expert model to identify model uncertainty for diffusion model training. Based on Lyapunov stability theory, the ℓ∞ gain performance can be guaranteed to minimize the impacts of uncertainty approximation errors and improve the estimation quality. Experimental results obtained from a real-world test track demonstrate the consistent effectiveness of the proposed framework across various driving scenarios and different SSM parameters, especially in extreme driving scenarios outside the training data distribution. The implementation code is available at https://github.com/tiancheng007/Diffusion-UIO
Circuit intégré d’intelligence artificielle pour capteurs intelligent ultra basse consommation
No full textIn a context where edge computing is becoming increasingly important, the transition from traditional Von Neumann-type architectures to Compute-In-Memory (CIM) architectures appears to be a promising solution for reducing the energy consumption of intelligent systems deployed in constrained environments.While most research focuses on optimising on-chip inference, more recent work explores the implementation of embedded learning, in particular error and gradient calculation, in order to bring greater flexibility to edge systems. Nevertheless, updating parameters remains a major challenge in terms of energy efficiency, due to the data movements in memory that are intrinsic to this operation.This thesis proposes a digital macro, implemented in 28nm FD-SOI ST Microelectronics, for in-memory computing, based on a bit-serial inference architecture with 8-bit inputs and weights, as well as an in-memory weight update architecture. The latter uses gradients encoded as powers of two, combined with a specially designed SRAM cell to enable local bit inversion.The results show energy efficiency of up to 58.9 TOPS/W for inference and up to 527 TOPS/W for weight updating. This work increases the energy efficiency of weight updating by a factor of 9 compared to the conventional Read-Compute-Write approach, without impacting the energy efficiency of inference.This thesis, which introduces the concept of Training-In-Memory (TIM), opens up prospects for intelligent embedded systems whose AI core can be updated according to the context without requiring the use of large server-type infrastructures.Dans un contexte où l'edge computing prend une importance croissante, le passage des architectures classiques de type Von Neumann vers des architectures de calcul en mémoire (Compute-In-Memory, CIM) apparaît comme une solution prometteuse pour réduire la consommation énergétique des systèmes intelligents déployés dans des environnements contraints.Alors que la majorité des recherches se concentrent sur l'optimisation de l'inférence sur puce, des travaux plus récents explorent l’implémentation de l’apprentissage embarqué, notamment le calcul de l'erreur et des gradients, afin d’apporter davantage de flexibilité aux systèmes en périphérie. Néanmoins, la mise à jour des paramètres reste un défi majeur en matière d'efficacité énergétique, en raison des mouvements de données en mémoire intrinsèques à cette opération.Ce travail de thèse propose une macro numérique, réalisée en 28nmFDSOI ST Microelectronics, pour le calcul en mémoire, reposant sur une architecture d’inférence de type bit-série avec des entrées et des poids sur 8 bits, ainsi qu’une architecture de mise à jour des poids en mémoire. Cette dernière utilise des gradients encodés sous forme de puissances de deux, associés à une cellule SRAM spécialement conçue pour permettre une inversion locale de bits.Les résultats montrent une efficacité énergétique pouvant atteindre 58.9 TOPS/W pour l’inférence et jusqu’à 527 TOPS/W pour la mise à jour des poids. Ce travail permet de multiplier par un facteur 9 l’efficacité énergétique pour la mise à jour de poids par rapport à l’approche classique Lecture-Calcul-Ecriture et ce sans impacter l’efficacité énergétique de l’inférence.Ce travail de thèse, qui introduit le concept de Training-In-Memory (TIM), ouvre des perspectives vers des systèmes embarquées intelligents dont le cœur IA peut se mettre à jour en fonction du contexte et sans nécessiter l’usage de grosses infrastructures de type serveurs
AlGaN channel high electron mobility transistors on Si for power electronics
No full textInternational audienc
Non-linear elastic-plastic behavior of the invert glass lithium phosphorous oxynitride (LiPON)
No full textInternational audienceMicro and nano-scale mechanical behavior of network/ionic glasses is dictated by their composition and the number of constraints per structural unit. From this perspective, the glass LiPON presents the opportunity for enhancement of the microscale ductility due to its rather unconstrained orthophosphate structure. We use instrumented nanoindentation with different tip geometries to investigate the mechanical response of LiPON glass. The results reveal that the elastic modulus of LiPON is not constant and depends on pressure. With the method utilizing spherical nanoindentation and continuous stiffness measurement (CSM) we determine the yield point of LiPON. We propose the Drucker-Prager type of yield criterion for LiPON and estimate its yield stress in compression as 2.4 GPa. There exists another stress threshold, however, around 490 MPa, at which the elastic deformation becomes non-linear, and this can be mistaken for the yield stress
In Situ Multiscale Investigation of Capillary-Force-Induced Cold-Welding of Silver Nanowire Networks
No full textInternational audienceSilver nanowire (AgNW) networks are in the spotlight as flexible transparent electrodes (TEs) thanks to their combination of high optical transmittance, low electrical resistance, and excellent flexibility. As such network is formed from a multitude of interwoven AgNW, a postdeposition treatment is needed to get the best electrical conductivity from each interconnect. Traditionally, thermal annealing above 200 °C enhances AgNW junction contacts but restricts compatibility with heat-sensitive substrates such as polymers and perovskites, common in flexible electronics and solar cells. The present study explores the potential of capillary-force-induced cold-welding, an already reported low-temperature alternative operating at or below 100 °C. Better insight into the morphological and electrical properties of coldwelded AgNWs is essential for their large-scale integration as flexible and stable TEs in devices. In this work, the effective welding of AgNW junctions is directly demonstrated by in situ and nanoscale transport measurements. The junction resistance is drastically reduced while preserving both optical transparency and nanowire morphology at significantly lower temperatures than those used with conventional treatments such as thermal annealing. These results strengthen cold welding as a promising approach for overcoming the challenges of AgNW network integration in flexible electronics
Influence of passivation, doping and geometrical parameters on the avalanche breakdown of GaN SBDs
No full textInternational audienceThe breakdown of GaN-based Schottky barrier diodes associated with impact ionization events initiated by electrons injected by tunneling is physically analyzed by means of a Monte Carlo simulator self-consistently coupled with a two-dimensional solution of the Poisson equation. Simulations of a realistic topology where different geometrical parameters are modified allow to identify their influence on the breakdown voltage. The correct physical modeling of two-dimensional effects is essential for a proper prediction of the breakdown. Epilayer doping and thickness, dielectric used for the passivation and lateral extension of the epilayer are analyzed. As expected, the lower the doping and the thicker the epilayer, the higher the value found for the breakdown voltage, but, interestingly, the results also indicate that the peak electric field present at the edge of the Schottky contact, which may be reduced by means of high-k dielectric passivation and a short lateral extension of the epilayer, plays a key role in the breakdown
A spiking coincidence detector for the ITD and ILD mechanisms of auditory localization
No full textInternational audienceSpike-based neuromorphic technology is spreading into artificial neural network applications, carrying the promise of improved energy efficiency. Indeed, spike neural networks offer the advantage of sparse and energy-efficient information representation. Direct encoding of sensor measurements is generally required for these networks to process information, as is done in human physiology—for example, in the cochlea and hair cells for hearing, or On-Off cells for vision. In auditory localization, mechanisms such as Interaural Level Difference (ILD) and Interaural Time Difference (ITD), also known as Time Difference of Arrival (TDOA), are encoded as spike sequences. By analyzing ILD and ITD, it is then possible to estimate the position of a sound source. This study presents a preprocessing model designed to encode auditory localization mechanisms into spike sequences, with frequency variations corresponding to ILD or ITD values. The output of this encoding can be visualized as spatial iso-contours representing the potential locations of the sound source. This model makes novel use of a coincidence detector topology-typically found in artificial vision-as an innovative and efficient estimator of auditory localization mechanisms
Étude expérimentale des modulations en régime sonique et de la génération d’interfaces temporelles dans un cristal phononique piézoélectrique contrôlable
No full textMétamatériaux électroacoustiques; GTEA - Transducteurs et Électroacoustique: GAPSUS - Acoustique Physique, Sous-Marine et Ultra-SonoreNational audienceLes propriétés de dispersion des ondes élastiques se propageant dans des cristaux phononiques piézoélectriques peuvent être modifiées grâce à de simples changements des conditions électriques imposées. Par exemple, pour un empilement d’anneaux piézoélectriques séparés par de fines électrodes, ces propriétés diffèrent si ces électrodes sont laissées en potentiel flottant ou sont reliées à la masse, avec l’apparition d’une bande interdite de Bragg d’origine électrique, absente dans le premier cas. La fréquence de cette bande interdite étant liée à la distance entre deux conditions de mise à la masse, cet effet permet la réalisation d’un milieu de propagation à modulation spatio-temporelle. En pratique, un sous-ensemble périodique d’électrodes est mis à la masse, et cet ensemble est décalé dans l’espace en fonction du temps (1). Un avantage important de ce type de système réside dans la facilité d’accès à des vitesses de modulation comparables (régime sonique) ou supérieures (régime super-sonique) à la vitesse de propagation des ondes dans le matériau, puisque la modulation est réalisée grâce à des circuits électroniques contrôlant les conditions électriques. Dans cette communication des résultats expérimentaux pour ces régimes sont présentés, discutés et confrontés à des prédictions théoriques. En plus des modulations ininterrompues, le système étudié permet la réalisation d’interfaces temporelles (changements globaux et abrupts de conditions) qui les rattachent à la classe des métamatériaux de Floquet (2). La diffraction de paquets d’ondes par une ou plusieurs interfaces temporelles de types différents est étudiée expérimentalement. En plus des effets généraux liés à la génération d’interfaces temporelles, des effets spécifiques aux systèmes piézoélectriques sont observés et discutés. In particulier, la possibilité d’un piégeage des ondes élastiques sous forme électrique est étudiée. (1) S. Tessier Brothelande et al., Appl. Phys. Lett. 123, 201701, 2023. (2) S. Yin et al., eLight 2, 8, 2022