Добірка наукової літератури з теми "Physics-based invertible models"

Particle accelerators are enabling tools for scientific exploration and discovery in various disciplines. However, finding optimised operation points for these complex machines is a challenging task due to the large number of parameters involved and the underlying non-linear dynamics. Here, we introduce two families of data-driven surrogate models, based on deep and invertible neural networks, that can replace the expensive physics computer models. These models are employed in multi-objective optimisations to find Pareto optimal operation points for two fundamentally different types of particle accelerators. Our approach reduces the time-to-solution for a multi-objective accelerator optimisation up to a factor of 640 and the computational cost up to 98%. The framework established here should pave the way for future online and real-time multi-objective optimisation of particle accelerators.

Abstract Information hiding technology is a technology that transmits secret information through a carrier, and is a research hotspot in the field of information security. The traditional information hiding algorithm relies on the meticulous design of human beings, and obtains the dense image through modification. With the development of deep learning, the integration of information hiding technology and deep learning has resulted in many information hiding technologies based on deep learning. Among them, image hiding has become a research hotspot due to its large steganographic capacity. Therefore, this paper reviews the information hiding technology based on deep learning. According to the difference of hidden models, it is analysed from four aspects: (1) information hiding model based on encoder-decoder; (2) information hiding model based on generative adversarial network; (3) information hiding model based on invertible network; (4) information hiding model based on neural network information hiding models for style transfer. Finally, these models are analysed and compared, and the future development direction is discussed and prospected.

The concept presented in this paper is based on previous dynamical methods to realize a time-varying matrix inversion. It is essentially a set of coupled ordinary differential equations (ODEs) which does indeed constitute a recurrent neural network (RNN) model. The coupled ODEs constitute a universal modeling framework for realizing a matrix inversion provided the matrix is invertible. The proposed model does converge to the inverted matrix if the matrix is invertible, otherwise it converges to an approximated inverse. Although various methods exist to solve a matrix inversion in various areas of science and engineering, most of them do assume that either the time-varying matrix inversion is free of noise or they involve a denoising module before starting the matrix inversion computation. However, in the practice, the noise presence issue is a very serious problem. Also, the denoising process is computationally expensive and can lead to a violation of the real-time property of the system. Hence, the search for a new ‘matrix inversion’ solving method inherently integrating noise-cancelling is highly demanded. In this paper, a new combined/extended method for time-varying matrix inversion is proposed and investigated. The proposed method is extending both the gradient neural network (GNN) and the Zhang neural network (ZNN) concepts. Our new model has proven that it has exponential stability according to Lyapunov theory. Furthermore, when compared to the other previous related methods (namely GNN, ZNN, Chen neural network, and integration-enhanced Zhang neural network or IEZNN) it has a much better theoretical convergence speed. To finish, all named models (the new one versus the old ones) are compared through practical examples and both their respective convergence and error rates are measured. It is shown/observed that the novel/proposed method has a better practical convergence rate when compared to the other models. Regarding the amount of noise, it is proven that there is a very good approximation of the matrix inverse even in the presence of noise.

Abstract 
It is found during the tests that the damping characteristics of the magnetorheological (MR) damper vary with the excitation amplitude and frequency. However, the existing MR damper models are not able to accommodate the change of excitation amplitude and frequency, which will lead to significant modeling errors. To deal with this problem, this paper analyzes the experimental data and obtains the regularity of the damping characteristics varying with the excitation. Subsequently, an excitation-adaptive MR damper model is constructed based on the hyperbolic tangent function. The proposed model is not only able to adapt to the change of excitation amplitude and frequency but also able to inverse, which is essential for MR damper controller construction. The fitting results show that compared with the existing models, the three normalized errors of the proposed model are improved from 22.61%, 13.96%, and 19.42% to 6.30%, 3.81%, and 6.97%, respectively, indicating that the model excitation adaptivity is significantly improved. Furthermore, this study also proposed a damper controller based on the new model, and the simulation results verify the effectiveness of the controller. The proposed model brings the acceleration signal into the model to improve the model adaptivity, which introduces a novel approach to enhance the adaptivity of MR damper models.

L’imagerie conventionnelle, qui se limite aux formes et couleurs des objets, montre ses limites en matière de reconnaissance. Pour améliorer les performances des systèmes d’imagerie, l’imagerie hyperspectrale et polarimétrique apporte une richesse d’informations, notamment des grandeurs physiques difficiles à obtenir autrement. Cela permet d’améliorer la détection, la caractérisation quantitative et la classification des objets. Cependant, le traitement des données complexes de ces modalités reste un défi. L’objectif de ce travail est de proposer une méthodologie générique pour analyser les signaux optiques, en se concentrant sur l’imagerie hyperspectrale (HSI) en premier terme. Une classification originale des modèles hyperspectraux inversibles basés sur la physique est présentée, avec description des modèles variés les plus récents pour des applications diverses : MPBOM pour le biofilm d’algues et de bactéries, MARMIT pour le sol, PROSPECT pour les feuilles de plantes, Farrell pour les tissus biologiques turbides, Schmitt pour la peau humaine et Hapke pour les objets du système solaire. Une convergence entre les modèles PROSPECT et Farrell pour des objets intermédiaires (pomme verte et poireau) ouvrant la voie au développement d’une nouvelle modélisation générique et complète. Notamment dans le domaine de la biologie, par une collaboration avec le laboratoire de l’ANSES, nous avons procédé à une détection précoce suivie d’une quantification du biofilm qui se forme dans les bassins d’élevage de poissons en utilisant l’imagerie hyperspectrale et polarimétrique du fait que sa détection actuelle est visuelle et n’est pas assez efficace pour prévenir son accumulation et pour mettre en place des procédures de nettoyage et de désinfection. Ainsi une première version d’une modélisation physique propre nommée "DNA-HSI" a été mise en place
Conventional imaging, limited to object shapes and colors, faces limitations in object recognition. To enhance imaging system performance, hyperspectral and polarimetric imaging provides a wealth of information, includingchallenging-to-obtain physical parameters. This facilitates improved object detection, quantitative characterization, and classification. However, the processing of complex data from these modalities remains a challenge. The aim of this work is to propose a generic methodology for the analysis of optical signals, with a primary focus on hyperspectral imaging (HSI). An original classification of invertible physics-based hyperspectral models is presented, along with descriptions of recent diverse models for various applications: MPBOM for algae and bacteria biofilm, MARMIT for soil, PROSPECT for plant leaves, Farrell for turbid biological tissues, Schmitt for human skin, and Hapke for objects in the solar system. A convergence between the PROSPECT and Farrell models for intermediate objects (green apple and leek) paves the way for the development of a new generic and comprehensive modeling approach.Particularly in the field of biology, in collaboration with the ANSES laboratory, we conducted early detection ollowed by quantification of biofilms forming in fish farming basins using hyperspectral and polarimetric imaging. This is crucial as the current visual detection method is not efficient in preventing biofilm accumulation and implementingcleaning and disinfection procedures. Hence, an initial version of a dedicated physical modeling approach called "DNA-HSI" has been established

Structural systems are inevitably subject to degradation that evolves progressively over time. Developing a degradation model to capture the physics of damage evolution is essential for failure prognostics, i.e., remaining useful life (RUL) prediction, to enable individualized predictive maintenance. Due to the lack of runto- failure data for large structural systems and natural variability across physical systems, uncertainty is inherent in the degradation model even if a degradation model can be constructed based on the physics of a certain damage mechanism. It is therefore necessary to update the degradation model over time based on measurements of quantities that are directly measurable. With the development of sensing and image processing techniques, it is possible to derive structural strain response from videos, which overcomes the limitations of the cumbersome and costly deployment of conventional contact sensors. While the strain video monitoring data provide rich information for structural health monitoring, the usage of this information for degradation model updating is challenging due to the implicit connection between the degradation model parameters and strain video monitoring data and the highly complicated model architectures. This research proposes a novel sequential Bayesian model updating framework for a degradation model using a likelihood-free Bayesian inference method and strain video monitoring data. In the proposed framework, strain video monitoring data are first compressed into lowdimensional latent time-series features using a convolutional autoencoder. Subsequently, a likelihood-free Bayesian inference method is employed to update the degradation model using a given time duration of the monitoring data. To enable continuous monitoring and model updating over a long time period, a sequential Bayesian model updating scheme is developed. Based on the updated degradation model, failure prognostics are performed sequentially and the associated uncertainty on RUL estimation is also quantified. The application of the developed framework to a miter gate structure demonstrates the efficacy of the proposed framework. Keywords: Remaining useful life; Degradation model; Likelihood-free Bayesian inference; Conditional invertible neural network