Optical Memory and Neural Networks最新文献

Modern combined heat and power (CHP) plants are one of the main sources of energy and thermal resources for the private and industrial sectors. They are also a source of a large amount of data, which is sufficient for the application of machine learning methods. For accurate optimization of power equipment, it is necessary to predict the generated power to ensure sufficient electrical and thermal energy. This study develops an LSTM-based forecasting model. The model is trained on a time series collected from a combined heat and power (CHP) plant over 15 months of operation. Compared with a gradient-boosting model, the LSTM achieves higher accuracy across regression metrics. The developed model can serve as a component of decision-support systems and for subsequent optimization of equipment operation at power stations, thereby improving overall plant efficiency.

We investigate the optical properties of guided-mode resonant gratings with the thickness of the waveguide layer spatially varying according to a parabolic law. We show that an efficient approach for describing such structures is the coupled-mode theory (CMT) with varying parameters. The CMT simulations are quite fast and its predictions turn out to be in a good agreement with time-consuming rigorous simulations based on the aperiodic Fourier modal method. We demonstrate that the optical properties of gratings with convex and concave waveguide layers are quite different, which is explained in terms of the local photonic band gap of the considered structures. In our opinion, the obtained results are promising for the design of optical filters and resonators based on quasiperiodic photonic structures with nonlinearly varying parameters.

The comparative efficiency of wave aberration recognition using a convolutional neural network (CNN) with the VGG architecture based on diffraction patterns on linear and binary axicons is investigated. A dataset of 90 000 images was calculated for training the CNN, in which diffraction patterns on two types of axicons in the focal plane were simulated for each of the selected 8 types of aberrations. Based on CNN training in 40 epochs, it was shown that the achieved average absolute error in recognizing aberrations using a binary axicon does not exceed 0.0138, which is significantly less than for a linear axicon (0.193).

This paper presents a theoretical and experimental investigation of the influence of sector perturbations on the polarization structure of azimuthal and radial modes. To quantify the degree of polarization restoration, an integral metric based on normalized Stokes parameters is introduced. Numerical simulations demonstrate that for small sector angles, the polarization structure recovers at propagation distances exceeding the Rayleigh length. As the sector angle increases, only partial restoration is observed. Although the radial mode exhibits locally more stable polarization behavior, the integral restoration metric shows only minor differences compared to the azimuthal mode. The results are supported by experimental data and are in good agreement with the theoretical model.

We propose an approach for obtaining analytical expressions approximating asymmetric resonances in the spectra of multimode layered structures, where elements of the structure support excitation of plasmonic modes. The 2 × 2 transfer matrix method is used to form a composite matrix describing the field amplitudes in such a structure element. The coefficients of the composite matrices are represented by analytical expressions describing the interference of the resonance and non-resonance components. The applicability of the proposed approach is demonstrated in the case of coupling of two plasmonic modes in a multilayer metal-dielectric structure.

Besides scalar optical vortices that have a topological charge (TC), helical wave front, and carry an orbital angular momentum (OAM) that can be transferred to particles and rotate them along circular trajectories, polarization optical vortices are also known, whose polarization state in the beam section changes with the azimuthal angle. Such vortices are polarization singularities that are described by indices, similar to the TC. However, polarization OAM for polarization vortices still has not been considered, although laser beams with inhomogeneous polarization can perform spiral mass transport in polarization-sensitive media. In this work, we introduce and consider two analogues of the OAM for vector fields. One polarization OAM analogue is the polarization orientation azimuthally changing velocity, or polarization OAM (pOAM), whereas the hybrid OAM analogue is the polarization ellipticity azimuthally changing velocity, or hybrid OAM (hOAM). For instance, we demonstrate that the normalized pOAM is equal to the order of a cylindrical vector beam and also equals the order of Poincaré beam. In optical material processing, the considered OAM analogues describe how light fields with different polarization affect polarization-sensitive media. In optical data transmission, these OAM analogues allow distinguishing different light fields that are indistinguishable by their conventional OAM.

The paper investigates the formation of polarization singularities arising from the non-axial interference of vector optical modes of different symmetry. An analytical model of superposition of two spatially shifted beams with arbitrary topological charges and phase parameters is proposed. It is shown that in the region of their overlap polarization features of the V- and C-points type as well as stable configurations of the lemon and star types can arise. On the basis of the Stokes parameters, the features are classified and the conditions of their formation are determined. An analytical relation determining the optimal value of beam displacement for the appearance of a symmetric singularity is proposed. The results are confirmed by numerical modelling and experiment.

When considering optical beams propagating in media with optical activity, using the example of a structurally stable Bessel-Gauss launch, it is shown that in media without active birefringence, it is possible to create a single optical vortex in one of the polarization components of the beam. Mutual transformation between the components is due to the optical activity of the medium for both azimuthal and radial field components. The stable state of the isotropic point is due to the absence of external disturbances of the beam, and in the presence of external disturbances by the active medium, the isotropic point is destroyed. Linear activity and the presence of birefringence are related to each other and affect the isotropic point. The structure of fields in a medium with optical activity is very similar to the structure of Bessel-Gauss beams and can be controlled by them using external disturbances.

This paper explores the application of computer vision for quality control of products, focusing on the challenge of surface defect detection with a limited dataset. Substrate made of AlSiC composite material is a good example. To address the small sample size, data augmentation and transfer learning techniques have been employed, pre-training a model on a public crack dataset. The core of approach is utilization of the YOLOv8-OBB object detector, chosen for its support of oriented bounding boxes, which are crucial for accurately capturing elongated defects like cracks. Furthermore, to enhance detection reliability, a method that combines results from multiple images of the same object captured from different angles has been proposed. This multi-view analysis allows for a reduction in the detection confidence threshold, increasing the true positive rate. Therefore, offered technique in article is a combination of YOLOv8-Obb, Augmentation, Transfer Learning and Multi-View Analysis. The proposed system was tested on a dedicated dataset of AlSiC products, achieving a defect detection rate of over 80% with a false alarm probability of approximately 1%. The results demonstrate the feasibility of using modern neural network-based detectors for automated visual inspection in specialized industrial applications.

Three methods of spectral matching for remote sensing data are studied: a pixelwise linear method, a pixelwise nonlinear method and a generalized nonlinear method. Nonlinear methods are implemented as a pair of multilayer perceptrons and a pair of convolutional neural networks respectively. Training and comparison of methods are performed using Landsat-8 and Sentinel-2 remote sensing images from 2021 IEEE GRSS Data Fusion Contest dataset. The root mean squared error (RMSE), the normalized mutual information (NMI) and the structural similarity index measure (SSIM) are used as metrics. A generalized nonlinear method demonstrates the best quality of spectral matching, achieving average values of RMSE = 0.048, NMI = 1.194 and SSIM = 0.887 over the testing set. A linear pixelwise method achieves RMSE = 0.075, NMI = 1.118 and SSIM = 0.847, a nonlinear pixelwise method achieves RMSE = 0.074, NMI = 1.117 and SSIM = 0.843. All methods show a significant improvement when compared to results without spectral matching (RMSE = 0.158, NMI = 0.119, SSIM = 0.585).