Programmable optoelectronic neural network for optimization

An optoelectronic neural network is presented that is designed to solve the assignment problem--or any similar optimization task given minimal adjustment--in both crossbar and banyan packet switches. We examine the design decisions made at the hardware, software, and algorithmic levels and indicate the associated effect on the system as a whole. Clearly detailed experimental results show the system's robustness and performance due to the particular optoelectronic-algorithm combination used. The integration and packaging of such a system are also briefly discussed.     

High-capacity photorefractive neural network implementing a kohonen topological map

We designed and built a high-capacity neural network based on volume holographic interconnections in a photorefractive crystal. We used this system to implement a Kohonen topological map. We describe and justify our optical setup and present some experimental results of self-organization in the learning database.     

Design considerations and algorithms for partitioning optoelectronic multichip modules

There is considerable interest in the development of optical interconnects for multichip modules (MCM's) to improve their performance. For effective utilization of the optical and electronic technologies, a methodology for partitioning the system is required. The key question to be answered is which technology should be used for each interconnect in a given netlist: optical or electronic. We introduce the computer-aided design approach for partitioning optoelectronic systems into optoelectronic MCM's. We first discuss the design trade-off issues in an optoelectronic system design, including speed, power dissipation, area, and diffraction limits for free-space optics. We then define a formulation for optoelectronic MCM partitioning and describe new algorithms for optimizing this partitioning based on the minimization of the power dissipation. The models for the algorithms are discussed in detail, and an example of a multistage interconnect network is given. Different results, with the number and size of chips being variable, are presented in which improvement for the system packaging has been observed when the partitioning algorithms are applied.     

Performance of an optoelectronic neural network in the presence of noise

Optoelectronic neural networks must not only be highly parallel but also fast to compete with electrical systems. Receiver noise becomes an important consideration at high data rates; so the limits set by noise to network size and speed are analyzed. A network incorporating an array of high-speed multi-quantum-well modulators was constructed. It employed a general method for optical representation of bipolar values, which required only a minimal increase in network dimensions and gave the network immunity to common-mode parameter variations. Different ways of partitioning pattern-recognition problems were compared, and the accuracy of one configuration was tested with the experimental network over a range of noise levels.     

Stochastic sensitivity analysis and Langevin simulation for neural network learning

A comprehensive theoretical framework is proposed for the learning of a class of gradient-type neural networks with an additive Gaussian white noise process. The study is based on stochastic sensitivity analysis techniques, and formal expressions are obtained for stochastic learning laws in terms of functional derivative sensitivity coefficients. The present method, based on Langevin simulation techniques, uses only the internal states of the network and ubiquitous noise to compute the learning information inherent in the stochastic correlation between noise signals and the performance functional. In particular, the method does not require the solution of adjoint equations of the back-propagation type. Thus, the present algorithm has the potential for efficiently learning network weights with significantly fewer computations. Application to an unfolded multi-layered network is described, and the results are compared with those obtained by using a back-propagation method.     

Design and analysis of a diffractive optical filter for use in an optoelectronic error-diffusion neural network

The design, fabrication, experimental characterization, and system-performance analysis of a diffractive optical implementation of an error-diffusion filter for use in digital image halftoning is reported. A diffractive optical filter was fabricated as an eight-level phase element that diffuses the quantization error nonuniformly in both the weighting and the spatial dimensions, according to a prescribed algorithm. Ten identical diffractive elements were fabricated on ten different wafers and subsequently characterized experimentally. A detailed error analysis including both fabrication and instrumentation errors was carried out to quantify the performance of the fabrication process as well as the expected system performance of the filters. Halftone system performance was evaluated by use of the experimental filter's performance and both quantitative and qualitative performance metrics. The results of this analysis demonstrate that multiple identical copies of a diffractive optical filter can be produced with sufficient accuracy that no loss in the halftoning system performance results.     

超弹性形状记忆合金的神经网络连续本构模型

对超弹性形状记忆合金(SMA)丝在不同应变幅值和荷载速率下进行加卸载单轴拉伸试验,分析其滞回特性随环境因素的变化规律。将径向基函数神经网络(RBFNN)和Graesser模型结合起来,Graesser模型参数取自试验曲线,能由数学式确定的模型参数和应变幅值、荷载速率一起作为网络的输入信息,不能由数学式确定的模型参数作为输出神经元。数值计算表明,RBFNN可以精确地预测Graesser模型参数,且计算的SMA应力-应变曲线与Graesser模型结果吻合很好。     

形状记忆合金超弹性本构关系的神经网络模型

通过试验，研究了受过循环变形、具有稳定超弹性变形性能的形状记忆合金丝在拉伸到不同应变幅值条件下卸载的超弹性变形行为。根据试验测得的结果，提出了基于神经网络的形状记忆合金超弹性本构关系模型，并把模型计算的结果和实验数据进行了比较分析，结果表明，该模型具有很高的精度。该模型避免了已有模型在参数确定上的困难，具有一定的工程应用价值，为建立形状记忆合金本构模型提供了一个新的思路。     

Optical learning neural network with a pockels readout optical modulator

We have constructed an optical neural-network system with learning capability by using a Pockels readout optical modulator. The system has a two-dimensional structure that permits easy optical alignment and can handle images without scanning. Learning signals are calculated optically with two liquid-crystal devices by a matrix-matrix outer-product method. The calculated learning signals are added directly to the weights memorized on the Pockels readout optical modulator. A two-layer network is implemented, and the learning and recognition of four alphabetical characters are realized according to the delta rule.     

Specification for a reconfigurable optoelectronic VLSI processor suitable for digital signal processing

A concept for a parallel digital signal processor based on opticalinterconnections and optoelectronic VLSI circuits is presented. Itis shown that the proper combination of optical communication, architecture, and algorithms allows a throughput that outperformspurely electronic solutions. The usefulness of low-level algorithmsfrom the add-and-shift class is emphasized. These algorithms leadto fine-grain, massively parallel on-chip processor architectures withhigh demands for optical off-chip interconnections. A comparativeperformance analysis shows the superiority of a bit-serialarchitecture. This architecture is mapped onto an optoelectronicthree-dimensional circuit, and the necessary optical interconnectionscheme is specified.     

数字化光电跟踪伺服系统定位极限环的抑制

数字光电跟踪系统定位时将产生自振荡,在相平面下可描述为极限环。自振荡对定位精度影响较大,因此必须对其进行抑制。针对实际的数字光电跟踪伺服系统,提出一种描述方法,即用等效间隙非线性来描述其定位时非线性特性,且该间隙主要由光电传感器的滞后效应和伺服系统的摩擦死区引起。理论分析表明,双极PWM功率放大器的零点振颤特性可以抑制伺服系统的定位极限环。实验也证明,通过软件补偿以及PWM驱动模式设计来改善光电跟踪系统定位精度是可行的,定位精度由1.77提高到了22.23,大大抑制了定位非线性特性。     

一种改进的人工神经网络学习算法及其在超声检测中的应用

本文用多层感知器 (MLP)与误差反向传播算法 (errorback propagationalgorithm)构造训练人工神经网络 ,提出了新的误差反向传播改进算法。试验结果表明 ,改进的BP算法收敛速度较之常规BP算法明显加快 ,因而在工业现场的超声检测领域有广阔应用前景。     

Dynamic holographic interferometry by photorefractive crystals for quantitative deformation measurements

A holographic interferometer that uses two-wave mixing in a photorefractive (Bi12SiO20) crystal under an applied ac field is described. The interferometer uses a repetitive sequence of separate record and readout times to obtain quasi real-time holographic interferograms of vibrating objects. It is shown that a good signal-to-noise ratio of the interferometer is obtained by turning off the object illumination and the applied ac field during readout of the hologram. The good signal-to-noise ratio of the resulting holographic interferograms enables phase measurement, which allows for quantitative deformation analysis.     

Performance degradation assessment of rolling bearing based on convolutional neural network and deep long-short term memory network

Many traditional approaches for performance degradation assessment of rolling bearings, using sensor data, make assumptions about how they degrade or fault evolve. However, the sequential sensor data cannot be directly taken as input in the traditional models since the data always contain noise and change in length. To solve these problems, a convolutional neural network and deep long-short term memory (CNN-DLSTM) based architecture is proposed to obtain an unsupervised H-statistic for performance degradation assessment of rolling bearing using sensor time-series data. Firstly, a CNN is applied to extract local abstract features from raw sensor data. Secondly, a deep LSTM is explored to extract temporal features. CNN-DLSTM is trained to reconstruct the time-series sensor signal reflecting the health condition of rolling bearing. The D- and Q-statistic are used to compute H-statistic which is then used for performance degradation assessment. The proposed approach is evaluated on an experiment with rolling bearings and the results are presented on a public dataset of rolling bearing, verifying that the proposed approach outperforms several state-of-the-art methods.     

Fast image vector quantization using a modified competitive learning neural network approach

The basic goal of image compression through vector quantization (VQ) is to reduce the bit rate for transmission or data storage while maintaining an acceptable fidelity or image quality. The advantage of VQ image compression is its fast decompression by table lookup technique. However, the codebook supplied in advance may not handle the changing image statistics very well. The need for online codebook generation became apparent. The competitive learning neural network design has been used for vector quantization. However, its training time can be very long, and the number of output nodes is somewhat arbitrarily decided before the training starts. Our modified approach presents a fast codebook generation procedure by searching for an optimal number of output nodes evolutively. The results on two medical images show that this new approach reduces the training time considerably and still maintains good quality for recovered images. © 1997 John Wiley & Sons, Inc. Int J Imaging Syst Technol, 8, 413–418, 1997     

人工神经网络在电积锌工艺中的应用

本文提出了一种基于人工神经网络分析电积锌工艺的方法 ,通过仿真计算 ,表明这种方法切实可行     

Dynamic analysis of hydrodynamic bearing–rotor system based on neural network

Feed-forward neural network is employed to model the nonlinear oil-film force database of a finite-length hydrodynamic journal bearing, which is constructed by continuous transformation of Reynolds equation. Neural network models trained are utilized to investigate motion characteristics of a rigid unbalanced rotor supported on elliptical bearings in 300 MW steam turbine generator set. There exist various forms of periodic, quasi-periodic and chaotic motions at different rotating speeds. Periodic doubling bifurcation and quasi-periodic routes to chaos may be found when rotating speed is used as the control parameter. Computational results show that there exist similar motion behaviors between neural networks and numerical method. It is available for neural network models of oil-film forces to research nonlinear dynamic problems of rotating machinery.     

Lymph node detection method based on multisource transfer learning and convolutional neural network

Recently years, convolutional neural networks (CNNs) have proven to be powerful tools for a broad range of computer vision tasks. However, training a CNN from scratch is difficult because it requires a large amount of labeled training data, which remains a challenge in medical imaging domain. To this end, deep transfer learning (TL) technique is widely used for many medical image tasks. In this paper, we propose a novel multisource transfer learning CNN model for lymph node detection. The mechanism behind it is straightforward. Point-wise (1 × 1) convolution is used to fuse multisource transfer learning knowledge. Concretely, we view the transferred features as priori domain knowledge and 1 × 1 convolutional operation is implemented after pre-trained convolution layers to adaptively combine the transfer information for target task. In order to learn non-linear transferred features and prevent over-fitting, we present an encode process for the pre-trained convolution kernels. At last, based on convolutional factorization technique, we train the proposed CNN model and the encoder process jointly, which improves the feasibility of our approach. The effectiveness of the proposed method is verified on lymph node (LN) dataset: 388 mediastinal LNs labeled by radiologists in 90 patient CT scans, and 595 abdominal LNs in 86 patient CT scans for LN detection. Our method demonstrates sensitivities of about 85%/71% at 3 FP/vol. and 92%/85% at 6 FP/vol. for mediastinum and abdomen respectively, which compares favorably to previous methods.