Ultrasonic signal/image restoration for quantitative NDE

Our overall objective is to improve the spatial resolution of ultrasonic images used for quantitative non-destructive evaluation of materials. These images contain distortion due to the band-pass frequency responses of the associated transducers and propagation paths. Our specific objective is to estimate the impulse responses of flaws and interfaces in material samples. Thus, we are attacking the restoration problem, which is, in general, ill-posed and difficult to solve in practice. We have applied several techniques for regularizing the problem, including both time- and frequency-domain methods, and some constrained optimization algorithms. Our new results demonstrate that regularization techniques can allow successful, practical solutions to this ill-posed problem. Our work concentrates mainly upon the problem in which the impulse response is expected to be impulsive (consisting of a weighted sum of delayed impulses). Our results show that some problems of practical interest are tractable by using a two-step procedure involving first system identification and then spectrum extrapolation to obtain sharper spatial impulses. The paper presents the results of experiments using actual laboratory signals, in which we demonstrate improved spatial resolution for impulsive signals, allowing more meaningful interpretation of ultrasonic A- and B-scans.     

Antinoise approximation of the lidar signal with wavelet neural networks

We propose a new, to our knowledge, denoising method for lidar signals based on a regression model and a wavelet neural network (WNN) that permits the regression model not only to have a good wavelet approximation property but also to make a neural network that has a self-learning and adaptive capability for increasing the quality of lidar signals. Specifically, we investigate the performance of the WNN for antinoise approximation of lidar signals by simultaneously addressing simulated and real lidar signals. To clarify the antinoise approximation capability of the WNN for lidar signals, we calculate the atmosphere temperature profile with the real signal processed by the WNN. To show the contrast, we also demonstrate the results of the Monte Carlo moving average method and the finite impulse response filter. Finally, the experimental results show that our proposed approach is significantly superior to the traditional methods.     

Modelling metabolic energy by neural networks

The apparent metabolic energy (EMA) of barley is modelled as a function of 12 easily obtainable analytical parameters by applying neural networks with the error back-propagation learning strategy. Kohonen maps and Ward's clustering technique have been used to define the objects for the training and test sets. The architecture of the neural network and the relevant parameters of error back-propagation learning have been optimised providing a RMS of 1.081 and a correlation coefficient (predicted versus found values) of 0.82. Contour maps of all variables including the output EMA value have been obtained by applying the counter-propagation learning strategy in a two-layer neural network. The responses yielded by the networks show that this method is capable of establishing a quantitative relationship between EMA and the original variables.     

基于神经网络和功率谱的水中目标信号检测方法

将功率谱和神经网络相结合，应用于高海况、低信噪比条件下，水中目标信号的特征提取中．文中首先对信号进行功率谱估计，利用目标信号功率主要集中在低频部分的特点，提取低频信号的能量作为特征，然后利用人工神经网络对目标信号进行检测．利用不同浪级情况下海洋水压场的仿真信号数据，对某型目标舰船的水压信号进行了检测计算，验证了该方法的有效性，尤其是达到了在高海况、低信噪比条件下，对目标信号检测率比较高、虚警率比较低的效果．     

Ultrasonic signal processing using Born inversion

A pulse-echo (1D) Born inversion technique, which utilizes digital signal processing on ultrasonic back-scattered signals to size defects, is briefly outlined together with its application to the measurement of spheroidal defects in metals. The theoretical derivation of the algorithm is valid for application to weak scatterers; however, it is shown experimentally that this technique can give accurate radius predictions for strong scatterers such as voids. This apparent anomaly is examined using theoretically generated exact scattering data. A methodology for testing a specific transducer's suitability for sizing a specific defect is reported and demonstrated.     

Damage detection in aluminum and composite elements using neural networks for Lamb waves signal processing

One of the important factors in the structural health monitoring systems is the amount of data that need to be analysed in real time. This study investigated the use of artificially deteriorated signals of Lamb waves in training the novelty detection (ND) system for the early damage detection. In this system Auto-associative Neural Networks were trained using principal components calculated on the basis of experimentally measured signals. The specimens studied relate to two different materials commonly used in the aerospace industry, i.e. aluminium and glass fibre reinforced polymer. Lamb waves measured in these specimens are a good example that the ND algorithm works correctly in case of simple as well as complex signals. Furthermore, it was found that the designed ND system remained sensitive and robust even when it used raw signals with a relatively low sampling rate, on a fairly narrow time window and even noised signals.     

Hybrid discrete-time neural networks

Hybrid dynamical systems combine evolution equations with state transitions. When the evolution equations are discrete-time (also called map-based), the result is a hybrid discrete-time system. A class of biological neural network models that has recently received some attention falls within this category: map-based neuron models connected by means of fast threshold modulation (FTM). FTM is a connection scheme that aims to mimic the switching dynamics of a neuron subject to synaptic inputs. The dynamic equations of the neuron adopt different forms according to the state (either firing or not firing) and type (excitatory or inhibitory) of their presynaptic neighbours. Therefore, the mathematical model of one such network is a combination of discrete-time evolution equations with transitions between states, constituting a hybrid discrete-time (map-based) neural network. In this paper, we review previous work within the context of these models, exemplifying useful techniques to analyse them. Typical map-based neuron models are low-dimensional and amenable to phase-plane analysis. In bursting models, fast-slow decomposition can be used to reduce dimensionality further, so that the dynamics of a pair of connected neurons can be easily understood. We also discuss a model that includes electrical synapses in addition to chemical synapses with FTM. Furthermore, we describe how master stability functions can predict the stability of synchronized states in these networks. The main results are extended to larger map-based neural networks.     

Structural optimization by gradient‐based neural networks

In this paper a neurocomputing strategy is presented which combines data processing capabilities of neural networks and numerical structural optimization. In this strategy, an improved counterpropagation neural network is used. Two artificial neural networks are trained, one for the constraints and the other for the gradients of the constraints and structural optimization is accomplished by using these nets. All required parameters such as weight matrices in the neural networks or the gradient computations are automated in this neuro‐optimizer strategy. Numerical examples are included to demonstrate the accuracy of the results. Copyright © 1999 John Wiley & Sons, Ltd.     

人工神经网络的单调序列逼近

设Nn,Ф是以Ф为激活函数的具有n+1个神经元的前向单隐层人工神经网络的全体.主要证明了,若f∈C[0,1],则对于任意的ε>0,存在两个神经网络序列{Pn,Ф}和{Qn,Ф},使得在[0,1]上Qn,Ф(x)≤Qn+1,Ф(x)≤f(x)≤Pn+1,Ф(x)≤Pn,Ф(x),而且Pn,Ф(x)-Qn,Ф(x)≤(6+4(2~(1/2)))En,Ф(f),这里的En,Ф(f)为Nn,Ф中的元对f的最佳逼近.     

Improvement in transient piezoelectric responses of NDE transceivers using selective damping and tuning networks

NDE ultrasonic applications for quality control purposes, based on piezoelectric devices working as pulsed ultrasonic probes, usually include some tuning circuit either across the pulser output connectors or close to the piezoelectric probe electrodes. Classic criteria to determine tuning parameters in ultrasonic transducers do not perform very well in transient regime under the typical electrical conditions and peculiar output impedances of the required high-voltage pulsers. In most practical situations, the tuning values are manually adjusted in specific circuits for each NDE case, testing each channel of a multiprobe inspection system to find the best sensitivity. In this paper, the positive influences of certain selective damping and tuning networks on the time and frequency behaviors of NDE piezoelectric transceivers are analyzed in detail. Different waveforms and spectra of received ultrasonic signals are comparatively evaluated for different real tuning conditions, after passing through two ultrasonic processes linked in cascade: a transmitter of industrial type and a broad-band ultrasonic receiver. Piezoelectric transducer impedances, transduction transfer functions, and HV output spikes from a piezoelectrically loaded NDE pulser, are computed, to identify separately the influence of each stage. In order to improve amplitude and axial resolution in the received transient responses, relative contributions from emission and reception tuning networks are individually evaluated for a particular NDE two-stage piezoelectric channel. Also shown are the final experimental results relative to the ultrasonic test pulse and detected in a through-transmission NDE configuration, gradually improved by using distinct tuning levels.     

Job-shop scheduling using neural networks

Complete enumeration of all sequences to establish global optimality is not feasible as the search space; for a general job-shop scheduling problem,PiG has an upper bound of (n!)m. Since the early fifties a great deal of research attention has been focused on solving PiG , resulting in a wide variety of approaches such as branch and bound, simulated annealing, tabu search, etc. However, limited success has been achieved by these methods due to the shear intractability of this generic scheduling problem. Recently, much effort has been concentrated on using neural networks to solve PiG as they are capable of adapting to new environments with little human intervention and can mimic thought processes. Major contributions in solving PiG using a Hopfield neural network, as well as applications of back-error propagation to general scheduling problems are presented. To overcome the deficiencies in these applications a modified back-error propagation model, a simple yet powerful architecture which can be successfully simulated on a personal computer, is applied to solve PiG.     

DeepLip: block-based lip pixel detection by deep neural networks

This paper presents an effective lip pixel detection method based on blocks and deep neural networks. Since only-rough localization of a pair of lips is a trivial task, we use a rectangle that loosely bounds two lips as an input region of interest for lip detection. For each pixel in the rectangle region we generate a block whose center is at the pixel, and the pixel is classified into either a lip or non-lip pixel by exploiting the pixels in the block. Deep neural networks are trained using a sufficient number of labeled blocks obtained from a quite tractable number of labeled images. As a result, lip pixels are detected with high accuracy despite negligible labeling effort. Experimental results demonstrate the effectiveness of the presented method. We show that even single-minute training can outperform the mouth map with the best threshold.     

一类神经网络逼近可积函数

用连续模刻画了实轴上Cardaliguet-Eurrard型神经网络算子逼近连续函数速度的上界估计,同时,对于Lebesgue可积函数的逼近,构造相应的神经网络算子,并且给出其逼近速度的Jackson型估计.     

Using neural networks to study networks of scientific journals

In this paper a new approach to study science dynamics is introduced. This approach is based in the use of Kohonen preserving topology maps, a kind of neural network. Four data set consisting in cross-citation matrix are studied using this approach. Relations maps and domains maps are computed for these data sets and interrelationships among journals are studied. This approach allow to stude both, hierarchical journal structure in a given time and evolution of relations among journals in a given time lag.     

Eddy-current signal analysis and inversion for semielliptical surface cracks

A scalar potential formulation of the Z formula for the change in impedance of an eddy-current probe caused by a surface-breaking flaw is developed. The resulting formula is evaluated using a finite-difference method, which permits calculation of Z for semielliptical flaws. The numerical results are checked by comparing calculations for rectangular-shaped flaws to previous calculations using an analytical solution for this geometry. Theoretical results are then verified by comparison with measurements on semielliptical fatigue cracks and EDM notches in aluminum alloy specimens using air-core eddy-current probes. An inversion method that compares features of the flaw profile, obtained by scanning the eddy-current probe along the length of the flaw, to a theoretical inversion chart (McFetridge chart) is demonstrated using the experimental data.     

Detection and classification mean-shifts in multi-attribute processes by artificial neural networks

To monitor the quality of a multi-attribute process, some issues arise. One of them being the occurrence of a high number of false alarms (type I error) and the other an increase in the probability of not detecting defects when the process is monitored by a set of independent uni-attribute control charts. In this paper, based upon the artificial neural network capabilities we develop a new methodology to overcome this problem. We design a perceptron neural network to monitor either the proportions of several types of product nonconformities (instead of using several np charts) or the number of different types of defects (instead of using several c charts) in a product. Moreover, while the proposed method possesses the ability to be applied for small sample sizes, it is also able to diagnose the mean shift online. We present two simulation experiments in which the proportions of several types of nonconformities are monitored. In addition, we present one more simulation experiment in which the number of different types of defect is controlled. We also compare the performance of the proposed methodology with the ones from the Mnp and T ² charts for multi-attribute processes. The results of the simulation studies are encouraging.     

基于神经网络的电机转矩软测量研究

径向基函数神经网络（RBFNN）具有最优逼近和全局逼近的特性，在函数拟合方面优于传统的BP网络，将在化工领域广泛使用的软测量技术应用于电机系统的转矩测量，该方法的可行性进行了论证，并运用RBF神经网络建立转矩的软测量模型。同时建立了基于BP神经网络的软测量模型，用改进的kvenberg—Marquardt算法对BP神经网络进行学习和训练，并对两种网络进行了对比。该方法只需要电流信息，辨识方法简单。研究表明，RBF神经网络辨识效果优于BP神经网络。     

Autonomous underwater vehicle guidance by integrating neural networks and geometric reasoning

This paper presents a method for guiding an autonomous underwater vehicle (AUV) during sea bottom inspection missions. The vehicle is equipped with several sensors (optical, sonar, acoustic) and is able to detect and follow a pipeline placed on the sea bottom. Neural networks and geometric reasoning methods are integrated to perform a real-time identification of pipeline borders in a complex underwater environment. Different scenarios characterized by both obstacles and/or artifacts (due to reflections of artificial light sources used by the vehicle to illuminate the scene) have been considered. Results focus on pipeline detection accuracy and on AUV missions in the absence or presence of down stream and/or obstacles. © 1999 John Wiley & Sons, Inc. Int J Imaging Syst Technol 10, 385–396, 1999