Analog fault diagnosis of actual circuits using neural networks

We have developed a neural-network based analog fault diagnostic system for actual circuits. Our system uses a data acquisition board to excite a circuit with an impulse and sample its output to collect training data for the neural network. The collected data is preprocessed by wavelet decomposition, normalization, and principal component analysis (PCA) to generate optimal features for training the neural network. This ensures a simple architecture for the neural network and minimizes the size of the training set required for its proper training. Our studies indicate that features extracted from actual circuits lie closer to each other and exhibit more overlap across fault classes compared to SPICE simulations. This implies that the neural network architecture which can most reliably perform fault diagnosis of actual circuits is one whose outputs estimate the probabilities that input features belong to different fault classes. Our work also shows that SPICE simulations can be used to select appropriate features for training the neural network. Reliable diagnosis of faults in an actual circuit, however, requires training data from the circuit itself. Our fault diagnostic system, trained and tested using data obtained from real sample circuits, achieves 95% accuracy in classifying faulty components     

Pattern recognition of visible and near-infrared spectroscopy from bayberry juice by use of partial least squares and a backpropagation neural network

Visible and near-infrared reflectance (visible-NIR) spectroscopy is applied to discriminate different varieties of bayberry juices. The discrimination of visible-NIR spectra from samples is a matter of pattern recognition. By partial least squares (PLS), the spectrum is reduced to certain factors, which are then taken as the input of the backpropagation neural network (BPNN). Through training and prediction, three different varieties of bayberry juice are classified based on the output of the BPNN. In addition, a mathematical model is built and the algorithm is optimized. With proper parameters in the training set, 100% accuracy is obtained by the BPNN. Thus it is concluded that the PLS analysis combined with the BPNN is an alternative for pattern recognition based on visible and NIR spectroscopy.     

基于卷积神经网络的光学遥感图像船只检测

目的研究无需进行复杂的图像预处理和人工特征提取,就能提高光学遥感图像的船只检测准确率和实现船只类型精细分类。方法对输入的检测图像,采用选择性搜索的方法产生船只候选区域,用已经标记好的训练样本对卷积神经网络进行监督训练,得到网络参数,然后使用经过监督训练的卷积神经网络提取抽象特征,并对候选区域进行分类,根据船只候选区域的分类概率同时确定船只的位置以及类型。结果与现有的2种检测方法进行对比,实验结果表明卷积神经网络能有效提高船只检测准确率,平均检测准确率达到了93.3%。结论该检测方法无需进行复杂的预处理,能同时对船只进行检测和分类,并能有效提高船只检测准确率。     

Neural Network Burst Pressure Prediction in Composite Overwrapped Pressure Vessels Using Mathematically Modeled Acoustic Emission Failure Mechanism Data

The purpose of this research was to predict burst pressures in composite overwrapped pressure vessels (COPVs) by using mathematically modeled acoustic emission (AE) data. Both backpropagation neural network (BPNN) and multiple linear regression (MLR) analyses were performed on various subsets of the low proof pressure AE data to predict burst pressures and to determine if the two methods were comparable. AE data were collected during hydrostatic burst testing on the 15-inch diameter COPVs. Once collected, the AE data were filtered to eliminate noise then classified into AE failure mechanism data using a MATLAB Kohonen self-organizing map (SOM). The matrix cracking only amplitude distribution data were mathematically modeled using bounded Johnson distributions with the four Johnson distribution parameters – ?, λ, γ, and η – employed as inputs to make both the BPNN and MLR predictions. The burst pressure predictions generated using a MATLAB BPNN resulted in a worst case error of 1.997% as compared to ?1.666% for the MLR analysis, suggesting comparability. However, the MLR analysis required the data from all nine COPVs to get approximately the same results as the BPNN training on just five COPVs; plus, MLR analyses are intolerant to noise, whereas BPNNs are not.     

Classification of audio radar signals using radial basis function neural networks

Radial basis function (RBF) neural networks are used to classify real-life audio radar signals that are collected by a ground surveillance radar mounted on a tank. Currently, a human operator is required to operate the radar system to discern among signals bouncing off tanks, vehicles, planes, and so on. The objective of this project is to investigate the possibility of using a neural network to perform this target recognition task, with the aim of reducing the number of personnel required in a tank. Different signal classification methods in the neural net literature are considered. The first method employs a linear autoregressive (AR) model to extract linear features of the audio data, and then perform classification on these features, i.e, the AR coefficients. AR coefficient estimations based on least squares and higher order statistics are considered in this study. The second approach uses nonlinear predictors to model the audio data and then classifies the signals according to the prediction errors. The real-life audio radar data set used here was collected by an AN/PPS-15 ground surveillance radar and consists of 13 different target classes, which include men marching, a man walking, airplanes, a man crawling, and boats, etc. It is found that each classification method has some classes which are difficult to classify. Overall, the AR feature extraction approach is most effective and has a correct classification rate of 88% for the training data and 67% for data not used for training.     

Characterization of random rough surfaces from scattered intensities by neural networks

Abstract

Optical scatterometry, a non-invasive characterization method, is used to infer the statistical properties of random rough surfaces. The Gaussian model with rms-roughness [sgrave] and correlation length σ is considered in this paper but the employed technique is applicable to any representation of random rough surfaces. Surfaces with wide ranges of Λ and σ, up to 5 wavelengths (λ), are characterized with neural networks. Two models are used: self-organizing map (SOM) for rough classification and multi-layer perceptron (MLP) for quantitative estimation with nonlinear regression. Models infer Λ and σ from scattering, thus involving the inverse problem. The intensities are calculated with the exact electromagnetic theory, which enables a wide range of parameters. The most widely known neural network model in practise is SOM, which we use to organize samples into discrete classes with resolution ΔΛ = Δσ = 0.5λ. The more advanced MLP model is trained for optimal behaviour by providing it with known parts of input (scattering) and output (surface parameters). We show that a small amount of data is sufficient for an excellent accuracy on the order of 0.3λ and 0.15λ for estimating Λ and σ, respectively.     

Analysis of one-dimensional stochastic finite elements using neural networks

This paper explores the applicability of neural networks for analyzing the uncertainty spread of structural responses under the presence of one-dimensional random fields. Specifically, the neural network is intended to be a partial surrogate of the structural model needed in a Monte Carlo simulation, due to its associative memory properties. The network is trained with some pairs of input and output data obtained by some Monte Carlo simulations and then used in substitution of the finite element solver. In order to minimize the size of the networks, and hence the number of training pairs, the Karhunen–Loéve decomposition is applied as an optimal feature extraction tool. The Monte Carlo samples for training and validation are also generated using this decomposition. The Nyström technique is employed for the numerical solution of the Fredholm integral equation. The radial basis function (RBF) network was selected as the neural device for learning the input/output relationship due to its high accuracy and fast training speed. The analysis shows that this approach constitutes a promising method for stochastic finite element analysis inasmuch as the error with respect to the Monte Carlo simulation is negligible.     

NEW REGULARIZATION BY TRANSFORMATION FOR NEURAL NETWORK BASED INVERSE ANALYSES AND ITS APPLICATION TO STRUCTURE IDENTIFICATION

The present authors have been developing an inverse analysis approach using the multilayer neural network and the computational mechanics. This approach basically consists of the following three subprocesses. First, parametrically varying model parameters of a system, their corresponding responses of the system are calculated through computational mechanics simulations such as the finite element analyses, each of which is an ordinary direct analysis. Each data pair of model parameters vs. system responses is called training pattern. Second, a neural network is iteratively trained using a number of training patterns. Here the system responses are given to the input units of the network, while the model parameters to be identified are shown to the network as teacher data. Finally, some system responses measured are given to the well-trained network, which immediately outputs appropriate model parameters even for untrained patterns. This is an inverse analysis. This paper proposes a new regularization method suitable for the inverse analysis approach mentioned above. This method named the Generalized-Space-Lattice (GSL) transformation transforms original input and/or output data points of all training patterns onto uniformly spaced lattice points over a multi-dimensional space. The topological relationships among all the data points are maintained through this transformation. The neural network is then trained using the GSL-transformed training patterns. Since this method significantly remedies localization of training patterns caused due to strong nonlinearity of problem, the neural network can learn the training patterns efficiently as well as accurately. Fundamental performances of the present inverse analysis approach combined with the GSL transformation are examined in detail through the identification of a vibrating non-uniform beam in Young's modulus based on the observation of its multiple eigenfrequencies and eigenmodes.     

Optimization of principal-component-analysis-applied in situ spectroscopy data using neural networks and genetic algorithms

A new model of multidimensional in situ diagnostic data is presented. This was accomplished by combining a back-propagation neural network (BPNN), principal component analysis (PCA), and a genetic algorithm (GA). The PCA was used to reduce input dimensionality. The GA was applied to search for a set of optimized training factors involved in BPNN training. The presented technique was evaluated with optical emission spectroscopy (OES) data measured during the etching of oxide thin films in a CHF(3)-CF(4) inductively coupled plasma. For a systematic modeling, the etching process was characterized by a face-centered Box Wilson experiment. The etch responses to be modeled include oxide etch rate, oxide profile angle, and oxide etch rate non-uniformity. In PCA, three types of data variances were employed and the reduced input dimensionality corresponding to 100, 99, and 98% are 16, 8, and 5. The BPNN training factors to be optimized include the training tolerance, number of hidden neurons, magnitude of initial weight distribution, gradient of bipolar sigmoid function, and gradient of linear function. The prediction errors of GA-BPNN models are 249 A/min, 2.64 degrees, and 0.439% for the etch rate, profile angle, and etch rate non-uniformity, respectively. Compared to the conventional and previous full OES models, the presented models demonstrated a significantly improved prediction for all etch responses.     

Identification of Materials Properties with the Help of Miniature Shear Punch Test Using Finite Element Method and Neural Networks

This paper describes an approach to identify the mechanical properties i.e. fracture and yield strength of steels. The study involves the FE simulation of shear punch test for various miniature specimens thickness ranging from 0.20mm to 0.80mm for four different steels using ABAQUS code. The experimental method of the miniature shear punch test is used to determine the material response under quasi-static loading. The load vs. displacement curves obtained from the FE simulation miniature disk specimens are compared with the experimental data obtained and found in good agreement. The resulting data from the load vs. displacement diagrams of different steels specimens are used to train the neural networks to predict the properties of materials i.e. fracture and yield strength. Two different feed forward neural networks have been created and trained in order to predict the Fracture toughness and yield strength values of different steels. L-M algorithm has been used in the networks to form an output function corresponding to the input vectors used in the network. The trained network provides the output values i.e., fracture toughness and yield strength of unknown input values, which are within in the range of data that is used for the training of network.     

基于自回归模型和学习矢量量化神经网络的手指动作识别

手指的力量和动作是反映手指协同运动、评价手部运动机能的重要参数。本文提出了一种以自回归（Auto-regressive,AR）模型和学习矢量量化（Learning Vector Quantization,LVQ）网络相结合的表面肌电信号处理方法。13名受试者参与了目标力量为4N、6N、8N等三个力量等级的指力跟踪实验,对指力信号和前臂指浅屈肌（flex digitorum superficials,FDS）、指伸肌（extensor digitorum,ED）的表面肌电信号进行了同步记录;通过对采集到的肌电信号进行预处理,提取AR系数作为其特征值;然后设计了一个LVQ神经网络,对同等力量水平下食指、中指的动作进行模式分类,分类正确率在80％以上。实验表明,表面肌电信号（surface Eleetromyography,sEMG）与手指动作具有相关性,使用AR结合LVQ的sEMG有较高的识别率。     

基于GA-BP神经网络的结构损伤位置识别

针对传统BP神经网络训练中存在的一些问题，提出了一种基于遗传算法(GA)-BP神经网络混合技术识别结构损伤位置的方法。该方法利用基因实数编码的遗传算法优化BP网络的结构及初始参数，从而大大提高了神经网络的训练精度。运用GA-BP网络与传统BP网络技术分别对两个算例进行了结构损伤定位的识别仿真，结果表明遗传BP稳定性好，精度高，对噪声有很好的鲁棒性，便于工程应用。     

Next-day electricity-price forecasting using a hybrid network

The paper proposes a novel method of forecasting short-term electricity price based on a two-stage hybrid network of self-organised map (SOM) and support-vector machine (SVM). In the first stage, a SOM network is applied to cluster the input-data set into several subsets in an unsupervised manner. Then, a group of SVMs is used to fit the training data of each subset in the second stage in a supervised way. With the trained network, one can predict straightforwardly the next-day hourly electricity prices. To confirm its effectiveness, the proposed model has been trained and tested on the data of historical energy prices from the New England electricity market.     

PSO-Gabor-CNN算法在印刷品套印缺陷的检测

目的二维Gabor滤波器含有多个参数,在印刷品套印缺陷检测中,二维Gabor滤波器使用不同参数增强图像特征的效果差别较大,为了获得二维Gabor在某印刷品套印缺陷检测下的优化参数。方法在印刷品套印缺陷检测中,提出一种PSO-Gabor-CNN算法,采用Sobel算子对印刷品图像进行边缘检测,以粒子群算法(PSO)对二维Gabor滤波器的中心最大频率kmax、带宽σ、模板窗口window进行参数寻优,处理后的图像与模板图像采用加权欧式距离进行评价。然后用优化后的Gabor滤波器对图像进行滤波,最后采用卷积神经网络(CNN)对印刷品套印缺陷进行检测和分类。结果通过粒子群算法,确定了二维Gabor中心最大频率kmax为6.0476、带宽σ为0.1444、模板窗口window为27×27取得最佳效果,此时加权欧式距离为1.1927×10-33。卷积神经网络经过70次训练的均方误差为0.0035,测试样本正确率为96.93%。该方法与无数据预处理的BP神经网络(BPNN)、Sobel预处理的BP神经网络(Sobel-BPNN)、无数据预处理的卷积神经网络(CNN)、Sobel预处理的卷积神经网络(Sobel-CNN)对比,表现出了较好的识别效果。结论该方法可以获取二维Gabor滤波器的较优参数,从而获得较好的滤波效果,将其应用于套印缺陷检测,具有一定的应用价值。     

Modeling of Magnetic Properties of Nanocrystalline La-doped Barium Hexaferrite

In this paper an artificial neural network (ANN) has been developed to compute the magnetization of the pure and La-doped barium ferrite powders synthesized in ammonium nitrate melt. The input parameters were: the Fe/Ba ratio, La content, sintering temperature, HCl washing and applied magnetic field. A total of 8284 input data set from currently measured 35 different samples with different Fe/Ba ratios, La contents and washed or not washed in HCl were available. These data were used in the training set for the multilayer perceptron (MLP) neural network trained by Levenberg–Marquardt learning algorithm. The hyperbolic tangent and sigmoid transfer functions were used in the hidden layer and output layer, respectively. The correlation coefficients for the magnetization were found to be 0.9999 after the network was trained.     

Intelligent detection of microcalcification from digitized mammograms

This paper reports the design and implementation of an intelligent system for detection of microcalcification from digital mammograms. A neuron based thresholding strategy has been developed to reduce the number of candidate pixels. A back propagation neural network (BPNN) classifier has been used to classify the pixels into positive (affected) and normal ones. The false positives generated in the process are eliminated using the connected component analysis and the elongated component removal algorithms in succession. Suspected areas of microcalcification are detected and marked on the mammogram. The system was rigorously tested for the available images and was found to be quite robust, consistent and fast in detection. The output image with prompts generated by the system can form an important input to a radiologist for the final diagnosis.     

基于遗传算法的神经网络被动声呐目标分类研究

被动声呐目标识别系统中目标分类器的设计和训练是一项重要内容，本文设计了目标分类器的神经网络结构，提出了一种用改进的遗传算法训练神经网络分类器的新方法，最后，对海上实录的Ａ，Ｂ，Ｃ三类目标噪声进行了分类识别，实验结果表明基于遗传算法的神经网络分类器比传统的基于ＢＰ算法的神经网络分类源泛化性能有明显提高。     

Understanding weld modelling processes using a combination of trained neural networks

In this paper, a set of neural networks has been trained for weld modelling processes with different architecture and training parameters. The set of neural networks is trained using actual weld data available in the literature. The performance of each neural network in this set is defined by two performance measures of interest, namely training error and generalization error. Instead of using one of the best networks from this set of trained networks, a method of combining the outputs of all the network from the set is proposed and is called the combined output (or output of the combined network). It is shown that the performance measures of interest obtained using this combined output is better than the performance measures of interest obtained by all the individual neural networks in the set.     

基于灰色 Elman 神经网络的爆破飞石距离预测研究

针对单一神经网络预测方法存在一些不足,将建立灰色关联分析法与 Elman 神经网络的耦合模型,对爆破飞石最大飞散距离进行预测研究。首先,利用灰色关联分析方法对数据进行预处理,确定各影响因素与爆破飞石距离之间的关联度;然后,根据关联度的大小,选择关联度较大的影响因素作为 Elman 神经网络的输入层数据;最后,用神经网络的功能对数据进行训练和预测。研究结果表明：利用灰色关联分析方法确定主要影响因素作为输入层,比单一使用 Elman 神经网络的预测精度更高,达到95％以上。     

Improved neural network model for reverse engineering

While conventional engineering transforms engineering concepts into real parts, in reverse engineering real parts are transformed into engineering models. The construction of a surface from three-dimensional (3D) measuring data points is an important problem in reverse engineering. This paper presents a reconstruction method for the sculptured surfaces from the 3D measuring data points. The surface reconstruction scheme is presented based on a neural network. The reconstruction of the existing surfaces is realized by training the network. A series of measuring points from existing sculptured surfaces is used as a training set. Once the neural network has been trained, it serves as a geometric model to generate all the points that are needed. However, the learning rate for the neural network is relatively slow, and the learning accuracy is often unacceptably low. In this paper, to improve the performance of the neural network, a pre-processor is proposed before the input layer. The pre-processor maps the input into the larger space by generating a set of linearly independent values. The effect of the pre-processor is to increase modelling accuracy, and reduce learning time. Based on this method, experimental results are given to show that the reconstructed surfaces are faithful to the original data points. The proposed scheme is useful for regular or irregular digitized data.