基于神经网络的贮仓结构参数识别

利用神经网络技术，提出了识别结构物理参数的一种方法。用单元刚度矩阵基本值和模态应变能来选择基本模态，用修正的Latin超立方采样技术和模态准入准则来产生网络的输入数据。贮仓在动载作用下的自振频率和模态作为网络的输入，子矩阵参与系数作为网络的输出，用Levenberg-Marquardt算法训练网络。仿真计算表明，方法是可行的。     

Photon-limited image classification with a feedforward neural network

An application of neural networks to the classification of photon-limited images is reported. A three-level feedforward network architecture is employed in which the input units of the network correspond to the pixels of a two-dimensional image. The network is trained in a minicomputer by the use of the backpropagation technique. The statistics of the network components are analyzed, resulting in a method by which the probability of correct classification of a given input image can be calculated. Photon-limited images of printed characters are obtained with a photon-counting camera and are classified. The experimental results are in excellent agreement with theoretical predictions.     

Evaluation of neural network models with generalized sensitivity analysis

A sensitivity analysis method for discovering characteristic features of the input data using neural network classification models has been devised. The sensitivity is the gradient of the neural network model response function, and because neural network models are nonlinear, the gradient depends on the point where it is evaluated. Two criteria are used for measuring the sensitivity. The first criterion calculates the sensitivity or gradient of the neural network output with respect to the average of the objects that comprise each class. The second criterion measures the average sensitivity of the class objects. The sensitivity analysis was applied to temperature-constrained cascade correlation network models and evaluated with sets of synthetic data and experimental mobility spectra. The neural network models were built using temperature-constrained cascade correlation networks (TCCCNs). A weight constraint was devised for the output units of the network models. This method implements weight decay with conjugate gradient training and yields more sensitive neural network models. Temperature-constrained hidden units furnish more sensitive network models than networks without constraints. By comparing the sensitivities of the class mean input and the mean sensitivity for all the inputs of a class, the individual input variables may be assessed for linearity. If these two sensitivities for an input variable differ by a constant factor, then that variable is modeled by a simple linear relationship. If the two sensitivities vary by a nonconstant scale factor, then the variable is modeled by higher order functions in the network. The sensitivity method was used to diagnose errors in the training data, and the test for linearity indicated a TCCCN architecture that had better predictability.     

一种粗模糊神经分类器

介绍一种新的粗集编码模糊神经分类器。基于粗集理论的概念，讨论了知识编码、属性简化、分类系统简化的方法；并利用模糊隶属度函数将输入精确信息映射为模糊变量信息，解决分类中病态定义的数据问题和提高系统非线性映射的分类能力；提出了结合系统参数的重要性因子的网络的模糊推理方法和粗模糊神经分类器的网络结构以及有导师的最小平方误差学习训练算法。实现的粗集编码模糊神经分类器具有网络结构空间维数低、学习算法简单、网络训练时间短、非线性特性丰富等优点。     

用BP神经网络诊断结构破损

提出了一种基于BP神经网络的结构破损诊断方法,该方法以结构残余力向量作为破损诊断的网络输入。对网络训练样本采用广义空间格点法进行了变换,从而较好地解决了由于系统响应样本在数据空间分布不均对网络收敛速度及网络诊断精度的影响问题。应用实例表明,本文方法能准确诊断结构破损位置与严重程度,是一种有效的结构破损诊断方法。     

基于卷积神经网络与门控循环单元的气液两相流流型识别方法

提出了一种基于卷积神经网络(CNN)与门控循环单元(GRU)的垂直管道气液两相流流型识别方法。该方法基于电阻层析成像(ERT)系统的重建图像,对其填充处理后进行离散余弦变换(DCT),求取最大、最小 DCT 系数的差值,选取一定帧数长度数据作为网络输入,对流型进行识别。分析了输入序列长度对CNN-GRU、CNN 及 GRU 网络分类准确的影响,确定了最佳输入向量维度分别为 60、65 及 50,使用实验数据对3种网络进行训练、测试,结果表明,CNN-GRU网络分类准确率最高,平均流型识别准确率可达 99.40%。     

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.     

On-line identification of control chart patterns using self-organizing approaches

Effective identification of unnatural control chart patterns (CCPs) is an important issue in statistical process control, as unnatural CCPs can be associated with specific assignable causes adversely affecting the process. The intention of this paper is to develop an automatic CCP identification system using self-organizing approaches—neural network and decision tree (DT) learning. Recently, back-propagation networks (BPNs) have been widely used in the research field of CCP identification. However, one of the major limitations of conventional BPN is in dealing with dynamic patterns that vary over time, such as CCPs. This limitation is one of the major reasons for the false classification problem commonly encountered in the BPN-based CCP identification schemes in the literature. A time-lagging input algorithm is proposed in this research to enhance the performances of the BPN-based CCP identifiers. Additionally, DT learning is employed as a novel approach to the CCP identification problem. The simulation experiments demonstrate that both the BPN-based system with time-lagging input and the DT-based system perform better than the conventional BPN-based system in terms of identification accuracy and speed. The proposed time-lagging input algorithm can greatly improve the identification speed and stability of the BPN-based CCP identifier. Besides, the empirical comparison indicates that the DT-based system outperforms the BPN-based system with respect to classification capability in an on-line CCP identification scheme. Moreover, the learning time of the DT-based system is much shorter than that of the BPN-based system.     

Feature selection and combination criteria for improving accuracy in protein structure prediction

The classification of protein structures is essential for their function determination in bioinformatics. At present, a reasonably high rate of prediction accuracy has been achieved in classifying proteins into four classes in the SCOP database according to their primary amino acid sequences. However, for further classification into fine-grained folding categories, especially when the number of possible folding patterns as those defined in the SCOP database is large, it is still quite a challenge. In our previous work, we have proposed a two-level classification strategy called hierarchical learning architecture (HLA) using neural networks and two indirect coding features to differentiate proteins according to their classes and folding patterns, which achieved an accuracy rate of 65.5%. In this paper, we use a combinatorial fusion technique to facilitate feature selection and combination for improving predictive accuracy in protein structure classification. When applying various criteria in combinatorial fusion to the protein fold prediction approach using neural networks with HLA and the radial basis function network (RBFN), the resulting classification has an overall prediction accuracy rate of 87% for four classes and 69.6% for 27 folding categories. These rates are significantly higher than the accuracy rate of 56.5% previously obtained by Ding and Dubchak. Our results demonstrate that data fusion is a viable method for feature selection and combination in the prediction and classification of protein structure.     

基于BP神经网络的盲均衡器设计

为构建用于某语音信号传输系统盲均衡器的BP神经网络模型,编写了基于BP神经网络的盲均衡算法伪代码,计算了算法的时间复杂度,分析了BP神经网络输入层神经元个数、隐含层神经元个数和隐含层层数对盲均衡算法性能的影响,评估了基于Sigmoid的变步长算法、基于误差补偿的变步长算法和基于误差的变步长算法对基于BP神经网络的盲均衡器性能的改进效率,据此设计了一种含双隐层结构的BP神经网络盲均衡器,并对其性能进行了数值仿真分析,明确了其适用范围,为该语音信号传输系统设计提供了技术支撑。     

Intelligent sensors using neural networks: the example of amicrosystem for visual inspection

The aim of this article is to show how artificial neural networks and 3D packaging technology have a major role to play in the development of microsystems. A visual inspection system for real-time identification of objects in a scene is described. The system comprises a CMOS or CCD imager, an analogue preprocessing stage that includes a learning mechanism for adapting the system to images of different contrast, and a neural classification stage. The detection of a matrix code using as the classifier a vector support machine is illustrated. As the latter is difficult to realise in VLSI the author has turned to the threshold neural network `Offset', which constructs a parity machine, i.e. a network comprising a single layer of neurons, the output being obtained with the help of a simple exclusive-OR logic gate. Unfortunately the parity machine suffers from overtraining, as the OffSet algorithm converges to a zero error over the entire training base. Nevertheless, if good implementation strategies are available, it is possible to improve the performance in general by combining a large number of classifiers by majority voting. A CMOS VLSI circuit, called SysNeuro, has been fabricated which integrates a parity machine in a square systolic architecture of 4×4 processors. This circuit has variable precision. The number of neurons has been increased by combining 4 SysNeuro chips in a multichip module and stacking three of the modules to form a 3D structure-SysNeuro3D     

Control chart pattern recognition using learning vector quantization networks

Pattern recognition systems using neural networks for discriminating between different types of control chart patterns are discussed. A class of pattern recognizers based on the Learning Vector Quantization (LVQ) network is described. A procedure to increase the classification accuracy and decrease the learning time for LVQ networks is presented. The results of control chart pattern recognition experiments using both existing LVQ networks and an LVQ network implementing the proposed procedure are given.     

基于卷积神经网络和Transformer网络的鸟声识别

针对传统鸟声识别算法中特征提取方式单一、分类识别准确率低等问题,提出一种结合卷积神经网络和Transformer网络的鸟声识别方法。该方法综合考虑网络局部特征学习和全局上下文依赖性构造,从原始鸟声音频信号中提取短时傅里叶变换(Short Time Fourier Transform,STFT)语谱图特征,将其输入到卷积神经网络(ConvolutionalNeural Network,CNN)中提取局部频谱特征信息,同时提取鸟声信号的对数梅尔特征及一阶差分、二阶差分特征用于合成梅尔频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC)混合特征向量,将其输入到Transformer网络中获取全局序列特征信息,最后融合所提取的特征可得到更丰富的鸟声特征参数,通过Softmax分类器得到鸟声识别结果。在Birdsdata和xeno-canto鸟声数据集上进行实验,平均识别准确率分别达到了97.81%和89.47%。实验结果表明该方法相较于其他现有的鸟声识别模型具有更高的识别准确率。     

Determination of closed form solution for acceptance sampling using ANN.

Tabled sampling schemes such as MIL-STD-105D offer limited flexibility to quality control engineers in designing sampling plans to meet specific needs. We describe a closed form solution to determine the AQL indexed single sampling plan using an artificial neural network (ANN). To determine the sample size and the acceptance number, feed-forward neural networks with sigmoid neural function are trained by a back propagation algorithm for normal, tightened, and reduced inspections. From these trained ANNs, the relevant weight and bias values are obtained. The closed form solutions to determine the sampling plans are obtained using these values. Numerical examples are provided for using these closed form solutions to determine sampling plans for normal, tightened, and reduced inspections. The proposed method does not involve table look-ups or complex calculations. Sampling plan can be determined by using this method, for any required acceptable quality level and lot size. Suggestions are provided to duplicate this idea for applying to other standard sampling table schemes.     

Neural Networks in Applied Statistics

Artificial neural networks are computer algorithms or computer programs derived in part from attempts to model the activity of nerve cells. They have been applied to pattern recognition, classification, and optimization problems in the physical and chemical sciences, as well as in other fields. We introduce the principles of the multilayer feedforward network that is among the most commonly used neural networks in practical problems. The relevance of neural network models for the applied statistician is considered using a time series prediction problem as an example. The multilayer feedforward neural network uses a nonlinear function of the predictors to obtain predictions for future time series values. We illustrate the considerations involved in specifying a neural network model and evaluate the accuracy of neural network models relative to the accuracy obtained using other computer-intensive, nonmodel-based techniques.     

Failure diagnosis of a servovalve by neural networks with new learning algorithm and structure analysis

This paper considers a failure diagnosis of a pneumatic servovalve used in automated production systems. The valve is monitored by an accelerometer. Six parameters characterizing the vibration data are extracted, and fed into neural networks to solve four types of diagnosis problems. A conjugate gradient followed by a variable metric method is demonstrated as an effective learning algorithm. Neural network structures are analysed through Boolean expressions summarizing network simulation results for given learning patterns. The neural networks are found to utilize majority voting mechanisms. Irrelevant neurons can be identified and removed without degrading the diagnosis performance.     

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.     

Evolutionary neural network modeling for software cumulative failure time prediction

An evolutionary neural network modeling approach for software cumulative failure time prediction based on multiple-delayed-input single-output architecture is proposed. Genetic algorithm is used to globally optimize the number of the delayed input neurons and the number of neurons in the hidden layer of the neural network architecture. Modification of Levenberg–Marquardt algorithm with Bayesian regularization is used to improve the ability to predict software cumulative failure time. The performance of our proposed approach has been compared using real-time control and flight dynamic application data sets. Numerical results show that both the goodness-of-fit and the next-step-predictability of our proposed approach have greater accuracy in predicting software cumulative failure time compared to existing approaches.     

Bidirectional multiple‐valued neural network for pattern recognition and associative recall

A new neural network algorithm based on the counter‐propagation network (CPN) architecture, named MVL‐CPN, is proposed in this paper for bidirectional mapping and recognition of multiple‐valued patterns. The MVL‐CPN is capable of performing a mathematical mapping of a set of multiple‐valued vector pairs by self‐organization. The use of MVL‐CPN reduces considerably the number of nodes required for the input layers as well as the number of synaptic weights compared to the binary CPN. The training of the network is stable because all synaptic weights are monotonically nonincreasing. The bidirectional mapping and associative recall features of the MVL‐CPN are tested by using various sets of quaternary patterns. It is observed that the MVL‐CPN can converge within three or four iterations. The high‐speed convergence characteristics of the network can lead to the possibility of using this architecture for real‐time applications. An important advantage of the proposed type of neural network is that it can be implemented in VLSI with reduced number of neurons and synaptic weights when compared to a larger binary network needed for the same application. © 2000 John Wiley & Sons, Inc. Int J Imaging Syst Technol 11, 125–129, 2000     

IWO与BPNN混合智能算法在 转子故障诊断中的应用

在经典IWO杂草算法的基础上提出一种适用于神经网络优化的新算法。该算法将多种结构的神经网络权值阈值编码为不同维度的杂草种子，以神经网络均方误差作为种子适应度的统一评价标准，同时对多个维度的杂草种子进行排序筛选，实现了神经网络权值阈值与结构同时优化的目的。应用该方法于转子系统故障分类问题，实验结果表明该方法可以在结合BP算法优势的同时有效优化神经网络各参数，可以得到分类精度高、结构最简且泛化能力强的神经网络故障分类器。