人工神经网络在建筑声学中的应用

本文对人工神经网络的历史作了简单回顾，并对它在建筑声学领域中的应用状况作了阐述，同时提出了有待进一步探讨的若干问题。     

The investigation of model selection criteria in artificial neural networks by the Taguchi method

Artificial neural networks (ANNs) have been successfully used for solving variety of problems. One major disadvantage of ANNs is that there is no formal systematic model building approach. This paper presents the application of the Taguchi method in the optimization of the design parameters of the ANNs. The performances of the ANNs were determined by the Taguchi method considering factors relevant for ANNs’ performance. The properties affecting the performance of the ANNs and their levels on the peak analytical function were determined by performing computational experiments. After training the network, the values of the statistical data criteria were determined and the optimum parameter levels were obtained in terms of the performance statistics. The performance of ANNs is shown to be better in the case of the application of the Taguchi method rather than in the case of random choice of factor values.     

Application of artificial neural networks for the determination of proteins with CPA-pI by rayleigh light scattering technique

The determination of proteins with 2-(4-chloro-2-phosphonophenylazo)-7-(4-iodophenylazo)-1,8-dihydroxynaphthalene-3,6-disulfonic acid (CPA-pI) by Rayleigh light scattering (RLS) was studied in this paper. The weak RLS of CPA-pI and BSA can be enhanced greatly by the addition of Al³⁺ at the pH 5.6 and an enhanced RLS signal was produced at 365-385 nm. Based on the reaction of CPA-pI, Al³⁺ and proteins, a new quantitative determination method for proteins has been developed. The effect of three variables for the determination of proteins was optimized by means of artificial neural networks (ANNs) using extended delta-bar-delta (EDBD) algorithms with the optimal network structure of 3-5-1. This method is very sensitive (2.5-35.4 μg/ml for bovine serum albumin (BSA)), rapid (<2 min), simple (one step) and tolerance of most interfering substances. Six samples of protein in human serum were determined and the maximum relative error is no more than 2% and the recovery is between 95% and 105%.     

Estimating the sound absorption coefficients of perforated wooden panels by using artificial neural networks

Developing efficient sound absorption materials is a relevant topic for large scale structures such as gymnasiums, shopping malls, airports and stations. This study employs artificial neural network (ANN) algorithm to estimate the sound absorption coefficients of different perforated wooden panels with various setting combinations including perforation percentage, backing material and thickness. The training data sets are built by carrying out a series of experimental measurements in the reverberation room to evaluate the sound absorption characteristics of perforated wooden panels. A multiple linear regression (MLR) model is also developed for making comparisons with ANN. The analytical results indicate that the ANN exhibits satisfactory reliability of a correlation between estimation and truly measured absorption coefficients of approximately 0.85. However, MLR cannot be applied to nonlinear cases. ANN is a useful and reliable tool for estimating sound absorption coefficients estimation.     

神经元传输函数对人工神经网络动力学特性的影响

以神经元局域场分布为基础，重新研究了连续神经元传输函数对具有联想记忆的人工神经网络功能的影响.与以往的认识不同的是，研究发现连续传输函数与硬极限传输函数相比并不存在明显的优越性，相反，连续传输函数对网络的某些功能，如最大存储率具有负面影响.研究表明神经网络的特性主要决定于网络的动力学结构(具体体现为网络吸引子对应的神经元局域场分布)，网络的动力学结构可以通过选择合适的设计规则进行有效控制，不同的传输函数虽然也能影响到网络的动力学结构，但是它所带来的影响是被动的，可控性很差. 关键词： 联想记忆 神经网络 吸引子 局域场分布     

���˹�������Ԥ��HL-2A��������ŵ�����

使用人工神经网络(ANN)对HL-2A装置破裂放电进行了离线预测研究。采用了两种方法训练网络,一种方法是采用原始实验数据作为网络输入训练网络,另一种是把训练样本中的Mirnov原始实验信号进行预处理,目的是突出Mirnov原始信号隐含的破裂信息。比较这两种方法,结果表明第二种方法获得的网络对破裂放电能够做出更加准确的预测。     

Prediction of airblast-overpressure induced by blasting using a hybrid artificial neural network and particle swarm optimization

Blasting is an inseparable part of the rock fragmentation process in hard rock mining. As an adverse and undesirable effect of blasting on surrounding areas, airblast-overpressure (AOp) is constantly considered by blast designers. AOp may impact the human and structures in adjacent to blasting area. Consequently, many attempts have been made to establish empirical correlations to predict and subsequently control the AOp. However, current correlations only investigate a few influential parameters, whereas there are many parameters in producing AOp. As a powerful function approximations, artificial neural networks (ANNs) can be utilized to simulate AOp. This paper presents a new approach based on hybrid ANN and particle swarm optimization (PSO) algorithm to predict AOp in quarry blasting. For this purpose, AOp and influential parameters were recorded from 62 blast operations in four granite quarry sites in Malaysia. Several models were trained and tested using collected data to determine the optimum model in which each model involved nine inputs, including the most influential parameters on AOp. In addition, two series of site factors were obtained using the power regression analyses. Findings show that presented PSO-based ANN model performs well in predicting the AOp. Hence, to compare the prediction performance of the PSO-based ANN model, the AOp was predicted using the current and proposed formulas. The training correlation coefficient equals to 0.94 suggests that the PSO-based ANN model outperforms the other predictive models.     

变参数混沌时间序列的神经网络预测研究

研究一类复杂变参数混沌系统时间序列的预测问题.首先构造一个变参数Logistic映射,分析变参数混沌系统的特点,指出动力学特征不断变化的这类系统不存在恒定形状的吸引子;结合Takens嵌入定理和神经网络理论,阐述神经网络方法预测具有恒定吸引子形状的混沌系统可行的原因,分析研究其用于预测变参数混沌系统的潜在问题.变参数Ikeda系统的神经网络预测试验验证了理论分析结果,试验还表明,简单增大预测训练样本数可能降低泛化预测精度,训练集的选择对这类系统的泛化预测效果影响极大,指出混沌时间序列预测实用化必须研究解决这类变参数混沌系统的预测. 关键词： 混沌 预测 神经网络 变参数系统     

用人工神经网络预测HL-2A等离子体放电破裂

使用人工神经网络(ANN)对HL-2A装置破裂放电进行了离线预测研究。采用了两种方法训练网络,一种方法是采用原始实验数据作为网络输入训练网络,另一种是把训练样本中的Mirnov原始实验信号进行预处理,目的是突出Mirnov原始信号隐含的破裂信息。比较这两种方法,结果表明第二种方法获得的网络对破裂放电能够做出更加准确的预测。     

Particle identification using artificial neural networks at BESⅢ

A multilayered perceptrons' neural network technique has been applied in the particle identification at BESⅢ. The networks are trained in each sub-detector level. The NN output of sub-detectors can be sent to a sequential network or be constructed as PDFs for a likelihood. Good muon-ID, electron-ID and hadron-ID are obtained from the networks by using the simulated Monte Carlo samples.     

Particle sizing by multiangle light-scattering data using the high-order neural networks

The problem of retrieval of size and refractive index of a spherical particle by angular dependence of scattered light in scanning flow cytometry is considered. For its solution, the high-order neural networks are used. We restricted the range of angles available for measurement from 10° to 60°. The retrieval errors of characteristics of nonabsorbing particles were investigated at the ranges of the radius and relative refractive index 0.6–10.6 μm, and 1.02–1.38, respectively.     

Prediction of concrete strength using ultrasonic pulse velocity and artificial neural networks

Ultrasonic pulse velocity technique is one of the most popular non-destructive techniques used in the assessment of concrete properties. However, it is very difficult to accurately evaluate the concrete compressive strength with this method since the ultrasonic pulse velocity values are affected by a number of factors, which do not necessarily influence the concrete compressive strength in the same way or to the same extent. This paper deals with the analysis of such factors on the velocity-strength relationship. The relationship between ultrasonic pulse velocity, static and dynamic Young’s modulus and shear modulus was also analyzed. The influence of aggregate, initial concrete temperature, type of cement, environmental temperature, and w/c ratio was determined by our own experiments. Based on the experimental results, a numerical model was established within the Matlab programming environment. The multi-layer feed-forward neural network was used for this purpose. The paper demonstrates that artificial neural networks can be successfully used in modelling the velocity-strength relationship. This model enables us to easily and reliably estimate the compressive strength of concrete by using only the ultrasonic pulse velocity value and some mix parameters of concrete.     

Study of particle tracking algorithms based on neural networks for stereoscopic tracking velocimetry

Stereoscopic-tracking velocimetry can offer an excellent potential for continuously monitoring three-dimensional flow fields for all three-component velocities in near-real-time. Requiring only the deployment of two solid-state cameras with directional freedom in test-section illumination and observation, the system can be built to be compact and robust. For flow velocimetry measurement, increasing the number density in particle seeding is much desirable for maximizing spatial resolution, that is, number of velocity data points of the captured field. The challenge, however, is how to successfully track numerous crisscrossing particles with speed and reliability. In our approach, the task of particle tracking is converted to an optimization problem for which neural networks are applied. Here we present the design and development of the neural networks for particle tracking as well as the investigation results based on both computer simulations and real experiments. The approach appears to be appropriate for near-real-time velocity monitoring, being able to provide reliable solutions achieved by the massive parallel-processing power of the neural networks.     

Effect of damage in neural networks

The effect of damage on the pattern recognition in the Hopfield-model of neural networks is studied. It is assumed that in a damaged network the synaptic efficaciesJ _i,j=J_j,i, between pairs of neuronsS _i andS _j follow the Hebb rule with probability (1–p) and are equal to zero with probabilityp. Numerical simulations are performed for a net consisting of 400 neurons. It is investigated in detail how the retrieval of noisy patterns and the storage capacity of the net depends, for varying initial noise, on the concentrationp of the damaged synaptic efficacies.     

Optimization of the operational parameters in a fast axial flow CW CO2 laser using artificial neural networks and genetic algorithms

This paper presents an artificial intelligence approach for optimization of the operational parameters such as gas pressure ratio and discharge current in a fast-axial-flow CW CO₂ laser by coupling artificial neural networks and genetic algorithm. First, a series of experiments were used as the learning data for artificial neural networks. The best-trained network was connected to genetic algorithm as a fitness function to find the optimum parameters. After the optimization, the calculated laser power increases by 33% and the measured value increases by 21% in an experiment as compared to a non-optimized case.     

多光谱成像技术诊断植物病虫害的人工神经网络模型

为了实现可靠的植物病虫害诊断,提出把人工神经网络和多光谱成像技术结合的方法,并将该方法用于常见的三种黄瓜病害的识别研究。在此基础上,实验采用窄带多光谱成像技术获取患病黄瓜叶面的14个可见光通道和近红外通道、全色通道的多光谱图像。利用BP网络对病斑样本的光谱信息进行学习分类。和14通道训练结果比较,增加850nm的近红外通道和全色通道,使网络的训练时间缩短、预测能力提高。实验结果表明,这种方法对植物进行快速、准确和非破坏性诊断提供可靠的技术支持。     

时滞离散神经网络的同步控制

利用既有效又便于实施的时滞状态反馈控制器,根据所给定的条件构造相应的不等式,研究了带有时滞的离散神经网络模型的同步控制问题,给出了该离散系统指数同步的充分条件.在设计同步控制的时候,没有假设激励函数的有界性、可微性和单调性,给出的条件简便易实施.数值结果进一步证明了该控制方法的有效性. 关键词： 离散神经网络 时滞 同步     

Evaluation of thermal conductivity of MgO-MWCNTs/EG hybrid nanofluids based on experimental data by selecting optimal artificial neural networks

In this paper, the thermal conductivity ratio of MgO-MWCNTs/EG hybrid nanofluids has been predicted by an optimal artificial neural network at solid volume fractions of 0.05%, 0.1%, 0.15%, 0.2%, 0.4% and 0.6% in the temperature range of 25–50 °C. In this way, at the first, thirty six experimental data was presented to determine the thermal conductivity ratio of the hybrid nanofluid. Then, four optimal artificial neural networks with 6, 8, 10 and 12 neurons in hidden layer were designed to predict the thermal conductivity ratio of the nanofluid. The comparison between four optimal ANN results and experimental showed that the ANN with 12 neurons in hidden layer was the best model. Moreover, the results obtained from the best ANN indicated the maximum deviation margin of 0.8%.     

Parameter estimation of continuous variable quantum key distribution system via artificial neural networks

Continuous-variable quantum key distribution(CVQKD)allows legitimate parties to extract and exchange secret keys.However,the tradeoff between the secret key rate and the accuracy of parameter estimation still around the present CVQKD system.In this paper,we suggest an approach for parameter estimation of the CVQKD system via artificial neural networks(ANN),which can be merged in post-processing with less additional devices.The ANN-based training scheme,enables key prediction without exposing any raw key.Experimental results show that the error between the predicted values and the true ones is in a reasonable range.The CVQKD system can be improved in terms of the secret key rate and the parameter estimation,which involves less additional devices than the traditional CVQKD system.     

Overview of hybrid optical neural networks

This paper reviews optical architectures for the implementation of hybrid neural networks. Optics is mainly applied to implementing the matrix-vector or tensor-matrix multiplication. In addition, the general background of neural networks as well as a brief discussion on holographic associative memory are also given.