Back propagation genetic and recurrent neural network applications in modelling and analysis of squeeze casting process

Today, in competitive manufacturing environment reducing casting defects with improved mechanical properties is of industrial relevance. This led the present work to deal with developing the input-output relationship in squeeze casting process utilizing the neural network based forward and reverse mapping. Forward mapping is aimed to predict the casting quality (such as density, hardness and secondary dendrite arm spacing) for the known combination of casting variables (that is, squeeze pressure, pressure duration, die and pouring temperature). Conversely, attempt is also made to determine the appropriate set of casting variables for the required casting quality (that is, reverse mapping). Forward and reverse mapping tasks are carried out utilizing back propagation, recurrent and genetic algorithm tuned neural networks. Parameter study has been conducted to adjust and optimize the neural network parameters utilizing the batch mode of training. Since, batch mode of training requires huge data, the training data is generated artificially using response equations. Furthermore, neural network prediction performances are compared among themselves (reverse mapping) and with those of statistical regression models (forward mapping) with the help of test cases. The results shown all developed neural network models in both forward and reverse mappings are capable of making effective predictions. The results obtained will help the foundry personnel to automate and précised control of squeeze casting process.     

Fabrication at wafer level of miniaturized gas sensors based on SnO2 nanorods deposited by PECVD and gas sensing characteristics

SnO₂ nanorods were successfully deposited on 3″ Si/SiO₂ wafers by inductively coupled plasma-enhanced chemical vapour deposition (PECVD) and a wafer-level patterning of nanorods layer for miniaturized solid state gas sensor fabrication were performed. Uniform needle-shaped SnO₂ nanorods in situ grown were obtained under catalyst- and high temperature treatment-free growth condition. These nanorods have an average diameter between 5 and 15 nm and a length of 160-300 nm. The SnO₂-nanorods based gas sensors were tested towards NH₃ and CH₃OH and gas sensing tests show remarkable response, showing promising and repeatable results compared with the SnO₂ thin films gas sensors.     

The application of an artificial immune system-based back-propagation neural network with feature selection to an RFID positioning system

This study uses the Received Signal Strength Indication (RSSI) values of RFID to predict the position of picking staff for warehouse management. A proposed feature selection-based back-propagation (BP) neural network that uses an artificial immune system (AIS) (FSBP-AIS) to determine the connecting weights of a neural network learns the relationship between the RSSI values and the position of the picking staff. In addition, the proposed FSBP-AIS is able to determine the representative features, or inputs, during training. Once a picking staff's position is known, this information is used to plan the picking route for picking staff if a new order arrives. The computational results indicate that the proposed FSBP-AIS can provide better predictions than a traditional BP neural network, BP neural network with stepwise regression to determine the important inputs, and regression method.     

VLSI implementation of transcendental function hyperbolic tangent for deep neural network accelerators

Extensive use of neural network applications prompted researchers to customize a design to speed up their computation based on ASIC implementation. The choice of activation function (AF) in a neural network is an essential requirement. Accurate design architecture of an AF in a digital network faces various challenges as these AF require more hardware resources because of its non-linear nature. This paper proposed an efficient approximation scheme for hyperbolic tangent (tanh) function which purely based on combinational design architecture. The approximation is based on mathematical analysis by considering maximum allowable error in a neural network. The results prove that the proposed combinational design of an AF is efficient in terms of area, power and delay with negligible accuracy loss on MNIST and CIFAR-10 benchmark datasets. Post synthesis results show that the proposed design area is reduced by 66% and delay is reduced by nearly 16% compared to state-of-the-art.     

基于人工神经网络的动态过程聚类研究

为了有效解决传统人工神经网络对于时变函数的聚类问题,以及提高在大样本下网络的学习和泛化能力,提出了基于离散余弦变换的传统人工神经网络动态过程聚类方法。通过离散余弦变换将样本函数降维映射到由对应余弦参数所张成的模式特征空间,满足了传统人工神经网络对输入样本的要求,使传统人工神经网络实现动态过程的聚类成为可能。给出了实现算法,分析了计算复杂度,并使用基本竞争型人工神经网络对特征样本向量进行聚类,实验结果表明该方法是正确、有效的。与过程人工神经网络相比,该方法具有运算简单、物理意义明确等优点。     

A neural network-based approach for dynamic quality prediction in a plastic injection molding process

This paper presents an innovative neural network-based quality prediction system for a plastic injection molding process. A self-organizing map plus a back-propagation neural network (SOM-BPNN) model is proposed for creating a dynamic quality predictor. Three SOM-based dynamic extraction parameters with six manufacturing process parameters and one level of product quality were dedicated to training and testing the proposed system. In addition, Taguchi’s parameter design method was also applied to enhance the neural network performance. For comparison, an additional back-propagation neural network (BPNN) model was constructed for which six process parameters were used for training and testing. The training and testing data for the two models respectively consisted of 120 and 40 samples. Experimental results showed that such a SOM-BPNN-based model can accurately predict the product quality (weight) and can likely be used for various practical applications.     

基于模糊聚类神经网络的化工过程预报模型研究

本文提出了在化工过程预报中的模糊聚类神经网络模型,该模型具有提取典型数据、优化模糊规则及优化参数的优点,在化工过程预报实验中与传统方法相比预报结果的精度提高,计算时间缩短。     

基于神经网络的多媒体通信控制机制研究

在Internet网络的多媒体通信系统中需要解决QoS控制问题,如视频、音频等多媒体数据的同步、网络拥塞控制、多媒体数据传输的QoS协商控制、视频平滑,以及连续多媒体系统的CPU调度等.为了解决好这些控制问题,提出一种基于神经网络的多媒体通信控制机制,把人工智能与多媒体通信技术紧密结合起来,并在Internet网络环境下开发了实用的多媒体通信系统.运行结果表明,该系统效果优越.     

Hybrid deep neural network model for human action recognition

In this paper, we propose a hybrid deep neural network model for recognizing human actions in videos. A hybrid deep neural network model is designed by the fusion of homogeneous convolutional neural network (CNN) classifiers. The ensemble of classifiers is built by diversifying the input features and varying the initialization of the weights of the neural network. The convolutional neural network classifiers are trained to output a value of one, for the predicted class and a zero, for all the other classes. The outputs of the trained classifiers are considered as confidence value for prediction so that the predicted class will have a confidence value of approximately 1 and the rest of the classes will have a confidence value of approximately 0. The fusion function is computed as the maximum value of the outputs across all classifiers, to pick the correct class label during fusion. The effectiveness of the proposed approach is demonstrated on UCF50 dataset resulting in a high recognition accuracy of 99.68%.     

卷积神经网络的FPGA并行加速方案设计

根据卷积神经网络的特点,提出了深度流水的FPGA加速方案,设计了卷积层的通用卷积电路。该卷积电路可以在一个时钟周期内获得一个计算结果。理论上,该方案对于MNIST数据集,在28×28个时钟周期内可以获得一幅图片的运算结果。针对网络训练过程的前向传播阶段,在网络结构和数据集相同的情况下,对GPU,FPGA,CPU进行了在计算效率和能耗之间的比较。其中在计算效率方面,50 MHz频率的FPGA就可以相较于GPU实现近5倍的加速,相较于12核的CPU实现8倍的加速。而在功耗方面,该FPGA的实现方案只有GPU版本的26.7%。     

选择性神经网络二次集成方法在定量构效关系建模中的应用研究

针对在建立定量构效关系(QSAR)模型中,单个人工神经网络模型难以确定参数,容易产生“过拟合”;一般神经网络集成模型虽然建立过程简单,但由于个体差异度小而导致泛化能力相对单个神经网络没有明显改善等问题,提出了一种基于随机梯度法的选择性神经网络二次集成方法。在建立除草剂(苯乙酰胺类化合物)的QSAR模型的实验研究中表明,该方法设计过程简单,能够以较小的运算代价明显地提高了模型的泛化能力,是建立QSAR模型的一个有效方法。     

A feasible intrusion detector for recognizing IIS attacks based on neural networks

Most activities on the Internet can be recorded as log files of websites and website administrators can inspect log files to locate problems after any network intrusion occurs. However, since log files usually contain a huge quantity of data, without effective methods, it is generally not feasible for administrators to determine the concealed meanings within log files. One method for dealing with this issue is to use neural networks; this is an effective means to distinguish and classify abnormal data in log files, thus alleviating the administrator's burden. This paper presents the results of a study on intrusion detection on IIS (Internet information services) utilizing a hybrid intrusion detection system (IDS). The feasibility of the hybrid IDS is validated based on the Internet scanner system (ISS). In the intrusion detection system proposed, we used four different training data sets: 200, 800, 1400, and 2000. The system is trained either by Taguchi's experimental design or full factorial experimental design under different training data sets; the former can save much more time than the latter. Under Taguchi's experimental design, the best results are obtained when the training data set is of size 1400; overall accuracy in this case is 97.5%. On the contrary, for the full factorial experimental design, the best results are reached when the training data set is of size 2000; overall accuracy is 97.6%. Our study indicates that when to retrain the detector and how much time to allow for this training fully depend on the downgrade percentage of the detection rate, which determines the size of the retraining data set. To reduce the void time for updating the detector, the downgrade percentage should be restricted.     

基于遗传算法的BP神经网络技术的应用

针对BP网络的不足,提出了基于遗传算法的神经网络技术。将两者有机的融合在一起,充分利用了GA算法的全局搜索能力和BP算法的局部搜索能力,加快了收敛速度,提高了收敛精度,将其应用于高速公路动态称重系统的神经网络控制器的训练中,取得了较好的效果。     

Simulation of an adaptive artificial neural network for power system security enhancement including control action

This paper presents a new method for enhancing power system security, including a remedial action, using an artificial neural network (ANN) technique. The deregulation of electricity markets is still an essential requirement of modern power systems, which require the operation of an independent system driven by economic considerations. Power flow and contingency analyses usually take a few seconds to suggest a control action. Such delay could result in issues that affect system security. This study aims to find a significant control action that alleviates the bus voltage violation of a power system and to develop an automatic data knowledge generation method for the adaptive ANN. The developed method is proved to be a steady-state security assessment tool for supplying possible control actions to mitigate an insecure situation resulting from credible contingency. The proposed algorithm is successfully tested on the IEEE 9-bus and 39-bus test systems. A comparison of the results of the proposed algorithm with those of other conventional methods reveals that an ANN can accurately and instantaneously provide the required amounts of generation re-dispatch and load shedding in megawatts.     

SAR图像目标识别的卷积神经网模型

目的 合成孔径雷达图像目标识别可以有效提高合成孔径雷达数据的利用效率。针对合成孔径雷达图像目标识别滤波处理耗时长、识别精度不高的问题,本文提出一种卷积神经网络模型应用于合成孔径雷达图像目标识别。方法 首先,针对合成孔径雷达图像特点设计特征提取部分的网络结构;其次,代价函数中引入L2范数提高模型的抗噪性能和泛化性;再次,全连接层使用Dropout减小网络的运算量并提高泛化性;最后研究了滤波对于网络模型的收敛速度和准确率的影响。结果 实验使用美国运动和静止目标获取与识别数据库,10类目标识别的实验结果表明改进后的卷积神经网络整体识别率（包含变体）由93.76%提升至98.10%。通过设置4组对比实验说明网络结构的改进和优化的有效性。卷积神经网络噪声抑制实验验证了卷积神经网络的特征提取过程对于SAR图像相干斑噪声有抑制作用,可以省去耗时的滤波处理。结论 本文提出的卷积神经网络模型提高了网络的准确率、泛化性,无需耗时的滤波处理,是一种合成孔径雷达图像目标识别的有效方法。     

贝叶斯正则化神经网络预测金属晶体结合能的研究

采用贝叶斯正则化神经网络(BRNN)对61种金属晶体结合能进行了预测。对网络结构、训练集、预测集以及学习次数进行了优化,并用独立预测样本对贝叶斯正则化神经网络作了检验。预测结果表明,在推广能力方面,贝叶斯正则化神经网络优于熟知的反向传播(BP)神经网络和多元线性回归方法(MLR)。它可望成为元素和化合物构效关系研究的辅助手段。     

GPU加速的神经网络BP算法*

近年来图形处理器(GPU)快速拓展的可编程性能力加上渲染流水线的高速度及并行性,使得图形处理器通用计算(GPGPU)迅速成为一个研究热点。针对大规模神经网络BP算法效率低下问题,提出了一种GPU加速的神经网络BP算法。将BP网络的前向计算、反向学习转换为GPU纹理的渲染过程,从而利用GPU强大的浮点运算能力和高度并行的计算特性对BP算法进行求解。实验结果表明,在保证求解结果准确度不变的情况下,该方法运行效率有明显的提高。     

基于并列卷积神经网络的超分辨率重建

为提取更多有效特征并提高模型训练的收敛速度,提出一种基于并列卷积神经网络的超分辨率重建方法。该网络由两路不同结构的网络组成：一路为简单的残差网络,其优化残差映射比原始的映射更容易实现;另一路为增加了非线性映射的卷积神经网络,增强了网络的非线性能力。随着并行网络结构的复杂化,收敛速度慢成为突出问题。针对这个问题,在卷积层后添加正则化处理,以简化模型参数、增强特征拟合能力,最终达到加快收敛的目的。实验结果表明,与基于深度卷积神经网络算法相比,该网络结构收敛速度更快,主观视觉效果更好,峰值信噪比（PSNR）平均提高了0.2 dB。     

卷积神经网络在心拍类识别中的应用

心电图（ECG）心拍分类对心脏疾病的临床诊断具有重要意义,但是ECG四类心拍间数据不平衡问题严重制约着心拍分类性能的提升。针对这一问题,以卷积神经网络（CNN）为基础,首先在组合四类心拍等量数据基础上构建用于表达噪声及四类心拍间共性信息的通用CNN模型,接着以通用CNN模型为基础分别在四类心拍数据上构建四个更为有效表达对应心拍类别倾向性信息的类别CNN模型,最后综合四个类别CNN模型的输出判别心拍类型。在MIT-BIH心电图数据库上的实验结果显示,该方法的平均灵敏度为99.68%、平均阳性检测率是98.58%、综合指标是99.12%,显著优于二级联合聚类法在MIT-BIH心电图数据库上的分类性能。     

多尺度卷积神经网络显著物体检测

目的 传统显著性检测模型大多利用手工选择的中低层特征和先验信息进行物体检测,其准确率和召回率较低,随着深度卷积神经网络的兴起,显著性检测得以快速发展。然而,现有显著性方法仍存在共性缺点,难以在复杂图像中均匀地突显整个物体的明确边界和内部区域,主要原因是缺乏足够且丰富的特征用于检测。方法 在VGG（visual geometry group）模型的基础上进行改进,去掉最后的全连接层,采用跳层连接的方式用于像素级别的显著性预测,可以有效结合来自卷积神经网络不同卷积层的多尺度信息。此外,它能够在数据驱动的框架中结合高级语义信息和低层细节信息。为了有效地保留物体边界和内部区域的统一,采用全连接的条件随机场（conditional random field,CRF）模型对得到的显著性特征图进行调整。结果 本文在6个广泛使用的公开数据集DUT-OMRON（Dalian University of Technology and OMRON Corporation）、ECSSD（extended complex scene saliency dataset）、SED2（segmentation evalution database 2）、HKU、PASCAL-S和SOD（salient objects dataset）上进行了测试,并就准确率—召回率（precision-recall,PR）曲线、F测度值（F-measure）、最大F测度值、加权F测度值和均方误差（mean absolute error,MAE）等性能评估指标与14种最先进且具有代表性的方法进行比较。结果显示,本文方法在6个数据集上的F测度值分别为0.696、0.876、0.797、0.868、0.772和0.785;最大F测度值分别为0.747、0.899、0.859、0.889、0.814和0.833;加权F测度值分别为0.656、0.854、0.772、0.844、0.732和0.762;MAE值分别为0.074、0.061、0.093、0.049、0.099和0.124。无论是前景和背景颜色相似的图像集,还是多物体的复杂图像集,本文方法的各项性能均接近最新研究成果,且优于大多数具有代表性的方法。结论 本文方法对各种场景的图像显著性检测都具有较强的鲁棒性,同时可以使显著性物体的边界和内部区域更均匀,检测结果更准确。