基于深度神经网络的语种识别*

语音段的有效表示方法存在易混淆语种和短时语音段识别率较低等问题,为满足不同时长和方言的识别要求,提出基于深度神经网络不同层的有效语音段表示方法.采用含有中间瓶颈层的深层神经网络作为前端特征提取,综合利用该网络的输出层和中间瓶颈层输出结果,得到不同形式的语音段表示并用于语种识别.在美国国家标准技术局语种识别评测2009年和2011年阿拉伯方言数据集上验证了方法的有效性.     

Adoption of combined detection technology of tumor markers via deep learning algorithm in diagnosis and prognosis of gallbladder carcinoma

This study was to explore the application value of back propagation (BP) neural network (BPNN) and genetic algorithm (GA) in the combined detection and prognosis of tumor markers in patients with gallbladder cancer. 446 patients with gallbladder cancer were included in the experimental group, 279 patients with benign gallbladder disease were included in the control group, and 188 healthy people were selected and included in the blank group. Serum tumor markers (CA242, CA199, CEA, and CA125) of the three groups were detected by electrochemical luminescent immune analyzer, and follow-up data for 5 years after surgery were collected. Based on BPNN and GA, an optimization algorithm for multi-tumor markers was constructed and applied to the combined detection of tumor markers in patients. The artificial neural network (ANN), dynamic network biomarker (DNB), auxiliary diagnosis algorithm of the support vector machine (SVM) based on the particle swarm optimization (PSO) (PSO-SVM), matched-pairs feature selection (MPFS) based on the machine learning, and the BPNN were introduced to compare with the algorithm constructed. The diagnostic performances of the algorithms were evaluated with the fivefold cross-validation method. The results showed that the levels of CanAg (CA) 242, carcinoma embryonic antigen (CEA), CA199, and CA125 and positive rates in the experimental group were significantly higher than those in the control group and the blank group (P?<?0.05); but the differences between control group and blank group were not visible (P?>?0.05). The sensitivity (91.72%) and specificity (87.49%) in detecting CA242 and CA199 based on the proposed algorithm were the highest; the sensitivity (0.9186), specificity (0.8622), and accuracy (94.94%) of the proposed algorithm were higher than those of the conventional algorithms. The postoperative follow-up survival rate of patients in the experimental group was reduced from 41.72% in the first year to 4.28% in the fifth year; tumor node metastasis (TNM) stage IV, neck gallbladder cancer, and CA199 were significantly correlated with the survival rate of patients in the experimental group (P?<?0.05). In summary, the combined detection technology of multiple tumor markers based on deep learning algorithms showed excellent diagnostic and prognostic performance for gallbladder cancer. The occurrence of gallbladder cancer was related to the tumor markers CA242, CA199, CEA, and CA125, showing better detection effects by combination of CA242 and CA199. The TNM stage IV, neck gallbladder cancer, and CA199 were independent risk factors for the decrease in survival rate of patients with gallbladder cancer.

     

Controllers of Ziegler-Nichols type for unstable process with time delay

Proportional (P), proportional plus integral (PI) and proportional plus integral plus derivative (PID) controllers are designed for a time-delayed process having latency L and a single right half-plane pole at s=λ, with λL < 1. Explicit designs are derived from the concept of phase margin, and an optimum stability approach for PI controllers is based on a parameter plane study. An optimum gain margin design is also given for P control.     

Short-term changes in upper extremity dynamic mechanical response parameters following power hand tool use

Dynamic mechanical response parameters (stiffness, damping and effective mass), physiological properties (strength and swelling) and symptoms of the upper limb were measured before power tool operation, immediately following and 24?h after power tool operation. Tool factors, including peak torque (3 Nm and 9 Nm) and torque build-up time (50?ms and 250?ms), were controlled in a full factorial design. Twenty-nine inexperienced power hand tool users were randomly assigned to one of four conditions and operated a pistol grip nutrunner four times per min for 1?h in the laboratory. Isometric strength decreased immediately following tool use (15%) (p?<?0.01) and 24?h later (9%) (p?<?0.05). Mechanical parameters of stiffness (p?<?0.05) and effective mass (p?<?0.05) were affected by build-up time. An average decrease in stiffness (43%) and effective mass (57%) of the upper limb was observed immediately following pistol grip nutrunner operation for the long (250?ms) build-up time. A previously developed biomechanical model was used to estimate handle force and displacement associated with the tool factors in the experiment. The conditions associated with the greatest predicted handle force and displacement had the greatest decrease in mechanical stiffness and effective mass, and the greatest increase in localized discomfort.     

基于深度学习的航空发动机传感器故障检测

针对传统反向传播(BP)神经网络和支持向量机(SVM)存在的过拟合、维数灾难、参数选择困难等问题,提出了一种基于深度学习算法的航空发动机传感器故障检测方法.对发动机参数记录仪采集的多维数据进行预处理,建立基于深度置信网络(DBN)的故障检测模型,利用预处理后的数据对检测模型进行训练,经过DBN故障检测模型逐层特征学习实现了传感器故障检测.仿真结果表明:在无人工特征提取和人工特征提取的情况下,基于DBN故障检测的准确率均高于BP神经网络和SVM模型.     

基于机器学习的平行语料库应用效果评价方法

针对传统翻译系统在时态翻译中不准确的问题,结合当前的机器学习算法,提出一种基于DBN的平行语料库时态翻译方法。为实现该方法,首先对时态标注模型和DBN基本理论进行介绍,并提出汉英语句时态翻译的思路;而在进行DBN平行语料库特征提取的过程中,采用自动时态标注算法对时态进行标注,并对得到的数据进行时态树编码;然后以编码数据作为输入,运用DBN网络对时态进行训练预测,得到中文语句中可能的时态;最后通过Transformer翻译器对语句进行翻译,得到对应的翻译时态句子;最后结合LDC等平行语料库中的翻译句子,采用上述方法进行翻译,结果表明提出的时态翻译方法无论是在准确率,还是在召回率等方面,都有很大的优势,说明提出的机器学习算法在平行语料库的翻译中更具有准确性。     

Comparison of two different cooling systems in alleviating thermal and physiological strain during prolonged exercise in the heat

Abstract

This study compared the efficacy of an ice vest comprising of water (WATER) or a water-carbon (CARBON) emulsion on thermophysiological responses to strenuous exercise in the heat. Twelve male cyclists completed three 50-minute constant workload trials (55% of peak power output, ambient temperature 30.4?±?0.6°C) with WATER, CARBON, and without ice vest (CONTROL), respectively. The increase in core body temperature (Tcore) was lower in WATER at 40 (?0.49?±?0.34 °C) and 50?minutes (?0.48?±?0.48 °C) and in CARBON at 30 (?0.41?±?0.48 °C), 40 (?0.54?±?0.51 °C), and 50?minutes (?0.67?±?0.62 °C) as compared to CONTROL (p?<?0.05, ES > 0.8). While heart rate and blood lactate kinetics did not differ between the conditions, statistical main effects in favour of both WATER and CARBON were found for thermal sensation (condition p?<?0.001 and interaction p?<?0.01) and rating of perceived exertion (condition p?<?0.05). Per-cooling with CARBON and WATER similarly reduced Tcore but not physiological strain during prolonged exercise in the heat.

Practitioner Summary: Exercise in the heat is characterised by increases in thermophysiological strain. Both per-cooling with a novel carbon-based and a conventional water-based ice vest were shown to reduce core temperature significantly. However, due to its lower mass, the carbon-based system may be recommended especially for weight-bearing sports.     

Effect of backpack fit on lung function

Carrying loads close to the trunk with a backpack causes a restrictive type of change in lung function in which Forced Vital Capacity (FVC) and Forced Expiratory Volume in 1?s (FEV₁) are reduced without a corresponding decrement in the FEV₁.FVC^???1 %. It is not known whether this is due to the weight of the load acting on the chest or to the tightness of fit of the shoulder and chest straps and waist belt of the pack harness. This study examined FVC, FEV₁, FEV₁.FVC^???1 %, peak expiratory flow (PEF), forced expiratory flow between 0.2 and 1.2?s (FEF_0.2?–?1.2) after the start of expiration and between 25 and 75% of each FVC (FEF_25?–?75%) in 12 healthy males wearing a 15?kg backpack in which the shoulder and chest straps and hip belt were loosened by 3?cm from a ‘comfort fit’ to achieve a ‘loose pack’ fit (LPF) and tightened by 3?cm from CF to achieve a ‘tight pack’ fit (TPF). In comparison with the control condition of no pack, a loose pack fit significantly reduced FVC (by 3.6%, p?<?0.01), FEV₁ (by 4.3%, p?<?0.01) and FEF_25?–?75% (by 8.4%, p?<?0.01). A tight pack fit significantly reduced FVC (by 8.1%, p?<?0.01) and FEV₁ (by 9.1%, p?<?0.001). It also significantly reduced FEF_0.2?–?1.2 (by 7.3%, p?<?0.05) and FEF_25?–?75% (by 21%, p?<?0.01). In comparison with a loose pack fit, the tight pack fit was associated with a significantly lower FVC (by 4.6%, p?<?0.01), FEV₁ (by 5.0%, p?<?0.01), FEF_25?–?75% (by 13.8%, p?<?0.01) and a fall in FEF_0.2?–?1.2 (by 5.5%). The latter was approaching significance (p?=?0.077). There were no significant changes in FEV₁.FVC^???1% and PEF. It is concluded that tightening the fit of a backpack significantly affects lung function in a manner that is typical of a restrictive change in lung function and is very similar in pattern to that of wearing a loosely fitted loaded backpack. The effect of tightness of fit is additional to that due to the weight of the load alone and may also reduce expiratory flow at low lung volumes.     

基于超声造影的颈部淋巴结灌注流向可视化

对于颈部淋巴结的超声造影视频病例, 可利用其时间强度曲线提取灌注特征进行病情诊断. 现有的研究方法对感兴趣区域进行像素级的分析可以更准确地描述灌注特征, 然而很少有深入研究灌注流向可视化的方法. 本文利用了像素级的时间强度曲线TIC分析, 采取双重筛选方式对TIC曲线进行筛选, 进而针对TIC曲线提取二维灌注参数对灌注流向进行可视化. 特征提取后生成的流线图像能够一定程度反映血管的分布, 对医生病情诊断有一定的辅助价值, 也可以对微血管重构有一定的启发价值.     

卷积神经网络和深度置信网络在SAR影像冰水分类的性能评估

目的 海冰分类是海冰监测的主要任务之一。目前基于合成孔径雷达SAR影像的海冰分类方法分为两类：一类是基于海冰物理特性与SAR成像特征等进行分类,这需要一定的专业背景;另一类基于传统的图像特征分类,需要人为设计特征,受限于先验知识。近年来深度学习在图像分类和目标识别方面取得了巨大的成功,为了提高海冰分类精度及海冰分类速度,本文尝试将卷积神经网络（CNN）和深度置信网络（DBN）用于海冰的冰水分类,评估不同类型深度学习模型在SAR影像海冰分类方面的性能及其影响因素。方法 首先根据加拿大海冰服务局（CIS）的冰蛋图构建海冰的冰水数据集;然后设计卷积神经网络和深度置信网络的网络架构;最后评估两种模型在不同训练样本尺寸、不同数据集大小和网络层数、不同冰水比例的测试影像以及不同中值滤波窗口的分类性能。结果 两种模型的总体分类准确率达到93%以上,Kappa系数0.8以上,根据分类结果得到的海冰区域密集度与CIS的冰蛋图海冰密集度数据一致。海冰的训练样本尺寸对分类结果影响显著,而训练集大小以及网络层数的影响较小。在本文的实验条件下,CNN和DBN网络的最佳分类样本尺寸分别是16×16像素和32×32像素。结论 利用CNN和DBN模型对SAR影像海冰冰水分类,并进行性能分析。发现深度学习模型用于SAR影像海冰分类具有潜力,与现有的海冰解译图的制作流程和信息量相比,基于深度学习模型的SAR影像海冰分类可以提供更加详细的海冰地理分布信息,并且减小时间和资源成本。     

基于深度置信网络风电机组变桨系统的故障诊断

针对风电机组变桨系统常见故障,提出一种基于深度置信网络(DBN)的故障诊断方法。设计出基于DBN的变桨系统故障诊断框架;通过堆叠多层受限玻尔兹曼机(RBM),对比重构数据与原始输入数据差异,研究了DBN故障特征自提取能力;将堆叠RBM提取的故障特征输入到顶层分类器中进行训练,得到故障诊断模型;最后采用风场真实故障数据集进行了验证测试。实验结果表明,采用该方法进行风电机组变桨系统故障诊断相比其他方法具有更高的准确率。     

一类基于非线性PCA和深度置信网络的混合分类器及其在PM2.5浓度预测和影响因素诊断中的应用

传统的深度置信网络（Deep brief networks,DBN）在建立高维数据分类模型时,往往存在网络负荷大,运算复杂度高等问题.本文首先基于非线性PCA（NPCA）对高维样本数据进行降维,然后以提取到的非线性特征作为DBN的网络输入,构建了一类含非线性特征提取预处理机制的DBN分类器.并从信息熵理论的角度出发,证明了所提改进DBN分类器在网络结构和算法复杂度方面的优势.通过一个PM2.5浓度预测与影响因素诊断实例,验证了所提改进DBN在一类分类和影响因素诊断问题中的应用,并与传统的分类器进行对比,显示了所提方法在建模精度及收敛速度上的优势.     

The utility of heart rate and minute ventilation as predictors of whole-body metabolic rate during occupational simulations involving load carriage

The utility of cardiac and ventilatory predictors of metabolic rate derived under temperate and heated laboratory conditions was evaluated during three fire-fighting simulations (70-mm hose drag, Hazmat recovery, bushfire hose drag; N = 16 per simulation). The limits of agreement for cardiac (temperate: ? 0.54 to 1.77; heated: ? 1.39 to 0.80 l min^? 1) and ventilatory surrogates (temperate: ? 0.19 to 1.27; heated: ? 0.26 to 1.16 l min^? 1) revealed an over-estimation of oxygen consumption that exceeded the acceptable limits required by occupational physiologists (N = 25; ± 0.24 l min^? 1). Although ventilatory predictions offered superior precision during low-intensity work (P < 0.05), a cardiac prediction was superior during more demanding work (P < 0.05). Deriving those equations under heated conditions failed to improve precision, with the exception of the cardiac surrogate during low-intensity work (P < 0.05). These observations imply that individualised prediction curves are necessary for valid estimations of metabolic demand in the field.     

基于深度置信网络的目标识别方法

针对合成孔径雷达图像预处理鲁棒性不足、特征提取及利用不充分等问题,提出了一种基于深度置信网络的合成孔径雷达(SAR)图像目标自动识别算法。首先提出一种基于双树复小波变换（DT-CWT）的非局部均值图像降斑算法,并结合目标方位角估计实现对原始数据鲁棒的预处理;最后,引入多层深度置信网络提取针对合成孔径雷达目标的深度抽象视觉信息作为特征并完成识别任务。采用3类运动与静止目标的获取与识别(MSTAR)实测数据进行的仿真实验结果表明,所提算法具有较高鲁棒性和识别率。     

深度置信网络在极化SAR图像分类中的应用

目的 深度置信网络能够从数据中自动学习、提取特征,在特征学习方面具有突出优势。极化SAR图像分类中存在海量特征利用率低、特征选取主观性强的问题。为了解决这一问题,提出一种基于深度置信网络的极化SAR图像分类方法。方法 首先进行海量分类特征提取,获得极化类、辐射类、空间类和子孔径类四类特征构成的特征集;然后在特征集基础上选取样本并构建特征矢量,用以输入到深度置信网络模型之中;最后利用深度置信网络的方法对海量分类特征进行逐层学习抽象,获得有效的分类特征进行分类。结果 采用AIRSAR数据进行实验,分类结果精度达到91.06%。通过与经典Wishart监督分类、逻辑回归分类方法对比,表现了深度置信网络方法在特征学习方面的突出优势,验证了方法的适用性。结论 针对极化SAR图像海量特征的选取与利用,提出了一种新的分类方法,为极化SAR图像分类提供了一种新思路,为深度置信网络获得更广泛地应用进行有益的探索和尝试。     

基于BTSM和DBN模型的唇读和视素切分研究

为实现文本/语音驱动的说话人头部动画,本文提出基于贝叶斯切线形状模型的口形轮廓特征提取方法和基于动态贝叶斯网络（Dynamic Bayesian Network, DBN）模型的唇读系统。在描述词与它的组成视素关系的基础上,得到视素时间切分序列。为比较性能,音素DBN模型和HMM的音素识别结果被影射成视素序列。在评价准则上,提出绝对视素切分正确性和基于图像与嘴唇几何特征两种相对视素切分正确性的评价标准。实验表明,DBN模型识别性能优于HMM,而基于视素的DBN模型能为说话人头部动画提供最好的口形。     

Sensitivity of large-scale vegetation greenup and dormancy dates to climate change in the north−south transect of eastern China

Phenology event responses, based on vegetation types, are strong indicators of climate variability and the ability of the vegetation to adapt to future climate changes. However, the sensitivity of phenology events to climate change along either environmental or vegetation type gradients is rarely examined. Phenological curves of major vegetation types along the North?South Transect of Eastern China (NSTEC) have been developed using wavelet and smooth-spline methods based on the normalized difference vegetation index from 1982 to 2006. Spatial?temporal patterns, trends of greenup-onset dates, dormancy dates, and growing season lengths (GSLs) during the period of 1982?2006 are presented.

The greenup-onset dates were most significantly and negatively related to the temperature in cold and humid areas, but insignificantly and positively in semi-arid regions. However, dormancy date showed a positive correlation with temperature. In populations of the same vegetation type, distributed along thermal gradients of NSTEC, the phenology sensitivities to warming were different. Greenup sensitivities of cold temperate coniferous forest (CTCF) and temperate meadow steppe (TMS) increased significantly from??6.0 to 0 days °C^?1 (p < 0.001) and from about??2.0 to 2.0 days °C^?1 (p < 0.001), respectively. In contrast, temperate grass steppe (TGS) and temperate deciduous shrubland (TDS) showed a decreased trend of greenup sensitivity from 2.0 to??4.0 days °C^?1 (p < 0.001) and from 2.0 to??6.0 days °C^?1 (p < 0.001), respectively. For the dormancy date sensitivity, CTCF showed a decreasing trend from about 6.0 to 0 days °C^?1 (p < 0.001), and subtropical evergreen-broadleaved forest (SEBF) decreased from 5.0 to??5.0 days °C^?1 (p < 0.05).     

Effect of carbohydrate supplementation on simulated exercise of rugby league referees

This investigation evaluated the effectiveness of supplementing eight elite rugby league referees with a 6% maltodextrin (Md) solution whilst undertaking a simulated rugby league game. The simulation was based on motion analysis of six rugby league matches. Subjects undertook two trials of repetitive 20-m shuttle activity on an indoor track. During one trial 200 ml of Md was ingested at eight time points and in the other trial a similarly tasting placebo (Pl) was administered. A single-blind, counterbalanced design was employed. The simulation involved subjects performing four, 10-min blocks of shuttle activity before a 10-min break was instigated. Three further 10-min blocks of shuttle activity were also performed before a performance test to volitional exhaustion involving 20-m shuttles at paces varying between 55 and 95% of a pre-determined [Vdot]O_{2 max} was undertaken. Timed 15-m sprints took place during each of the 10-min blocks. The rating of perceived exertion (RPE), and blood glucose and lactate concentrations were also determined throughout. The mean number of shuttles to exhaustion was significantly greater with Md ingestion than with Pl ingestion (57 ± 19 vs. 43 ± 15; p < 0.05), while the mean 15-m sprint times were significantly shorter for the Md than the Pl condition (2.40 ± 0.09 s vs. 2.51 ± 0.14 s; p < 0.01). The mean RPE was ～ 5.2% lower during Md than Pl ingestion, the values being significantly different (Md: 12.07 ± 0.32; Pl: 12.73 ± 0.28; p < 0.01). Maltodextrin ingestion significantly elevated blood glucose levels compared with placebo (F(1,7) = 18.07: p < 0.01), although no significant differences were apparent for blood lactate levels (F(1,7) = 4.39; p > 0.05). These results highlight the beneficial eVects of maltodextrin ingestion on work-rates of rugby league referees in a simulation of a game?s activity. The improvement may be related to higher circulating concentrations of blood glucose.     

Upper quadrant postural changes of school children in response to interaction with different information technologies

The objective of this study was to quantitatively analyse the sitting posture of school children interacting with both old (book) and new (laptop and desktop computers) information technologies to test the hypothesis that posture is effected by the type of information technology (IT) used. A mixed model design was used to test the effect of IT type (within subjects) and age and gender (between subjects). The sitting posture of 32 children aged 4?–?17 years was measured whilst they read from a book, laptop, and desktop computer at a standard school chair and desk. Video images were captured and then digitized to calculate mean angles for head tilt, neck flexion, trunk flexion, and gaze angle. Posture was found to be influenced by IT type (p?<?0.001), age (p?<?0.001) and gender (p?=?0.024) and significantly correlated to the stature of the participants. Measurement of resting posture and the maximal range of motion of the upper and lower cervical spines in the sagittal plane was also undertaken. The biophysical impact and the suitability of the three different information technologies are discussed.