融合声纹信息的能量谱图在鸟类识别中的研究

常用的梅尔倒谱系数结合高斯混合模型(MFCC+GMM)方法的鸟鸣声识别技术难适应噪声环境,模型难以收敛,且计算复杂度高。该文提出一种融合声纹信息的能量谱图的鸟类识别方法(VPS-BR),该方法利用鸟类鸣声在能量谱图上所表现的多维差异性,定量识别鸣声声纹特征。通过对分贝能量进行颜色映射得到能量谱图,提取其视觉特征所表达的声学特征,分析归纳得到鸟类特有鸣声模式。在特征提取步骤中,选用识别速度快的局部二值模式、识别鲁棒性高的方向梯度直方图两个参数表征鸟鸣声谱图的边缘声纹;在识别步骤中,用局部二值模式和方向梯度直方图两种特征分别与支持向量机、K最近邻和随机森林3种分类器算法进行两两组合构建识别模型测试。对15种原始带噪鸟类鸣声数据集进行交叉验证,VPS-BR模型的平均识别率比MFCC+GMM组合模型高出11.3%,方向梯度直方图特征与K最近邻分类器的组合模型识别率达90.5%,表现出较好的抗噪性能和识别性能。最后针对样本数据集缺乏问题,使用生成对抗网络进行图像增强,进一步将识别率提升1.48%。     

一种区分性互补系统构造与融合的语音识别方法

在区分性训练的框架下,提出了一种基于混淆信息加权的互补系统构造方法。首先通过统计音素对的混淆信息,利用混淆信息给音素对加以不同的惩罚权重,分别以基线系统中的3个最优识别结果作为参考,计算混淆信息加权后的音素准确率,同时以正确的标注为参考计算标准的音素准确率。然后通过同时最大化混淆信息加权后的音素准确率和最小化标准音素准确率,构建模型层互补系统,并进一步通过结合RDLT (region-dependent linear transform)特征变换过程构造特征层的互补系统。实验结果表明,与互补最小音素错误准则相比,融合模型层互补系统后识别率提高了0.76%,同时融合特征层和模型层的互补系统识别率提高了1.35%。本方法可以增大互补系统间的差异性,提高系统融合后的识别性能。      

电击死与死后电击的心脏组织光谱模式识别

在司法鉴定领域,涉及电击死亡的案件多发,鉴别死者是生前还是死后受到电击仍是法医病理学鉴定的难点问题之一。为此通过傅里叶变换红外光谱融合机器学习模型对心脏组织视角下的电击死和死后电击两种情况开展分类识别研究。将30只大鼠进行电击死、死后电击和对照处理,通过光谱仪扫描得到其心脏组织光谱,采用竞争性自适应重加权算法共提取到70个光谱特征波长,建立随机森林模型对特征波长提取前后的心脏组织光谱进行模式识别;结果表明,特征波长提取前后模型分类识别的准确率分别为34.9%和73.7%,验证了特征波长提取方法的有效性和必要性。同时建立偏最小二乘模型、传统支持向量机以及粒子群算法和灰狼算法优化的支持向量机模型进行分类识别,结果表明,模型分类识别的准确率分别为61.07%、 34.48%、 100%和98.46%,对比发现经特征提取后的粒子群优化支持向量机模型分类识别效果最好。为排除“生物学死亡期”的干扰,又取60只大鼠按同种方式对其处理,每组又分死后0.5 h和死后1 h 2个亚组,再次通过傅里叶变换红外光谱仪扫描得到光谱数据,数据预处理后将其与之前得到的数据进行一并处理并结合粒子群优化支持向量机模型分...     

基于Mel频谱值和深度学习网络的鸟声识别算法*

为了增强网络对鸟鸣声信号的特征学习能力并提高识别精度,提出一种基于深度残差收缩网络和扩张卷积的鸟声识别方法。首先,提取鸟鸣声信号的对数梅尔特征及其一阶和二阶差分系数组成logMel特征集作为网络模型的输入;其次,通过深度残差收缩网络自动学习噪声阈值,减少噪声干扰;然后,引入扩张卷积增大卷积核感受野并利用注意力机制使网络更关注关键帧特征;最后,通过双向长短时记忆网络从学到的局部特征中学习长期依赖关系。以百鸟数据birdsdata鸟声库中的19种中国常见鸟类作为实验对象,识别正确率可以达到96.58%,并对比模型在不同信噪比数据下的识别结果,结果表明该模型在噪声环境下的识别效果优于现有模型。     

基于加权特征融合的多尺度遥感影像分类

针对普通的三维卷积神经网络（3D CNN）从一个尺度上提取特征,会丢失部分细节信息,且对小样本任务表现一般的问题,本文提出了一种三支路的3D CNN,从不同尺度上提取特征后进行加权特征融合,从而获取了更为全面的特征;并引入数据增强技术,从而改善了小样本情形下的分类性能。现有特征融合方法通常对各个支路直接进行拼接,本文采用加权拼接的特征融合方法,将各特征分别乘以一个加权系数后再进行拼接,该系数通过模拟退火算法求取。本文方法在公开数据集Indian Pines,Pavia University,Salinas等上采用10%的数据进行训练,分别得到了98.60%、99.83%、99.97%的总体准确率,与各类对比方法相比,提升了高光谱遥感影像分类问题的准确率。     

圆形阵列无线传感器的鸟鸣声检测方法*

针对海岛湿地自然保护区鸟类等发声动物监测困难的问题,开发了基于物联网的圆形阵列无线传感器实时在线监测系统。鉴于保护区现场环境条件的特殊性,为保障传感器系统在野外长期稳定运行,需减少无鸣声数据的传输量来减轻系统负担,该文提出了基于子带能量谱熵比的动态双门限鸟类鸣声端点检测算法,对监测数据的有效鸣声段检测后实时回传到云平台。经过仿真和野外实测数据对比了不同信噪比下的鸣声检测性能,结果表明在暴风雨、风浪等海岛自然环境噪声下,鸟鸣声端点检测算法的准确率达90%以上,基本可实现海岛环境中的有效鸣声端点检测,提升了圆形阵列无线传感器系统的可靠性。     

磁瓦表面缺陷的机器视觉检测方法

为了提高磁瓦表面缺陷在线检测准确率并降低检测时间,提出一种基于机器视觉的检测方法。离线训练时,对经过Gabor小波处理后的子图进行融合并提取纹理特征,使用改进的Relief算法提取与类别相关性强的特征子集,并去除冗余特征。为降低缺陷磁瓦的漏检率,先进行偏向性分类处理,再采用最小二乘支持向量机进行分类预测。实验表明,整体预测准确率96.89%,缺陷磁瓦分类准确率达到99.09%,且在线预测时间减少了近1/4,只需67.4ms。     

偏振光谱多维信息的重构融合算法

针对传统光学手段难以实现复杂背景下光谱伪装目标的准确识别,同时,常规的数据融合方法易导致图像信息丢失的缺点,提出了一种基于非下采样轮廓波变换的偏振光谱多维信息融合方法。该方法结合自研的新型偏振光谱多维信息探测仪器,根据其获取的目标空间、光谱、偏振等多维信息,设计了多维信息重构算法流程,提取了偏振态基础数据斯托克斯参量以及偏振度和偏振角,利用NSCT对基础偏振参量进行图像融合,提升图像的信息含量以提高伪装物的识别准确率。先对具有相同边缘信息的图像Q和U采用NSCT分解,低通子带取均值,高通子带取最大值进行初步融合,获得偏振特征S,最后对偏振特征S、强度图像I以及偏振度DoLP进行NSCT分解,对分解所得低通子带进行区域能量加权融合;对高通子带,根据偏振特征图像具有灰度值小,受光照影响大等特点,采用LBP特征进行加权融合。同时,本方法与四类融合方法进行对比,据信息熵、标准差、平均梯度、对比度以及峰值信噪比五项指标对融合结果进行客观评价,并结合普通图像,偏振融合图像,偏振高光谱图像对目标识别精度进行对比。融合后的图像信息熵为6.998 6,标准差为45.599 8,平均梯度为19.808 6...     

基于多模态数据特征融合的舰船识别算法研究

在复杂海域场景下如何综合利用舰船监测的多模态数据进行高效特征提取和特征融合，以此来综合提升舰船识别精度仍存在巨大挑战。针对海域环境中舰船单一数据源识别准确率问题，提出一种有效的多模态数据特征提取和特征融合的舰船识别算法，然后基于深度残差网络模型进行特征融合以提升舰船识别准确率。通过实验结果对比，相比于其他算法基于多模态数据的舰船识别算法平均准确率提升约18%，有效地提升了舰船识别准确率，对相关船舶领域的研发工作具有借鉴意义。     

基于自适应无参经验小波变换和选择集成分类模型的运动想象

运动想象模式识别率的提高对脑机接口(BCI)技术的应用具有重要意义,本文采用自适应无参经验小波变换(APEWT)和选择集成分类模型相结合的方法提高脑电(EEG)信号的分类识别准确率.首先,通过APEWT将EEG信号分解成不同的模态;然后,使用最优模态重构后的信号计算其能量谱(ES)特征,使用最优模态分量计算其边际谱(MS)特征;最后,将不同时间段的ES特征和不同频段的MS特征输入到构建的选择集成分类模型中,从而得到其分类结果,并将该方法与其他4种组合方法进行比较.实验结果表明,本文方法具有较好分类准确率和实时性,其平均分类正确率高于其他4种方法,同时较近期使用相同数据的文献也有优势.本文为在线运动想象类BCI的应用提供了新的方法和思路.     

Aeroacoustic performance evaluation of axial flow fans based on the unsteady pressure field on the blade surface

Because of the European and global regulation concerning acoustic emission, the goal of manufacturers is to substantially decrease the noise radiated by turbomachines, and in particular axial fans, without degrading their aerodynamic performances. High rotation speed and increasing power add to the overall difficulties. The theoretical study of this paper consists of two parts: (1) an aerodynamic approach based on the vortex surface method and (2) an aeroacoustic approach which mainly concerns the prediction of the tonal noise using the Ffowcs Williams and Hawkings (FW-H) equation. One of the main goals is the evaluation of the unsteady aerodynamic forces applied on the fan blades. A 2D software analysis, based on the vortex surface method (or potential flow method), was carried out. That process enabled evaluation of the potential flow around a mobile grid; first in a steady mode, and secondly in an unsteady mode by introducing an upstream disturbance in the form of an inlet velocity variation. The sources of noise corresponding to the zones with high force fluctuation amplitude are located initially on the blade surface. These fluctuating forces are used to predict the tonal noise radiated by the fan in far field by means of the FW-H equation. Two axial flow fans were used in this study. The theoretical results will be compared to the experimental ones in order to evaluate the aeroacoustic performances of the fans.     

Application of frequency-domain linearized Euler solutions to the prediction of aft fan tones and comparison with experimental measurements on model scale turbofan exhaust nozzles

Although it is widely accepted that aircraft noise needs to be further reduced, there is an equally important, on-going requirement to accurately predict the strengths of all the different aircraft noise sources, not only to ensure that a new aircraft is certifiable and can meet the ever more stringent local airport noise rules but also to prioritize and apply appropriate noise source reduction technologies at the design stage. As the bypass ratio of aircraft engines is increased - in order to reduce fuel consumption, emissions and jet mixing noise - the fan noise that radiates from the bypass exhaust nozzle is becoming one of the loudest engine sources, despite the large areas of acoustically absorptive treatment in the bypass duct. This paper addresses this ‘aft fan’ noise source, in particular the prediction of the propagation of fan noise through the bypass exhaust nozzle/jet exhaust flow and radiation out to the far-field observer. The proposed prediction method is equally applicable to fan tone and fan broadband noise (and also turbine and core noise) but here the method is validated with measured test data using simulated fan tones. The measured data had been previously acquired on two model scale turbofan engine exhausts with bypass and heated core flows typical of those found in a modern high bypass engine, but under static conditions (i.e. no flight simulation). The prediction method is based on frequency-domain solutions of the linearized Euler equations in conjunction with perfectly matched layer equations at the inlet and far-field boundaries using high-order finite differences. The discrete system of equations is inverted by the parallel sparse solver MUMPS. Far-field predictions are carried out by integrating Kirchhoff's formula in frequency domain. In addition to the acoustic modes excited and radiated, some non-acoustic waves within the cold stream-ambient shear layer are also captured by the computations at some flow and excitation frequencies. By extracting phase speed information from the near-field pressure solution, these non-acoustic waves are shown to be convective Kelvin-Helmholtz instability waves. Strouhal numbers computed along the shear layer, based on the local momentum thickness also confirm this in accordance with Michalke's instability criterion for incompressible round jets with a similar shear layer profile. Comparisons of the computed far-field results with the measured acoustic data reveal that, in general, the solver predicts the peak sound levels well when the farfield is dominated by the in-duct target mode (the target mode being the one specified to the in-duct mode generator). Calculations also show that the agreement can be considerably improved when the non-target modes are also included, despite their low in-duct levels. This is due to the fact that each duct mode has its own distinct directionality and a non-target low level mode may become dominant at angles where the higher-level target mode is directionally weak. The overall agreement between the computations and experiment strongly suggests that, at least for the range of mean flows and acoustic conditions considered, the physical aeroacoustic radiation processes are fully captured through the frequency-domain solutions to the linearized Euler equations and hence this could form the basis of a reliable aircraft noise prediction method.     

Fan interaction noise reduction using a wake generator: experiments and computational aeroacoustics

A control grid (wake generator) aimed at reducing rotor-stator interaction modes in fan engines when mounted upstream of the rotor has been studied here. This device complements other active noise control systems currently proposed. The compressor model of the instrumented ONERA CERF-rig is used to simulate suitable conditions. The design of the grid is drafted out using semi-empirical models for wake and potential flow, and experimentally achieved. Cylindrical rods are able to generate a spinning mode of the same order and similar level as the interaction mode. Mounting the rods on a rotating ring allows for adjusting the phase of the control mode so that an 8 dB sound pressure level (SPL) reduction at the blade passing frequency is achieved when the two modes are out of phase. Experimental results are assessed by a numerical approach using computational fluid dynamics (CFD). A Reynolds averaged Navier-Stokes 2-D solver, developed at ONERA, is used to provide the unsteady force components on blades and vanes required for acoustics. The loading noise source term of the Ffowcs Williams and Hawkings equation is used to model the interaction noise between the sources, and an original coupling to a boundary element method (BEM) code is realized to take account of the inlet geometry effects on acoustic in-duct propagation. Calculations using the classical analytical the Green function of an infinite annular duct are also addressed. Simple formulations written in the frequency domain and expanded into modes are addressed and used to compute an in-duct interaction mode and to compare with the noise reduction obtained during the tests. A fairly good agreement between predicted and measured SPL is found when the inlet geometry effects are part of the solution (by coupling with the BEM). Furthermore, computed aerodynamic penalties due to the rods are found to be negligible. These results partly validate the computation chain and highlight the potential of the wake generator system proposed.     

上下并联轴流风机室外机气动声学预测模型

本文基于CFD模拟方法,分析了空调器室外机上下并联轴流风机系统噪声源分布,建立了室外机气动声学预测方法.研究发现,上下并联轴流风机系统由宽频和离散频率噪声组成,宽频噪声是影响室外机噪声总声压级的重要因素.涡声分析表明,涡脱落噪声是宽频噪声的主要影响因素.基于CFD的叶片尾缘涡脱落噪声预测方法计算得到宽频声压误差为2 dB,考虑离散频率影响时,室外机A计权总声压级预测误差小于2 dBA.基于CFD的点源时域预测模型,捕捉到了上下并联轴流风机系统的离散频率噪声峰值,且在上下叶轮前二阶谐波处预测值与实验值吻合较好.     

On the prediction of “buzz-saw” noise in acoustically lined aero-engine inlet ducts

Aero-engines operating with supersonic fan tip speeds generate an acoustic signature containing energy spread over a range of harmonics of the engine shaft rotation frequency. These harmonics are commonly known as the “buzz-saw” tones. The pressure signature attached to a supersonic ducted fan will be a sawtooth waveform. The non-linear propagation of a high-amplitude irregular sawtooth upstream inside the inlet duct redistributes the energy amongst the buzz-saw tones. In most modern aero-engines the inlet duct contains an acoustic lining, whose properties will be dependent on the mode number and frequency of the sound, and the speed of the oncoming flow. Such effects may not easily be incorporated into a time-domain approach; hence the non-linear propagation of an irregular sawtooth is calculated in the frequency domain, which enables liner damping to be included in the numerical model. Results are presented comparing noise predictions in hard-walled and acoustically lined inlet ducts. These show the effect of an acoustic liner on the buzz-saw tones. These predictions compare favourably with previous experimental measurements of liner insertion loss (at blade passing frequency), and provide a plausible explanation for the observed reduction in this insertion loss at high fan operating speeds.     

Experimental investigation of metal foam for controlling centrifugal fan noise

This paper presents an experimental investigation on the metal foam for controlling a centrifugal fan noise. Nine samples of metal foam with different types of cells, i.e., open, semi-open and close, are employed to compare their effects on the aerodynamic performance and noise level of the centrifugal fan. Experimental data confirms that the open cell metal foam is the most effective to control the fan noise because it not only significantly suppresses the tonal noise but also attenuates the broadband noise. Moreover, the geometrical parameters of the open cell metal foam, i.e., pores per inch and porosity, are studied to investigate their effects on the aerodynamic performance and noise level of the centrifugal fan.     

Application of theoretical modeling to multichannel active control of cooling fan noise

Multichannel active control has been applied to the global reduction of tonal noise from a cooling fan. In order to achieve consistent far-field attenuation of multiple harmonics of the blade passage frequency (BPF) of the fan, an analytical model has been applied to the control system in order to determine appropriate transducer configurations. The results of the modeling show that the additional global reduction possible by locating acoustically compact secondary sources coplanar with a compact primary source rapidly lessens as the number of symmetrically placed sources is increased beyond three. Furthermore, the model suggests that there are locations in the extreme near field of the sources that can be considered ideal for the minimization of far-field radiated power. Experiments carried out show that a four-channel control system is more effective than a two-channel system at achieving far-field attenuations, especially at the higher harmonics of the BPF for the fan tested. In addition, greater far-field mean-square pressure attenuations are achieved with the error microphones located along the calculated ideal regions than for nonideal placement.     

A study of active tonal noise control for a small axial flow fan

Sound radiated by a computer cooling fan consists of tones which are phase locked with the rotation, and other less deterministic tones and broadband random noise. This paper demonstrates the feasibility of globally eliminating the rotation-locked tones by applying a very simple destructive interference to a modified cooling fan with the number of struts equal to the number of rotor blades. The rig consists of a miniature electret microphone used as a rotation sensor, an ordinary loudspeaker, and a bandpass filter with adjustable amplitude and phase delay. The microphone is located at the inlet bellmouth of the fan to pick up the fluctuating aerodynamic pressure caused by the passing rotor blades. The pressure spectrum is rich in the blade passing frequency (BPF) and its low-order harmonics. It provides much better performance than a pulse-generating tachometer. Analysis of the original fan noise shows that about 90% of the radiated tonal sound is phase locked with rotation, and this portion is almost completely eliminated in all directions. The reductions of the radiated sound power in the first two BPFs are 18.5 and 13.0 dB, respectively, and the overall sound power reduction is 11.0 dB.     

An investigation into the noise generation of CPAP devices: Experimental and numerical study

Continuous positive airway pressure (CPAP) devices are popularly used for obstructive sleep apnoea (OSA) treatment. However, the noise level emission from these devices has been identified as a potential factor for patient’s discomfort and rejection. There is a need to obtain information on the noise characteristics and source locations in order to tackle the most serious noise source within these devices. A typical CPAP device was used for the investigation and its sound characteristics and sound power levels were determined. The noise generated from a centrifugal fan was also independently investigated to address its contribution to the overall noise of the device. Frequency analysis suggested that the noise generated from both the CPAP device and the fan is broadband in nature with discrete peaks containing rotational and non-rotational components. The broadband components were then studied in detail using numerical simulation approach. Computational aeroacoustics (CAA) method with hybrid approach was used to a three-dimensional (3-D) CPAP fluid model to predict the aerodynamic and aeroacoustics behaviours of the device. This showed a complicated flow structure involving flow separation, rotation, and vortices in several locations which resulted in high level of flow turbulence inside the device. The turbulent components were used to estimate the broadband noise level at source using the broadband noise source (BNS) models. It shows the most critical location is at the fan region and at the fan inlet.     

高速动车组气动噪声试验与仿真分析*

通过风洞试验对某高速动车组整车、受电弓及转向架远场气动噪声特性进行分析。试验结果表明,高速动车组远场气动噪声是一宽频噪声,总声能随速度的6.6次方增加;由受电弓引起的远场气动噪声主要集中在中高频,噪声峰值频率随速度变化线性增加;由转向架引起的远场气动噪声主要集中在中低频,噪声峰值频率与速度无关。在此基础上,通过大涡模拟和声扰动方程获得该高速动车组近场噪声。高速动车组远场噪声测点仿真结果与试验结果的最大差值2.2 dB(A),最大相对误差2.5%,表明仿真模型的准确性。仿真结果表明,车头近场噪声以车头鼻尖为界,底部气动噪声能量大于上部流线型气动噪声能量,其中转向架舱位置噪声能量最大,因此进行车内外降噪方案设计时,应重点关注车头转向架舱位置。