一种基于聚类的门控卷积网络语声分离方法*

基于深度聚类的语音分离方法已被证明能有效地解决混合语音中说话人输出标签排列的问题,然而,现有关于聚类进行说话人分离方法,大多数是优化嵌入使每个源的重建误差最小化。本文以时域卷积网络（ConvTasNet）为基础网络,设计了一种改进基于聚类的门控卷积（Gate-conv Cluster）语音分离方法,在时域上通过堆叠的门控卷积网络,实现端到端深度聚类的源分离。该框架将非线性门控激活用于时域卷积网络中,提取语音信号的深层次特征;同时在高维特征空间中聚类对语音信号的特征进行表示和划分,为恢复不同信号源提供了一个长期的说话者表示信息。该框架解决了说话人输出标签排列问题并对语音信号的长期依赖性进行建模。通过华尔街日报数据集进行实验得出,该方法在SDRi（信源失真比）和Si-SNR（尺度不变信源噪声比）指标上分别达到了16.72 dB和16.33 dB的效果。     

FCM-FastICA的点蚀声发射信号分离识别方法

金属点蚀是一种破坏性和隐患较大的设备损伤形式。点蚀会产生声发射信号。点蚀过程中产生的多种声源类型会造成信号混叠,影响腐蚀进程的判断。针对点蚀信号混叠问题,提出一种模糊C均值聚类与快速独立分量分析算法相结合的点蚀信号分离识别方法。通过分析单、双点蚀声发射数据将点蚀分为钝化膜破裂阶段、点蚀诱导成核及发展阶段,由聚类确定信号类别并用快速独立分量分析分离混合信号,利用相关性函数验证分离效果。结果表明：单点蚀过程存在3类原信号,双点蚀过程存在7类信号,其中包含单个信号与混合信号;单个信号与原信号相关性极高,达到0.8以上,混合信号的分离分量与原信号相关性达到0.6以上,分离效果较好。该方法可对点蚀混合信号进行有效分离和识别,为腐蚀进程判断提供支持。     

分形维数海豚哨声信号检测方法

针对海豚哨声信号自动检测的问题,提出一种基于分形维数的自适应阈值海豚哨声信号检测方法。对待检测声信号计算盒分形维数,根据得到的盒分形维数特征值,通过模糊C均值聚类自适应确定检测阈值,实现海豚哨声信号的自动检测。文中对水池录制的海豚声信号进行了数据分析,利用哨声信号盒分形维数对哨声信号段与非信号段进行检测,并与基于谱熵的方法进行对比,获得了较高的检测率以及较低的虚警率,可以适用于海豚哨声信号的自动检测与分割。     

桡动脉脉搏信号的无创检测与声信号处理

本文从根动与声学角度出发，用一种高性能的检测系统采集了挠动脉脉搏声信号，首次利用倒谱分析技术对两种脉搏信号进行了处理．获得了脉象功率谱和倒谱及其特征参数．分析结果表明：脉象功率谱可准确地判断人的呼吸频率特性；清脉信号在SHZ以内的谱能量约占总能量的83．7呢，而结代脉仅占47．2%；由倒港客观、准确地获得了脉搏周期值，倒谱可定量地给出脉搏强度大小，并可分析心律的稳定性等．本文的研究，为中医脉象的客观化研究和西医的心血管疾病的无创检测提供了一条新的途径．     

基于局部线性嵌入的特征融合方法在岩石破裂 状态分类的应用*

为了获取岩石破裂过程有效的声发射信号特征,更好的对岩石破裂状态进行分类,提出一种基于流形学习算法的LLE特征融合方法进行数据降维。以红砂岩为研究对象设计室内单轴压缩实验采集信号,然后对原始声发射信号预处理并对信号进行特征提取,以时域、频域下的特征向量重新组合成一组新的多维特征向量,采用线性主元(PCA)和流形学习LLE算法分别进行降维。比较两种算法降维后融合特征的聚类效果二维和三维分布图,使用LLE算法降维后,四种状态分布相对更近,呈一条水平线趋势,且各状态交叉混叠数目较少,第一状态没有一个样本错判,且四个状态相比于PCA降维后的聚类效果更集中。再比较两种算法降维后融合特征的敏感度之和,LLE算法融合特征敏感度之和远大于PCA算法,说明经过LLE算法降维后得到的融合特征更多地表征了原始信号包含的局部信息同时证明了LLE算法相比PCA算法具有更好的聚类效果。最后经LLE特征融合下的砂岩破裂状态分类实验验证,融合特征后的识别率相对单一的时域特征识别提高了6%。表明该方法能显著提高岩石破裂状态分类的识别率,降维性能相对突出。     

基于聚类技术的三维舰船模型特征库研究

存贮目标所有的视点图像,建立完备的特征库,或者提取能够抵抗视点变化的不变特征,是三维目标识别的常用方法。这两种方案都存在不足:要么特征库规模庞大,识别过程计算量大,识别效率低;要么难以找到鲁棒的识别特征。结合两种方案研究了基于聚类技术建立三维舰船模型特征库的方法。利用仿射传播聚类方法无需事先指定聚类中心的优点,将其应用于两型舰船模型的视点空间聚类。通过提取视点图像的Hu矩特征,进行了仿真实验,给出了聚类结果的有效性分析。     

贝叶斯优化模糊聚类地级行政区声环境

声环境功能区划是噪声污染管理的重要手段,当前声环境功能区划研究大多是基于某个特定的地级行政区来进行的,难以反映各地级行政区声环境的异同。 该文基于134个地级行政区的人口、面积、各声环境功能区面积和及面积占比,进行地级行政区声环境表征和归一化处理。 以轮廓系数作为聚类有效性评价指标,基于贝叶斯优化模糊聚类方法对地级行政区声环境表征进行聚类分析。 通过本文方法与谱聚类、K-medoids聚类、高斯混合模型聚类的聚类性能对比,验证了该方法的有效性。 结果表明我国地级行政区声环境分为9类,城市规模和用地情况发生显著变化导致其声环境表征和归类发生变化后应重新评估当前噪声污染管理政策,并借鉴同类的地级行政区的噪声污染管理政策做出必要的调整。     

基于深度聚类的目标细粒度分类方法北大核心CSCD

为提高光电系统对弱小目标的识别和分类能力,降低算法对硬件平台和数据的依赖,提出一种无监督分类方法−基于目标深度特征聚类的细粒度分类方法。该方法通过轮廓、颜色、对比度等浅层特征提取提示目标,经超分辨处理后,利用卷积神经网络对目标的深层特征进行编码,进一步采用基于注意机制的主成分分析方法进行降维生成表征矩阵,最后利用聚类的方式实现目标细粒度分类。实验验证了基于不同神经网络的深度聚类方法在不同数据集上的分类性能,其中采用ResNet-34聚类方法在CIFAR-10测试集上细粒度分类性能达92.71%,结果表明,基于深度聚类的目标细粒度方法能够取得与强监督学习方法相当的目标分类效果。此外,还可以根据不同簇数和聚类等级的选择实现不同细粒度的分类效果。     

基于核独立成分分析和聚类中值的水下传感器故障诊断

水下传感器所处环境的开放性以及其长期处于无人值守状态,极易收到环境作用而发生破坏,传统分簇式的分布式故障诊断方法无法有效对其进行故障诊断,因此,提出了一种基于核主成分分析和聚类中值的故障诊断方法。首先,设计了水下传感器节点故障诊断的模型,然后采用核主成分分析方法对节点采集的数据和来自邻居节点发送的数据进行数据降维,得到具有最小属性集的数据集,然后对节点的邻居节点集进行聚类,选出具有最多元素的聚类,并将聚类的中值作为参考数据,将各节点与其对应的参考数据进行比较从而确定节点是否发生故障,最后定义了基于核独立成分分析和聚类中值的传感器节点故障诊断算法。仿真实验证明了文中方法能有效地对水下传感器进行故障诊断,且与其它方法相比,具有误差检测率高和误检率低的优点。     

采用双谱多类小波包特征的雷达信号聚类分选

针对现有雷达脉内特征参量对噪音敏感,难以适应复杂体制雷达聚类分选的问题,应用双谱多类小波包特征实现了宽信噪比条件下未知复杂雷达信号的高准确率聚类分选.首先对接收到的雷达信号求得双谱归一化系数,然后利用多类小波包提取双谱归一化系数的特征参量,并选取最佳小波包特征作为分选参量,最后通过提取8类雷达辐射源信号的双谱小波包特征,采用核模糊C均值聚类算法实现聚类分选.仿真结果表明:提取的特征参量在宽信噪比范围内均具有很好的分离性和稳定性,可实现复杂雷达辐射源信号的准确聚类分选.     

Clustering diagnosis of rolling element bearing fault based on integrated Autoregressive/Autoregressive Conditional Heteroscedasticity model

Feature extraction plays an important role in the clustering analysis. In this paper an integrated Autoregressive (AR)/Autoregressive Conditional Heteroscedasticity (ARCH) model is proposed to characterize the vibration signal and the model coefficients are adopted as feature vectors to realize clustering diagnosis of rolling element bearings. The main characteristic is that the AR item and ARCH item are interrelated with each other so that it can depict the excess kurtosis and volatility clustering information in the vibration signal more accurately in comparison with two-stage AR/ARCH model. To testify the correctness, four kinds of bearing signals are adopted for parametric modeling by using the integrated and two-stage AR/ARCH model. The variance analysis of the model coefficients shows that the integrated AR/ARCH model can get more concentrated distribution. Taking these coefficients as feature vectors, K means based clustering is utilized to realize the automatic classification of bearing fault status. The results show that the proposed method can get more accurate results in comparison with two-stage model and discrete wavelet decomposition.     

基于自回归模型的光阱中粒子运动模拟

通过分析光阱中颗粒位移信号特性, 建立描述粒子受限布朗运动过程的自回归模型, 进而提出了一种基于自回归模型的光阱中颗粒运动信号模拟的新方法. 对半径为1 μm的粒子处于光阱刚度分别为10, 20, 50 pN/μm 光阱时的位移信号进行了模拟, 得到的模拟位移信号的自相关函数与理论值相一致. 为了进一步阐明自回归模型的有效性, 在相同光阱参数下, 分别采用自回归模型与蒙特卡罗方法模拟光阱中微粒的位移信号, 采用功率谱法分别对两种模拟方法所得的微粒位移标定光阱刚度, 结果表明自回归模型方法能够取得和蒙特卡洛法相同的精度. 因此, 本文为分析光阱中粒子的随机运动提出了一种新的模拟方法, 可以用来对光阱中的噪声及特性进行分析. 关键词： 光阱 布朗运动 信号模拟 自回归模型     

Measuring the similarity between trajectories using clustering techniques

A clustering method has been developed to group signals that display similar dynamic behavior. The procedure involves using the method of time delay embedding to construct a trajectory in state space from a time series. Certain features that characterize the geometry of the trajectory have been defined. These features were subjected to a series of statistical tests to determine their usefulness in a hierarchical clustering analysis. The latter is aimed at finding groups of similar trajectories. The trajectory-based clustering algorithm has been applied to simulated data, which included both stochastic data generated by a linear AR model, and nonlinear data generated by a Duffing oscillator. The results show that the algorithm works reliably in both cases.     

Fault Feature Extraction for Reciprocating Compressors Based on Underdetermined Blind Source Separation

In practical engineering applications, the vibration signals collected by sensors often contain outliers, resulting in the separation accuracy of source signals from the observed signals being seriously affected. The mixing matrix estimation is crucial to the underdetermined blind source separation (UBSS), determining the accuracy level of the source signals recovery. Therefore, a two-stage clustering method is proposed by combining hierarchical clustering and K-means to improve the reliability of the estimated mixing matrix in this paper. The proposed method is used to solve the two major problems in the K-means algorithm: the random selection of initial cluster centers and the sensitivity of the algorithm to outliers. Firstly, the observed signals are clustered by hierarchical clustering to get the cluster centers. Secondly, the cosine distance is used to eliminate the outliers deviating from cluster centers. Then, the initial cluster centers are obtained by calculating the mean value of each remaining cluster. Finally, the mixing matrix is estimated with the improved K-means, and the sources are recovered using the least square method. Simulation and the reciprocating compressor fault experiments demonstrate the effectiveness of the proposed method.     

一种抗血氧波动干扰的肝储备功能测量方法

肝储备功能参数是评估肝脏代谢作用是否正常的关键指标,也是判断切除肝叶手术能否进行的重要依据。当前临床上获取肝储备功能参数是通过脉搏色素分光光度法测量吲哚菁绿色素浓度实现的,但是该方法需要假设血氧值为100%,这将导致肝储备功能参数的计算值存在一定误差。针对这一问题,提出了一种抗血氧波动干扰的肝储备功能参数测量方法,以修正的朗伯-比尔定律为理论基础,实现了对脉搏色素分光光度法测量的吲哚菁绿色素浓度的修正。在人体注入吲哚菁绿后,利用自制的数据采集单元在指端皮肤处同步采集805和940 nm的双波长透射信号,以及730,805和890 nm的三波长反射信号,将收集到的五组数据依次上传至计算机,利用接收到的数据和人体注射色素前的血氧值绘制出吲哚菁绿色素的浓度曲线,并计算其特征参数,根据浓度曲线的特征参数计算出肝储备功能参数。以有效肝脏血流量为例,将所提出的方法和脉搏色素分光光度法的测量结果分别跟目前测量有效肝脏血流量最准确的电磁流量计法的测量结果相比较,测量误差得到了明显的改善。实验结果表明,该方法提高了肝储备功能参数测量的精确度,为临床提供了一种更加准确的肝储备功能参数检测方法。     

基于AR模型与神经网络的核爆与闪电电磁脉冲信号识别

 对平稳随机信号功率谱估计的AR模型,分别利用自相关函数法和Burg算法求该模型系数,作为核爆炸和闪电电磁脉冲信号的特征值;采用BP神经网络作为分类器以及不同的隐含层数和隐含层节点数,对核爆和闪电电磁脉冲实测数据进行识别研究。结果表明：AR参数模型法对两类信号特征值提取是非常有效的,采用Burg算法来求AR模型参数,其特征值提取效果优于自相关函数法。     

Fault Recognition of Rolling Bearings Based on Parameter Optimized Multi-Scale Permutation Entropy and Gath-Geva

To extract fault features of rolling bearing vibration signals precisely, a fault diagnosis method based on parameter optimized multi-scale permutation entropy (MPE) and Gath-Geva (GG) clustering is proposed. The method can select the important parameters of MPE method adaptively, overcome the disadvantages of fixed MPE parameters and greatly improve the accuracy of fault identification. Firstly, aiming at the problem of parameter determination and considering the interaction among parameters comprehensively of MPE, taking skewness of MPE as fitness function, the time series length and embedding dimension were optimized respectively by particle swarm optimization (PSO) algorithm. Then the fault features of rolling bearing were extracted by parameter optimized MPE and the standard clustering centers is obtained with GG clustering. Finally, the samples are clustered with the Euclid nearness degree to obtain recognition rate. The validity of the parameter optimization is proved by calculating the partition coefficient and average fuzzy entropy. Compared with unoptimized MPE, the propose method has a higher fault recognition rate.     

Formant frequencies in Estonian folk singing.

Formant frequencies in an old Estonian folk song performed by two female voices were estimated for two back vowels /a/ and /u/, and for two front vowels /e/ and /i/. Comparison of these estimates with formant frequencies in spoken Estonian vowels indicates a trend of the vowels to be clustered into two sets of front and back ones in the F1/F2 plane. Similar clustering has previously been shown to occur in opera and choir singing, especially with increasing fundamental frequency. The clustering in the present song, however, may also be due to a tendency for a mid vowel to be realized as a higher-beginning diphthong, which is characteristic of the North-Estonian coastal dialect area where the singers come from. No evidence of a "singer's formant" was found.     

Phase transitions for information diffusion in random clustered networks

We study the conditions for the phase transitions of information diffusion in complexnetworks. Using the random clustered network model, a generalisation of the Chung-Lurandom network model incorporating clustering, we examine the effect of clustering underthe Susceptible-Infected-Recovered (SIR) epidemic diffusion model with heterogeneouscontact rates. For this purpose, we exploit the branching process to analyse informationdiffusion in random unclustered networks with arbitrary contact rates, and provide noveliterative algorithms for estimating the conditions and sizes of global cascades,respectively. Showing that a random clustered network can be mapped into a factor graph,which is a locally tree-like structure, we successfully extend our analysis to randomclustered networks with heterogeneous contact rates. We then identify the conditions forphase transitions of information diffusion using our method. Interestingly, for variouscontact rates, we prove that random clustered networks with higher clustering coefficientshave strictly lower phase transition points for any given degree sequence. Finally, weconfirm our analytical results with numerical simulations of both synthetically-generatedand real-world networks.     

Synchronizability of Highly Clustered Scale-Free Networks

We consider the effect of clustering coefficient on the synchronizability of coupled oscillators located on scale-free networks. The analytic result for the value of clustering coefficient aiming at a highly clustered scale-free network model, the Holme-Kim model is obtained, and the relationship between network synchronizability and clustering coefficient is reported. The simulation results strongly suggest that the more clustered the network, the poorer the synchronizability.