基于特征空间轨迹匹配方式的语音关键词检测法

语音关键词识别是近年来颇受重视的一个研究领域,文章基于特征空间轨迹的时间规整化原理,提出了一种高性能的关键词检测法,并探讨了轨迹等分长度对该算法检测性能的影响。实验结果表明,基于特征空间轨迹匹配方式的关键词检测法的检测性能接近于人工检测,具有一定的实用性。     

DTD的规范化

一个设计良好的DTD对于XML应用来说是必须的，从消除文档内数据冗余的角度出发研究了这一问题。函数依赖是数据语义的重要组成部分，将它引入到XML的领域中。给出的函数依赖可以是绝对的，也可以是相对的，键只是它的一种特例。讨论了逻辑蕴涵及其相应的推理规则，并证明了推理规则集的正确性和完备性。基于函数依赖，提出了规范化的DTD概念，并给出了一个将DTD转化为规范化形式的算法。     

Evaluation of the fracture toughness properties of polytetrafluoroethylene

Polytetrafluoroethylene (PTFE) (Dupont Tradename Teflon) is a common polymer with many structural applications including sheet, gaskets, bearing pads, piston rings and diaphragms. The interest here developed because this polymer is being considered as the major component of a newly proposed `reactive' material with a possible application as a projectile to replace common inertial projectiles. Little mechanical property data is available on this material since it is commonly used only as a coating material with the dominant properties being its low friction coefficient and high application temperature. Previous work (Joyce, 2003) on commercially available sheet PTFE material has demonstrated the applicability of the normalization method of ASTM E1820 (1999), the elastic-plastic fracture toughness standard to develop fracture toughness properties of this material over a range of test temperatures and loading rates. Additional work on the aluminum filled `reactive' derivative of the basic PTFE polymer (Joyce and Joyce, 2004) has also recently been completed. In this work, standard ASTM E1820 fracture toughness specimens machined from sintered pucks of PTFE were tested at four test temperatures and at a range of test rates to determine the J _Ic and J resistance curve characteristics of the PTFE material. The major results are that while crack extension is difficult at standard laboratory loading rates at ambient (21 °C) temperature or above, for temperatures slightly below ambient or for elevated loading rates, a rapid degradation of fracture resistance occurs and cracking occurs in a ductile or even nearly brittle manner.     

人工神经网络用于电除盐产水水质预测研究

为了实现化工行业高纯水全自动生产,本文利用电除盐（electrodeionization,EDI）技术替代传统的混床技术,并研究了人工神经网络对EDI产水过程模拟仿真的可行性,采用误差反向传播网络（BP网络）建立了进水流量、电导率、pH值以及工作电压与EDI透过水电阻率之间关系的动态模型,并对不同的训练样本归一化方法和训练方法进行比较。结果表明,在网络隐含层层数为1、节点数为13时,采用归一化方法三能够较好的预测EDI透过水电阻率,且该模型可用于EDI除盐过程的动态描述,为实现化工行业高纯水全自动生产奠定了基础。     

An imbalance fault detection method based on data normalization and EMD for marine current turbines

This paper proposes an imbalance fault detection method based on data normalization and Empirical Mode Decomposition (EMD) for variable speed direct-drive Marine Current Turbine (MCT) system. The method is based on the MCT stator current under the condition of wave and turbulence. The goal of this method is to extract blade imbalance fault feature, which is concealed by the supply frequency and the environment noise. First, a Generalized Likelihood Ratio Test (GLRT) detector is developed and the monitoring variable is selected by analyzing the relationship between the variables. Then, the selected monitoring variable is converted into a time series through data normalization, which makes the imbalance fault characteristic frequency into a constant. At the end, the monitoring variable is filtered out by EMD method to eliminate the effect of turbulence. The experiments show that the proposed method is robust against turbulence through comparing the different fault severities and the different turbulence intensities. Comparison with other methods, the experimental results indicate the feasibility and efficacy of the proposed method.     

Partial aggregation of users for biometric scores normalization

In this paper, we analyze a new method to deal with users' biometric data, by incorporating them in an incomplete aggregation framework, to make an optimal decision for the normalization process. This latter invokes the experts' opinions concept, where the opinions are represented by the user-specific scores. This aggregation of opinions needs more experts for a trusted and informative partial consensus. The key idea of this new approach is based upon the employment of a convex mixture of users' score dispersion. This procedure is used to revise the users' scores, then to update specific statistical parameters which are the principal elements of our score normalization formula. Applying this concept to various scenarios gives an interesting outcome compared to other alternatives. The main contribution of this work for aggregation framework on native space is to introduce this new user-specific concept of scores' normalization, followed by a thorough evaluation of the performance. The validation of this approach is ensured by using a unique modality followed by the exploitation of the fusion of different modalities, to achieve a viable reliability of the final decision. The benchmarking was conducted under Nist2005, Biosecure DS2, and XM2VTS score databases.     

改进卷积输入的端到端普通话语音识别

主流神经网络训练的交叉熵准则是对声学数据的每个帧进行分类优化，而连续语音识别是以序列级转录准确性为性能度量。针对这个不同，构建基于序列级转录的端到端语音识别系统。针对低资源语料条件下系统性能不佳的问题，其中模型使用卷积神经网络对输入特征进行处理，选取最佳的网络结构，在时域和频域进行二维卷积，从而改善输入空间中因不同环境和说话人产生的小扰动影响。同时神经网络使用批量归一化技术来减少泛化误差，加速训练。基于大型的语言模型，优化解码过程中的超参数，提高模型建模效果。实验结果表明系统性能提升约24%，优于主流语音识别系统。     

基于贝叶斯信念网络的多案例库检索方法

为了克服在基于案例推理中单一案例库检索的局限性,提出了基于贝叶斯信念网络的多案例库检索方法.该方法仅要求用户采用关键词对目标案例进行描述,通过贝叶斯信念网络计算关键词与各案例库的特征属性的匹配概率,以获得适应各案例库的完整的目标案例描述,然后对相应的案例库进行案例检索;将从多案例库中检索到的案例通过相似度线性归一处理后,结合目标案例描述的匹配概率获得与目标案例的最终相似度,统一排序后返回检索结果.最后通过航空企业的实际应用与性能分析进一步验证了多案例库检索方法的有效性.     

基于LSTM的发电机组污染物排放预测研究

为了利用电力公司积累的海量历史污染物排放数据,形成可以减少污染物排放的调度框架。采用递归神经网络,结合发电机组输出功率与污染物排放量之间的关系,并使用批规范化等深度学习技术,对数据和模型进行学习和训练。实验结果表明,可以有效预测发电机组污染物排放量,解决传统回归分析方法无法适用的难以提取有效特征的问题。     

采用自适应归一化实数编码遗传算法优化鱼眼镜头系统

针对遗传算法(GA)优化超多参量光学系统时鲁棒性 较差的问题,在混入逃逸函数实数编码GA(MERCGA)的基础上,进一步结合参量归一化和自适 应变异概率的措施,提出了自适应归一化 实数编码GA(ANRCGA)。用ANRCGA对鱼眼镜头光学系统案例进行优化设计,并应用 评价函 数和Zemax光线追迹方法对MERCGA和ANRCGA的优化结果作比较。结果表明,应用本文的ANRCG A 比引自专利的参考设计及MERCGA优化得到光学系统的成像质量有明显提高,算法的鲁棒 性和计算效率也到了改善。