Improved consonant–vowel recognition for low bit‐rate coded speech

In this paper, we proposed a method for improving the recognition performance of 145 prominent consonant–vowel (CV) units in Indian languages for low bit‐rate coded speech. Proposed CV recognition method is carried out in two levels to reduce the similarity among a large number of CV classes. In the first level, vowel category of CV unit will be recognized, and in the second level, consonant category will be recognized. At each level of the proposed method, complementary evidences from support vector machine and hidden Markov models are combined to enhance the recognition performance. Effectiveness of the proposed two‐level CV recognition method is demonstrated by performing the recognition of isolated CV units and CV units collected from the Telugu broadcast news database. In this work, vowel onset point (VOP) is used as an anchor point for extracting accurate features from the CV unit. Therefore, a method is proposed for accurate detection of VOP in clean and coded speech. The proposed VOP detection method is based on the spectral energy in 500–2500 Hz frequency band of the speech segments present in the glottal closure region. Speech coders considered in this work are GSM full rate (ETSI 06.10), CELP (FS‐1016), and MELP (TI 2.4 kbps). Significant improvement in CV recognition performance is achieved using the proposed two‐level method compared with the existing methods under both clean and coded conditions. Copyright © 2011 John Wiley & Sons, Ltd.     

MSVQ-based speaker-adaptive Chinese syllable recognition based on discriminative training

In this paper we present two supervised speaker adaptation methods, including a feature normalization and an MCE/GPD algorithm, developed to implement an MSVQ-based adaptive Chinese syllable recognition system. In the MSVQ-based speech recognition, each recognition unit is represented as a time sequence of codebooks. The first proposed method is feature normalization, in which we model the inter-speaker variability as a linear transformation. By applying the feature normalization, the target speaker speech is normalized to reduce the inter-speaker acoustic variability. In the second adaptation method we first present an implementation of the MCE/GPD algorithm for discriminatively training an MSVQ-based speech recognizer. It is expected that this method can separate the confusion classes and can enhance speaker adaptation capability. By applying the MCE/GPD algorithm, the MSVQ-based recognizer parameters are adjusted iteratively to accomplish the objective of minimum classification error rate. We carried out recognition experiments of highly confusing Chinese syllables to assess its performance. Using the standard Chinese syllable database CRDB in China, the results show that when the two adaptation methods are combined, the error rate reduction on open data is over 62% with a single set of adaptation training data. Therefore, when the amount of adaptation data is limited, the adaptation methods can lead to substantial improvement. Upon increasing the training data, the capability of speaker adaptation is improved by using the MCE/GPD training only, so it can be used for tracking spectral evolution over time and provides a robust means for adaptive speech recognition. © 1997 John Wiley & Sons, Ltd.     

Discriminative training of GMM-HMM acoustic model by RPCL learning

This paper presents a new discriminative approach for training Gaussian mixture models (GMMs) of hidden Markov models (HMMs) based acoustic model in a large vocabulary continuous speech recognition (LVCSR) system. This approach is featured by embedding a rival penalized competitive learning (RPCL) mechanism on the level of hidden Markov states. For every input, the correct identity state, called winner and obtained by the Viterbi force alignment, is enhanced to describe this input while its most competitive rival is penalized by de-learning, which makes GMMs-based states become more discriminative. Without the extensive computing burden required by typical discriminative learning methods for one-pass recognition of the training set, the new approach saves computing costs considerably. Experiments show that the proposed method has a good convergence with better performances than the classical maximum likelihood estimation (MLE) based method. Comparing with two conventional discriminative methods, the proposed method demonstrates improved generalization ability, especially when the test set is not well matched with the training set.     

基于RBF网络的火电机组实时成本在线建模方法

针对火电厂实时成本建模问题，该文提出一种基于RBF网的在线建模方法，即资源优化网络，简称RON。RON能自动根据最近一段时间内的误差信息优化网络结构：如果当前网络不能实现新输入样本，则在线增加新隐节点，否则使用梯度法在线调节网络的隐节点数据中心和扩展常数。数据中心和扩展常数调整过程中还引入了隐节点的合并操作和删除操作。进一步介绍了火电厂实时成本建模方法，包括网络结构的确定，及如何获取训练样本。仿真实例表明，RON对能较好地适应对象参数的时变特性，并对训练样本集变化具有较好的鲁棒性。     

计及风光相位特性和机-网间元件的短路电流计算方法

以风电和光伏为代表的新能源机组大量接入电网使得短路电流的精确计算校核变得困难。针对最新研究中构建新能源并网点短路电流映射模型时,未考虑相位映射及未计及箱变、集电线路、主变及送出线路等机端到并网点间电气元件(机-网间元件)影响的理论缺陷,文中提出一种计及新能源相位特性和机-网间元件影响的改进工程化电网短路电流计算方法。首先,结合国家运行规程对新能源机端电压-短路电流相位映射进行有补充意义的理论建模。其次,对该模型进行工程化处理,并提出一种将所得映射逆推至并网点的迭代计算方法。继而,将工程化后的并网点短路电流幅相映射模型应用到现有局部迭代计算中,得到改进的计算方法。最后,在PSCAD仿真软件中搭建新能源并网模型,验证了新能源机端电压-短路电流理论幅相映射的准确性、机端映射模型逆推至并网点的迭代计算理论有效性;在此基础上,在IEEE 39节点系统中对所得映射模型进行实验验证。结果表明所提改进计算方法能在一定程度上提升短路电流计算精度。     

基于 FCN 的阀门内泄漏声发射信号识别方法

针对石化工业中输气管道阀门的内泄漏故障,将声发射检测技术与深度学习技术相结合,提出了一种基于全卷积神经 网络(FCN)的阀门内泄漏声发射信号识别方法。 该方法利用声发射技术采集阀门内泄漏的声发射信号,基于 FCN 搭建阀门内 泄漏分类诊断模型,充分发挥了声发射技术在阀门内泄漏检测领域的优越性,以及 FCN 在时间序列分类任务上的高性能。 该 方法相较于传统的识别方法,无需对原始采集数据进行特征提取或繁重复杂的预处理,而是将特征提取的任务也交于神经网络 模型来学习和完成,可实现端到端的阀门内泄漏声发射信号分类识别。 搭建阀门内泄漏检测实验平台,采集并制作阀门内泄漏 声发射信号数据集,建立了基于 FCN 的阀门内泄漏声发射信号的二分类模型,实验结果表明,该模型的分类识别准确率可达 98. 72%,相比较于其他先进的分类模型在数据集上表现出了更加优越的分类识别性能和训练效率,同时对环境噪声具有良好 的抗干扰性能。     

大型发电机组等效故障次数的分析与计算

大型发电机组在运行过程中存在大量的强迫降出力(IUD)事件。在利用随机过程模型分析大型发电机组的可靠性指标时,若将一次IUD事件计为1次故障。则对机组的可靠性水平估计偏低;若忽略IUD的影响,不计入故障,又使机组的可靠性估计偏高。文中针对这种情况,引入“分数次故障”的概念,将IUD等效为分数次故障,用分数次故障的大小来反映IUD的故障严重程度,从而使发电机组的可靠性评估更接近实际情况。此外,现有的随机过程模型都是基于整数次故障的,为便于利用这些模型,文中还介绍了将分数次故障进行整数化的方法。最后用实例简要说明了应用上述方法处理某地区一台300MW发电机组原始数据,并获得有效故障数据的情况。     

Lecture speech recognition using discrete‐mixture HMMs

Most of the state‐of‐the‐art speech recognition systems use continuous‐mixture hidden Markov models (CMHMMs) as acoustic models. On the other hand, it is well known that discrete hidden Markov model (DHMM) systems show poor performance because they are affected by quantization distortion. In this paper, we present an efficient acoustic modeling based on discrete distribution for large‐vocabulary continuous speech recognition (LVCSR). In our previous work, we proposed the maximum a posteriori (MAP) estimation of discrete‐mixture hidden Markov model (DMHMM) parameters and showed that the DMHMM system performed better in noisy conditions than the conventional CMHMM system. However, we conducted the recognition experiments on a read/speech task in which the vocabulary size was only 5k. In addition, the DMHMM was not effective in clean condition in that work. In this paper, we have developed a DMHMM‐based LVCSR system and evaluated the system on a more difficult task under clean condition. In Japan, a large‐scale spontaneous speech database ‘Corpus of Spontaneous Japanese’ has been used as the common evaluation database for spontaneous speech and we used it for our experiments. From the results, it was seen that the DMHMM system showed almost the same performance as the CMHMM system. Moreover, performance improvement could be achieved by a histogram equalization method. Copyright © 2010 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.     

基于深度神经网络和多元损失的说话人识别

生物特征识别技术相对于传统密码等方式具有更高的可靠性,而作为生物特征识别技术的重要研究方向之一的声纹识别方法,研究更精确的声纹识别方法具有更高的研究意义。随着深度学习的发展,深度学习应用在声纹识别技术上成为在声纹识别领域研究的重点。提出一种基于深度神经网络和beyond triplet loss相结合的说话人识别方法,模型通过梅尔频率倒谱系数（MFCC）提取MFCC声学特征,对MFCC声学特征提取说话人声纹特征,然后进行多元损失的模型训练。实验结果表明,DNN-BTL算法在说话人识别领域比高斯混合-隐马尔可夫模型（GMM-HMM）具有更好的识别效果。     

450t/h循环流化床锅炉机组动态仿真模型研究

采用模块化建模方法建立了我国设计制造的首台450t/h循环流化床锅炉机组实时仿真数学模型,介绍了锅炉机组模型的各子系统构成及其性能.以循环流化床燃烧系统为例,详细介绍了其主要仿真模块的数学模型,并据此仿真模块搭建了该子系统的仿真模型.仿真试验表明所建模型能够正确反映循环流化床锅炉机组的动态和静态特性,模型运算稳定可靠.所建锅炉机组仿真模型已用于多期仿真机培训,并可为循环流化床锅炉机组控制系统设计与分析提供对象模型.     

Surrogate modeling of microwave structures using kriging,co‐kriging,and space mapping

Space mapping (SM) is one of the most popular techniques for creating computationally cheap and reasonably accurate surrogates of electromagnetic‐simulated microwave structures (so‐called fine models) using underlying coarse models, typically equivalent circuits. One of the drawbacks of SM is that although good modeling accuracy can be obtained using a limited number of training points, SM is not capable of efficiently utilizing larger amount of fine model information, even if it is available. In this paper, we consider various ways of enhancing SM surrogates by exploiting additional training data as well as two function approximation methodologies, kriging and co‐kriging. To our knowledge, it is the first application of co‐kriging for microwave circuit modeling. With three examples of microstrip filters, we present a comprehensive numerical study in which we compare the accuracy of the basic SM models as well as SM enhanced by kriging and co‐kriging. Direct kriging interpolation of fine model data is used as a reference. Copyright © 2012 John Wiley & Sons, Ltd.     

基于控制变量法的含风电场发电系统可靠性评估

针对蒙特卡洛方法抽样容量大、效率低的问题,考虑大型风电场单机容量小、机组数多的特点,提出控制变量法抽样,以实现含风电场的发电系统可靠性评估的快速收敛。本文以常规机组的可靠性指标为控制变量,用解析法计算常规机组的可靠性指标,依次对常规机组和风电机组循环抽样。以IEEE RTS79标准测试系统为算例,对所提算法验证并与常规蒙特卡洛法、等分散抽样法对比,对影响系统抽样效率的因素进行分析。结果表明该方法能大幅减小抽样次数,提高抽样效率,同时保证系统精度。     

基于BERT-DSA-CNN和知识库的电网调控在线告警识别

电网调控告警识别是实现智能电网调度的重要环节。为提高电网调控告警识别的准确率,针对电网数据量庞大、有效信息提取困难、传统知识库知识迁移能力较差等问题,提出一种基于BERT-DSA-CNN和知识库的电网调控在线告警识别方法。首先在自然语言处理-深度学习的文本数据挖掘架构基础上,经过分词、去停用词等步骤,利用BERT模型获取电网调控告警信息词向量。然后将词向量输入CNN深度学习模型进行训练,并根据电网告警信息的特点引入DSA机制对CNN模型进行改进。最后提出了融合深度学习模型和传统知识库的电网调控在线告警识别方案。通过大量的算例结果分析得出,该方法相比Word2vec、传统CNN、传统知识库、离线学习等方法,具有更高的准确性和有效性,对不同的故障类型均具有较好的识别能力,为工程应用提供了一种思路。     

数据驱动的配电台区源荷污染源群体谐波排放建模

针对高密度、分散化、全网化电力电子非线性设备导致谐波污染难以有效估计问题,提出一种数据驱动的配电台区谐波污染源群体谐波排放水平建模方法.考虑多种谐波源模型特点和适用性,选取谐波Norton等效模型对谐波源负荷设备进行建模,形成设备典型谐波排放表征.利用非侵入式负荷监测(non-intrusive load monito...     

基于改进CatBoost的电力系统暂态稳定评估方法

在实际电网的运行过程中,通过同步相量测量单元实时采集到的电网动态参数通常含有部分噪声,且有时会因通信故障造成数值的随机缺失,对基于人工智能的电力系统暂态稳定评估模型造成很大影响.为此,提出一种基于改进CatBoost的暂态稳定评估方法.通过分箱算法对输入特征数据进行离散化处理,提高模型对噪声的鲁棒性;采用加权的焦点损失函数代替交叉熵损失函数,提升模型的可信度并减少模型对失稳样本的漏判;将量测数据部分缺失的样本划分到单独的节点中继续建模,从而充分挖掘不完整样本中的暂态信息.在新英格兰10机39节点上的实验结果表明,所提方法的准确率和查全率均优于其他几类机器学习算法,而且所提方法对噪声和数值缺失表现出良好的鲁棒性且具有较快的训练速度和预测速度.     

高海拔地区特高压换流站大尺寸复合支柱绝缘子直流污闪特性

在海拔2 100m的国家特高压工程实验室,采用恒压升降法全面深入地研究了一大一小和一大两小伞形、6.25m(5支串联组成)绝缘高度的两种大尺寸复合支柱绝缘子的直流污闪特性。试验表明:一大两小伞形大尺寸复合支柱绝缘子的直流污闪特性优于一大一小伞形;大尺寸试品的直流污闪电压与盐密、灰密均成幂函数关系,盐密对U50%的影响程度明显大于灰密;与短串试品相比,盐密对大尺寸试品的影响较大;随着上、下表面污秽比的增大,大尺寸试品直流污闪电压增加不超过10%;在绝缘高度6.25m范围内,直流污闪电压与复合支柱绝缘子串长呈良好线性关系,但由于试验数据存在一定的分散性,大尺寸试品线性推导值和试验值相差约15%;大尺寸试品的紫外放电光子数大于短串,同时,受污秽不均匀分布的影响,大尺寸试品的中、下部紫外光子数稍大于上部;大尺寸试品的中、下端电弧桥接现象剧烈,电弧大范围跨接更为明显。     

基于RetinaNet和类别平衡采样方法的销钉缺陷检测

传统的无人机巡检航拍图中的电力连接金具销钉缺陷检测依赖人工进行标注,针对此问题,借助深度学习缺陷检测算法RetinaNet自动提取正常、缺陷样本的特征,完成低层特征和顶层特征的融合,实现销钉缺陷的自动标注。考虑到现实情况中缺陷类别样本数量远少于正常类别样本数量,首先分析了缺陷数据不足引起的类别失衡对识别结果的影响,结果表明该情况下训练好的模型将会使得大量缺陷样本被错误地识别为正常类。于是,在数据层面采用类别平衡采样方法,确保每个类别参与训练的机会均衡,实验结果表明,所提的方法能够在维持销钉正常类的高识别率前提下,明显提高缺陷类别的平均准确率。     

基于SDAE-ELM伪量测建模的高容错快速状态估计方法

目前在电力系统中无法保证相量量测单元完全覆盖的情况下,状态估计需要采用相量量测单元（phasor measurement unit, PMU）与数据采集与监控（supervisory control and data acquisition, SCADA）混合量测进行传统非线性状态估计,但是SCADA数据精度低,含有较多不良数据,同时混合数据需要迭代求解,会导致计算效率低且存在截断误差。针对该问题,文章提出了一种基于堆叠去噪自编码器（stack denoising autoencoder, SDAE）与极限学习机（extreme learning machine, ELM）伪量测建模的电力系统高容错快速状态估计方法。其将含有不良量测的SCADA量测数据作为SDAE-ELM伪量测模型的输入,节点电压实部与虚部作为输出,根据历史数据进行训练得到伪量测值与伪量测误差模型,训练完成后得到精度较高的伪量测;将伪量测与PMU量测一起进行快速的线性状态估计。仿真结果表明,所提方法在保证估计精度的基础上,提高了计算效率,验证了所提方法的有效性。     

基于改进SinGAN的电力线巡检异物数据增强技术

针对电力线异物识别模型能使用的数据集较少,并且传统单幅自然图像的生成式模型(SinGAN)模型生成数据与异物识别模型匹配度不高、质量不佳、耗时过久的问题,提出了改进SinGAN模型。在改进SinGAN模型基础上加入仿射变换单元、大小变换单元进一步增强数据集,加入图像滤波单元提高电力线异物识别模型所需数据质量。并通过改进SinGAN反向传播训练过程和SinGAN的单精度生成器结构提升模型生成质量,减少所用时长。实验结果表明,经50次实验后,改进SinGAN的平均弗雷谢特起始距离(Fréchet inception distance, FID)为91.375,平均训练时长1.21 h。分别比传统SinGAN降低了27.247%和87.31%。改进SinGAN与其他主流生成式对抗网络相比有更好的异物数据生成能力,可以增强电力线异物识别模型所需数据,具有优越性。     

基于 AG-CNN 的轻量级调制识别方法

针对传统卷积神经网络在调制方式盲识别过程中,存在模型体积大、运算量高、无法部署至移动端等问题,提出了一种基于双注意力机制与Ghost模块的轻量级CNN模型AG-CNN(attention and Ghost convolution neural network)调制识别方法,该方法首先将调制信号映射至复空间,并根据归一化点密度对映射点进行颜色处理,得到高阶特征密度星座图;将该特征作为AG-CNN模型的输入进行学习训练,最后使用训练好的模型对接收端接收到的未知信号进行识别。实验表明,AG-CNN模型对散点为10 000的密度星座图识别率在99.95%以上,与相同层数的CNN模型相比,卷积层参数量压缩6.01倍,计算量压缩6.76倍,且相较于VGG-16、InceptionV3、ResNet-50、Shufflenet、Efficientnet等卷积网络模型,参数量与浮点数运算数下降明显,且在大幅节省学习参数量、降低模型复杂度的情况下,表现出优秀的分类性能。