基于DNN的半潜式平台系统稳定性预测模型

针对半潜式平台系统故障警报信号频发、系统运行稳定性差等问题，基于半潜式平台工作信号点位数据集研究影响半潜式平台稳定运行的重要因素。基于这些影响因素采用机器学习、深度学习算法构建基于深度神经网络（Deep Neural Network，DNN）的平台系统稳定性预测模型，该模型的曲线下面积（Area Under Curve，AUC）得分较逻辑回归、K近邻查询、支持向量机、朴素贝叶斯等传统机器学习模型的AUC得分提高1.0%~16.0%、准确率提高3.0%~25.6%，表明DNN模型具有较好的拟合能力和泛化能力，可以用于工业实践。     

WebRTC中一种基于DNN的噪声抑制算法的研究

WebRTC(Web Real-Time Communication)是谷歌开源的实时通信技术,包含了实现音视频会议系统所需的全部技术。文中对WebRTC的语音引擎中的降噪(Noise Suppression)算法进行了深入剖析,并且提出一种针对实际音视频会议系统环境下的基于DNN(Deep Neutral Network)的噪声抑制算法,实验结果表明提出的噪声抑制算法对于音视频会议系统中常见噪声有着很不错的噪声抑制效果,尤其对于人群嘈杂噪声。在低信噪比条件下的噪声抑制效果优于WebRTC中的降噪算法。     

Fast speaker adaptation using extended diagonal linear transformation for deep neural networks

This paper explores new techniques that are based on a hidden‐layer linear transformation for fast speaker adaptation used in deep neural networks (DNNs). Conventional methods using affine transformations are ineffective because they require a relatively large number of parameters to perform. Meanwhile, methods that employ singular‐value decomposition (SVD) are utilized because they are effective at reducing adaptive parameters. However, a matrix decomposition is computationally expensive when using online services. We propose the use of an extended diagonal linear transformation method to minimize adaptation parameters without SVD to increase the performance level for tasks that require smaller degrees of adaptation. In Korean large vocabulary continuous speech recognition (LVCSR) tasks, the proposed method shows significant improvements with error‐reduction rates of 8.4% and 17.1% in five and 50 conversational sentence adaptations, respectively. Compared with the adaptation methods using SVD, there is an increased recognition performance with fewer parameters.     

深度学习视阈下MOOC学习者流失预测及干预研究

MOOC（Massive Open Online Courses）在为学习者提供优质课程的同时，低完成率成为影响其有效推广的重要因素。通过对edX开放数据集分析发现，学习者的逐渐流失是导致MOOC课程低完成率的因素之一，且学习行为与成绩之间存在复杂的相关性；基于线性回归和深度神经网络（Deep Neural Network，DNN）预测学习者的成绩，实验证明，DNN能够更好地拟合学习行为与成绩之间复杂的相关性，实现对成绩更加精准的预测，预警学习者流失；对预测的潜在流失学习者迭代进行个性化的教学干预，提高MOOC课程的完成率。     

Real-time rock mass condition prediction with TBM tunneling big data using a novel rock–machine mutual feedback perception method

Real-time perception of rock mass information is of great importance to efficient tunneling and hazard prevention in tunnel boring machines (TBMs). In this study, a TBM–rock mutual feedback perception method based on data mining (DM) is proposed, which takes 10 tunneling parameters related to surrounding rock conditions as input features. For implementation, first, the database of TBM tunneling parameters was established, in which 10,807 tunneling cycles from the Songhua River water conveyance tunnel were accommodated. Then, the spectral clustering (SC) algorithm based on graph theory was introduced to cluster the TBM tunneling data. According to the clustering results and rock mass boreability index, the rock mass conditions were classified into four classes, and the reasonable distribution intervals of the main tunneling parameters corresponding to each class were presented. Meanwhile, based on the deep neural network (DNN), the real-time prediction model regarding different rock conditions was established. Finally, the rationality and adaptability of the proposed method were validated via analyzing the tunneling specific energy, feature importance, and training dataset size. The proposed TBM–rock mutual feedback perception method enables the automatic identification of rock mass conditions and the dynamic adjustment of tunneling parameters during TBM driving. Furthermore, in terms of the prediction performance, the method can predict the rock mass conditions ahead of the tunnel face in real time more accurately than the traditional machine learning prediction methods.     

Soft Sensor design for a Topping process in the case of small datasets

In this paper, a new strategy to cope with the identification of nonlinear models of industrial processes, when a limited number of experimental data is available, is proposed. The approach is intended to improve the generalization capabilities of the model and it is based on the integration of bootstrap resampling, noise injection and neural model stacking. A number of algorithms to stack the first level neural models are also compared. The method proposed has been applied to develop a Soft Sensor for the estimation of the Freezing Point of Kerosene in an atmospheric distillation unit (Topping) working in a refinery in Sicily, Italy. The improvements obtained thanks to the strategy proposed, with respect to a classical neural model, are shown in the paper.     

OPENID网站认证技术在地理信息系统中的应用

传统的身份密码网络认证方式非常容易被攻破,因此需要安全的、新的网络身份认证方式来防止网络钓鱼、网络欺骗和黑客破解等。OPENID用户认证方式就是其中一种较好的网络认证系统。OPENID身份验证技术与DNN门户网站相结合,应用于现有油田地理信息系统中取得了很好的效果。     

基于特征增强聚合的融合广告点击率预测模型

传统点击率（CTR）预测模型多在单一特征级上进行特征交互,未能充分利用不同特征级上的有效信息。基于特征增强聚合方法提出一种融合广告CTR预测（APNN）模型。在CTR预测模型的嵌入层中引入一阶信息重要性进行特征增强,通过注意力因子分解机（AFM）模型与基于乘积产生层的神经网络（PNN）模型融合不同特征级交互特征和增强的嵌入特征,并利用多个全连接层从融合特征中获得更多潜在的高阶信息。实验结果表明,相比AFM、PNN、FNN等模型,APNN模型的预测精度较高,其在Criteo数据集上的AUC和LogLoss指标较PNN模型分别提高1.74和1.42个百分点。     

基于平移随机变换的对抗样本生成方法

基于深度神经网络的图像分类模型能够以达到甚至高于人眼的识别度识别图像,但是因模型自身结构的脆弱性,导致其容易受对抗样本的攻击。现有的对抗样本生成方法具有较高的白盒攻击率,而在黑盒条件下对抗样本的攻击成功率较低。将数据增强技术引入到对抗样本生成过程中,提出基于平移随机变换的对抗样本生成方法。通过构建概率模型对原始图像进行随机平移变换,并将变换后的图像用于生成对抗样本,有效缓解对抗样本生成过程中的过拟合现象。在此基础上,采用集成模型攻击的方式生成可迁移性更强的对抗样本,从而提高黑盒攻击成功率。在ImageNet数据集上进行单模型和集成模型攻击的实验结果表明,该方法的黑盒攻击成功率高达80.1%,与迭代快速梯度符号方法和动量迭代快速梯度符号方法相比,该方法的白盒攻击成功率虽然略有降低,但仍保持在97.8%以上。     

基于BiLSTM模型的定义抽取方法

定义抽取是从非结构化文本中自动识别定义句的任务,定义抽取问题可建模为句子中术语及相应定义的序列标注问题,并利用标注结果完成抽取任务。针对传统的定义抽取方法在抽取定义特征过程中费时且容易造成错误传播的不足,提出一个基于双向长短时记忆(BiLSTM)的序列标注神经网络模型,对输入文本进行自动化定义抽取。通过将原始数据输入到BiLSTM神经网络中,完成输入句的特征表示,并采用基于LSTM的解码器进行解码得到标注结果。在Wikipedia英文数据集上的实验结果表明,该方法的精确率、召回率和F1值分别为94.21%、90.10%和92.11%,有效提升了基准模型效果。