基于Leap Motion的动态手势识别

动态手势识别作为人机交互的一个重要方向，在各个领域具有广泛的需求。相较于静态手势，动态手势的变化更为复杂，对其特征的充分提取与描述是准确识别动态手势的关键。为了解决对动态手势特征描述不充分的问题，利用高精度的Leap Motion传感器对手部三维坐标信息进行采集，提出了一种包含手指姿势和手掌位移的特征在内的、能够充分描述复杂动态手势的特征序列，并结合长短期记忆网络模型进行动态手势识别。实验结果表明，提出的方法在包含16种动态手势的数据集上的识别准确率为98.50%；与其他特征序列的对比实验表明，提出的特征序列，能更充分准确地描述动态手势特征。     

基于注意力机制的双向长短时记忆网络模型突发事件演化关系抽取

针对现有突发事件关系抽取研究多集中于因果关系抽取而忽略了其他演化关系的问题，为了提高应急决策中信息抽取的完备性，应用一种基于注意力机制的双向长短时记忆（LSTM）网络模型进行突发事件演化关系抽取。首先，结合突发事件演化关系的概念，构建演化关系模型并进行形式化定义，依据模型对突发事件语料进行标注；其次，搭建双向LSTM网络结构，并引入注意力机制计算注意力概率以突出关键词汇在文本中的重要程度；最终，使用搭建的网络模型进行演化关系抽取得到结果。在演化关系抽取实验中，相对于现有因果关系抽取方法，所提方法不仅抽取出更加充分的演化关系，为突发事件应急决策提供了更完善的信息；同时，在正确率、召回率和F1分数上分别平均提升了7.3%、6.7%和7.0%，有效提高了突发事件演化关系抽取的准确性。     

Soft sensing of water depth in combined sewers using LSTM neural networks with missing observations

Information and communication technologies combined with in-situ sensors are increasingly being used in the management of urban drainage systems. The large amount of data collected in these systems can be used to train a data-driven soft sensor, which can supplement the physical sensor. Artificial Neural Networks have long been used for time series forecasting given their ability to recognize patterns in the data. Long Short-Term Memory (LSTM) neural networks are equipped with memory gates to help them learn time dependencies in a data series and have been proven to outperform other type of networks in predicting water levels in urban drainage systems. When used for soft sensing, neural networks typically receive antecedent observations as input, as these are good predictors of the current value. However, the antecedent observations may be missing due to transmission errors or deemed anomalous due to errors that are not easily explained. This study quantifies and compares the predictive accuracy of LSTM networks in scenarios of limited or missing antecedent observations. We applied these scenarios to an 11-month observation series from a combined sewer overflow chamber in Copenhagen, Denmark. We observed that i) LSTM predictions generally displayed large variability across training runs, which may be reduced by improving the selection of hyperparameters (non-trainable parameters); ii) when the most recent observations were known, adding information on the past did not improve the prediction accuracy; iii) when gaps were introduced in the antecedent water depth observations, LSTM networks were capable of compensating for the missing information with the other available input features (time of the day and rainfall intensity); iv) LSTM networks trained without antecedent water depth observations yielded larger prediction errors, but still comparable with other scenarios and captured both dry and wet weather behaviors. Therefore, we concluded that LSTM neural network may be trained to act as soft sensors in urban drainage systems even when observations from the physical sensors are missing.     

面向限价指令簿趋势分析的网络集成模型

为了更好地分析限价指令簿(LOBs)的趋势,文中提出面向LOBs趋势分析的网络集成模型(NEM-LOB).模型融合2个长短期记忆(LSTM)子模型和1个卷积神经网络(CNN)子模型.一个LSTM子模型可通过LOBs的分布信息捕捉全局时间依赖性,另一个LSTM子模型可通过LOBs和订单流的动态信息捕捉全局动态性.CNN子模型通过LOBs的事实信息提取局部特征.最后,结合3个子模型,提取特征以获得预测结果.在FI-2010数据集上的实验表明NEM-LOB通过引入订单流信息,能对LOBs进行更好的趋势分析.     

基于长短期记忆神经网络的重大活动中鲜(冻)肉制品铅含量风险预测

目的 针对未来时间内重大活动举办地鲜(冻)肉制品铅含量风险预测的问题,建立基于LSTM的时间序列预测模型,对当地鲜(冻)肉制品铅含量进行风险评估与预测预警。方法 通过收集2011-2020年国家市场监督管理总局日常食品监督管理抽检数据,筛选出北京的鲜(冻)肉制品的抽检数据,构建数据集并进行预处理,按照7:2:1的比例划分训练集、验证集和测试集,基于Tensorflow平台构建4层LSTM模型并进行训练,基于重大活动举办前10天的鲜(冻)制品铅含量数据,对未来1天的鲜(冻)制品铅含量风险进行预测。结果 实验表明,经过50轮模型迭代训练,训练集和测试集Loss指标收敛至0.084,经过5次训练后的模型评估参数RMSE为0.192,R2_score为0.916,模型误差较小、准确度较高。结论 基于LSTM的鲜(冻)肉制品铅含量风险预测模型整体性能较好,可应用于重大活动举办地的食品风险预测,并精准指导监督抽检。     

基于XGBoost-LSTM组合模型的光伏发电功率预测

该文提出一种基于极端梯度提升（XGBoost）模型和长短期记忆网络（LSTM）模型的短期光伏发电功率预测组合模型。根据短期光伏发电特性,首先分别建立XGBoost模型和LSTM模型,然后利用XGBoost模型进行初步预测增加特征,并利用误差倒数法将两模型组合起来进行预测。选取2018年光伏电站人工智能运维大数据处理分析大赛的数据集进行实验评估,最终结果表明,该文所构建的XGBoost-LSTM组合模型的均方根误差（RMSE）为0.214,将上述方法与随机森林、GBDT模型和单一的XGBoost模型和LSTM模型相比较,该文提出的方法具有更高的预测精度。     

大坝变形的XGBoost-LSTM变权组合预测模型及应用

为了实现更高精度的大坝变形预报,提出了一种大坝变形的XGBoost-LSTM变权组合预测模型,即首先引入XGBoost模型和LSTM模型对大坝变形分别进行分析预测,然后采用变权组合方法将二者的分析预测结果进行有机融合,进而得到最终预测结果。以某混凝土重力坝为例,首先通过与随机森林、ELMAN以及逐步回归分析各模型的对比研究,论证了XGBoost与LSTM应用于大坝变形预测的优越性;进一步地,XGBoost与LSTM的变权组合预测效果相较于各单一模型取得了较大程度的提升,且相较于二者的等值赋权组合提升优势更加显著,变形预测结果与工程实际情况更加吻合,具有较好的适用性和可推广价值。     

Deep Neural Architecture for Recovering Dropped Pronouns in Korean

Pronouns are frequently dropped in Korean sentences, especially in text messages in the mobile phone environment. Restoring dropped pronouns can be a beneficial preprocessing task for machine translation, information extraction, spoken dialog systems, and many other applications. In this work, we address the problem of dropped pronoun recovery by resolving two simultaneous subtasks: detecting zero-pronoun sentences and determining the type of dropped pronouns. The problems are statistically modeled by encoding the sentence and classifying types of dropped pronouns using a recurrent neural network (RNN) architecture. Various RNN-based encoding architectures were investigated, and the stacked RNN was shown to be the best model for Korean zero-pronoun recovery. The proposed method does not require any manual features to be implemented; nevertheless, it shows good performance.     

基于自注意力机制和时空特征的Tor网站流量分析模型

不法分子利用洋葱路由器（Tor）匿名通信系统从事暗网犯罪活动,为社会治安带来了严峻挑战。Tor网站流量分析技术通过捕获分析Tor匿名网络流量,及时发现隐匿在互联网上的违法行为进行网络监管。基于此,提出一种基于自注意力机制和时空特征的Tor网站流量分析模型——SA-HST。首先,引入注意力机制为网络流量特征分配不同的权重以突出重要特征;然后,利用并联结构多通道的卷积神经网络（CNN）和长短期记忆（LSTM）网络提取输入数据的时空特征;最后,利用Softmax函数对数据进行分类。SA-HST在封闭世界场景下能取得97.14%的准确率,与基于累积量模型CUMUL和深度学习模型CNN相比,分别提高了8.74个百分点和7.84个百分点;在开放世界场景下,SA-HST的混淆矩阵各项评价指标均稳定在96%以上。实验结果表明,自注意力机制能在轻量级模型结构下实现特征的高效提取,SA-HST通过捕获匿名流量的重要特征和多视野时空特征用于分类,在模型分类准确率、训练效率、鲁棒性等多方面性能均有一定优势。