基于预训练模型的关系抽取研究综述

近年来随着深度学习技术的不断革新,预训练模型在自然语言处理中的应用也越来越广泛,关系抽取不再是单纯地依赖传统的流水线方法。预训练语言模型的发展已经极大地推动了关系抽取的相关研究,在很多领域已经超越了传统方法。首先简要介绍关系抽取的发展与经典预训练模型;其次总结当下常用的数据集与评测方法,并分析模型在各数据集上的表现;最后探讨关系抽取发展的挑战与未来研究趋势。     

Stakeholder requirement evaluation of smart industrial service ecosystem under Pythagorean fuzzy environment for complex industrial contexts: A case study of renewable energy park

This study focuses on ways to systematically evaluate stakeholder requirements when developing a smart industrial service ecosystem (SISE) in a complex industrial context. The SISE development requires considering the service requirement from both the complex industrial context and service ecosystem manners. This study proposes a systematic framework for stakeholder requirement evaluation in SISE. The first part of the framework is the industrial context-viable system model with ecological thinking (IC-VESM) to elicit the service requirements for the SISE, which facilitates a systematic analysis of the service value proposition and service requirement elicitation in the operational lifecycle of an entire industrial context. This second part of the framework proposes a method for evaluating service requirements that is both feasible and systematic. This is achieved by combining the Fuzzy Kano and AHP methods in a Pythagorean fuzzy (PF) environment. The PF Kano computes the categories and determines the weights of service requirements from a consumer perspective, while the PF AHP hierarchically analyzes the service requirements and provides pairwise comparison paths for design experts. Finally, an illustrative case study in a renewable energy context was used to demonstrate the feasibility and effectiveness of the methodology. The proposed theoretical model provides more reliable and systematic outcomes than traditional methods when eliciting service requirements and evaluating complex smart industrial service solutions. The study has practical implications by providing useful insights for companies to recognize key smart service requirements in complex industrial contexts and to improve sustainable development.     

Probabilistic prediction methods for nonlinear systems with application to stochastic model predictive control

The performance of modern control methods, such as model predictive control, depends significantly on the accuracy of the system model. In practice, however, stochastic uncertainties are commonly present, resulting from inaccuracies in the modeling or external disturbances, which can have a negative impact on the control performance. This article reviews the literature on methods for predicting probabilistic uncertainties for nonlinear systems. Since a precise prediction of probability density functions comes along with a high computational effort in the nonlinear case, the focus of this article is on approximating methods, which are of particular relevance in control engineering practice. The methods are classified with respect to their approximation type and with respect to the assumptions about the input and output distribution. Furthermore, the application of these prediction methods to stochastic model predictive control is discussed including a literature review for nonlinear systems. Finally, the most important probabilistic prediction methods are evaluated numerically. For this purpose, the estimation accuracies of the methods are investigated first and the performance of a stochastic model predictive controller with different prediction methods is examined subsequently using multiple nonlinear systems, including the dynamics of an autonomous vehicle.     

A marker-less assembly stage recognition method based on corner feature

Fast and robust product assembly stage recognition is a key step in human–machine cooperative assembly. To solve the recognition problem of similar adjacent assembly stages, a marker-less assembly stage recognition method is proposed based on corner feature between an assembling product and a digital model. Considering the geometric features of mechanical products, a corner identification method is proposed based on the circumferential angle difference (CADF). Then a corner matching method based on distance constraint is studied for ICP registration to realize the point cloud registration between the product and digital model. Based on the registration relationship, a similarity algorithm based on proximity point proportion is used to calculate the similarity between models and the input assembling product. The model with the greatest similarity is taken as the stage recognition result. In experiments on four group assembling products, the average stage recognition accuracy is 96.15%, which indicates that the proposed method can solve the stage recognition problem. The corner identification method based on the CADF outperforms the Harris-3D corner detection method in the efficiency of assembly stage recognition.     

基于摩擦模型的伺服系统模糊控制技术研究

针对非线性摩擦对伺服系统性能的影响,常规PID控制法不能兼顾快速性和稳定性的需求,并且正弦跟踪存在“平顶”现象,基于Stribeck摩擦模型[2],采用了模糊控制补偿策略,通过matlab/Simulink仿真,仿真结果表明,与常规PID控制[1]相比,模糊控制具有较强的鲁棒性和抗干扰性能,能够很好地解决阶跃跟踪带来超调量大和正弦跟踪带来的跟踪误差等问题,对伺服系统工程具有实际应用价值。     

基于云模型和余弦跳跃权重的改进蛙跳算法

标准蛙跳优化算法(SFLA)有寻优精度低和易陷入局部收敛区域的缺点, 为提高其性能, 提出一种基于云模型局部搜索和余弦跳跃权重更新位置的改进蛙跳算法(CSFLA). 首先通过Tent混沌映射和反向学习生成种群, 使种群的分布更均匀, 利用云模型的正态特性对子群中的优秀个体的所在区域进行探索. 同时, 对种群中其他个体引入基于余弦函数的跳跃步长权重, 使该权重在整个迭代过程中由高以不同的速率下降, 提高种群的全局搜索能力. 最后, 将CSFLA与多个优化算法在不同类型的测试函数上进行了比较. 结果表明, CSFLA具有更好的收敛速度和精度, 能有效找出全局最优解. 并且将改进算法应用至旅行商问题, 该算法可以找到总路程更短的路线.     

融合语音、脑电和人脸表情的多模态情绪识别

本文提出了一种多模态情绪识别方法, 该方法融合语音、脑电及人脸的情绪识别结果来从多个角度综合判断人的情绪, 有效地解决了过去研究中准确率低、模型鲁棒性差的问题. 对于语音信号, 本文设计了一个轻量级全卷积神经网络, 该网络能够很好地学习语音情绪特征且在轻量级方面拥有绝对的优势. 对于脑电信号, 本文提出了一个树状LSTM模型, 可以全面学习每个阶段的情绪特征. 对于人脸信号, 本文使用GhostNet进行特征学习, 并改进了GhostNet的结构使其性能大幅提升. 此外, 我们设计了一个最优权重分布算法来搜寻各模态识别结果的可信度来进行决策级融合, 从而得到更全面、更准确的结果. 上述方法在EMO-DB与CK+数据集上分别达到了94.36%与98.27%的准确率, 且提出的融合方法在MAHNOB-HCI数据库的唤醒效价两个维度上分别得到了90.25%与89.33%的准确率. 我们的实验结果表明, 与使用单一模态以及传统的融合方式进行情绪识别相比, 本文提出的多模态情绪识别方法有效地提高了识别准确率.     

基于特征建模的连续体机械臂自适应鲁棒控制技术研究

针对连续体机械臂动力学建模困难,控制难度大的问题,使用特征建模的方法,用简化等价的特征模型设计了自适应鲁棒控制器。首先,通过对机械臂系统进行分析,使用特征建模得到等价的动力学模型,然后采用递推最小二乘法进行参数辨识,得到该特征模型的特征参数,最后,利用模型与辨识得到的参数设计了一种结合滑模控制的自适应控制器。通过仿真实验对本文方法与传统PID方法进行对比,仿真结果表明,本文的方法对于末端角度控制的跟踪精度高,同时对于末端扰动有着良好的自适应能力和鲁棒性。     

基于MFE-BERT与FNNAttention的心理医学知识图谱构建

针对心理医学领域文本段落冗长、数据稀疏、知识散乱且规范性差的问题, 提出一种基于多层级特征抽取能力预训练模型(MFE-BERT)与前向神经网络注意力机制(FNNAttention)的心理医学知识图谱构建方法. MFE-BERT在BERT模型基础上将其内部所有Encoder层特征进行合并输出, 以获取包含更多语义的特征向量, 同时对两复合模型采用FNNAttention机制强化词级关系, 解决长文本段落语义稀释问题. 在自建的心理医学数据集中, 设计MFE-BERT-BiLSTM-FNNAttention-CRF和MFE-BERT-CNN-FNNAttention复合神经网络模型分别进行心理医学实体识别和实体关系抽取, 实体识别F1值达到93.91%, 实体关系抽精确率达到了89.29%, 通过融合文本相似度与语义相似度方法进行实体对齐, 将所整理的数据存储在Neo4j图数据库中, 构建出一个含有3652个实体, 2396条关系的心理医学知识图谱. 实验结果表明, 在MFE-BERT模型与FNNAttention机制的基础上构建心理医学知识图谱切实可行, 提出的改进模型所搭建的心理医学知识图谱可以更好地应用于心理医学信息管理中, 为心理医学数据分析提供参考.