融合约束学习的图像字幕生成方法

目的 图像字幕生成是一个涉及计算机视觉和自然语言处理的热门研究领域，其目的是生成可以准确表达图片内容的句子。在已经提出的方法中，生成的句子存在描述不准确、缺乏连贯性的问题。为此，提出一种基于编码器-解码器框架和生成式对抗网络的融合训练新方法。通过对生成字幕整体和局部分别进行优化，提高生成句子的准确性和连贯性。方法 使用卷积神经网络作为编码器提取图像特征，并将得到的特征和图像对应的真实描述共同作为解码器的输入。使用长短时记忆网络作为解码器进行图像字幕生成。在字幕生成的每个时刻，分别使用真实描述和前一时刻生成的字幕作为下一时刻的输入，同时生成两组字幕。计算使用真实描述生成的字幕和真实描述本身之间的相似性，以及使用前一时刻的输出生成的字幕通过判别器得到的分数。将二者组合成一个新的融合优化函数指导生成器的训练。结果 在CUB-200数据集上，与未使用约束器的方法相比，本文方法在BLEU-4、BLEU-3、BLEI-2、BLEU-1、ROUGE-L和METEOR等6个评价指标上的得分分别提升了0.8%、1.2%、1.6%、0.9%、1.8%和1.0%。在Oxford-102数据集上，与未使用约束器的方法相比，本文方法在CIDEr、BLEU-4、BLEU-3、BLEU-2、BLEU-1、ROUGE-L和METEOR等7个评价指标上的得分分别提升了3.8%、1.5%、1.7%、1.4%、1.5%、0.5%和0.1%。在MSCOCO数据集上，本文方法在BLEU-2和BLEU-3两项评价指标上取得了最优值，分别为50.4%和36.8%。结论 本文方法将图像字幕中单词前后的使用关系纳入考虑范围，并使用约束器对字幕局部信息进行优化，有效解决了之前方法生成的字幕准确度和连贯度不高的问题，可以很好地用于图像理解和图像字幕生成。     

User activities and the heterogeneity of urban space:The case of Dahiyat Al Hussein park

Urban researchers have maintained a constant interest in the complexity and continuity of urban space usage. Some have applied actor–network theory (ANT) to investigate the heterogeneity of spaces and present them through the networks of their users’ activities. However, these accounts are predominantly limited in examining the extent to which these spaces may be heterogeneous when exploring such networks. This paper draws on recent ANT scholarship, which employs an ethnographic research conducted in a main park in a housing project at Dahiyat Al Hussein in Amman, Jordan. The findings describe the complex and unpredictable negotiations that occur within spaces by documenting the varieties and interrelations among user activity networks within this common and shared urban space. This research reveals the extent to which spaces, parks in this case, may be heterogeneous by unpacking their usage. The conclusions and insights assert the necessity of paying attention to design detail and creating designs that are responsive to evolving user activities.     

Simultaneous synthesis of sub and above-ambient heat exchanger networks including expansion process based on an enhanced superstructure model

Synthesis of heat exchanger networks including expansion process is a complex task due to the involvement of both heat and work. A stream that expands through expanders can produce work and cold load, while expansion through valves barely affects heat integration. In addition, expansion through expanders at higher temperature produces more work, but consumes more hot utility. Therefore, there is a need to weigh work production and heat consumption. To this end, an enhanced stage-wise superstructure is proposed that involves synchronous optimization of expander/valve placement and heat integration for each pressure-change sub-stream in stages. A mixed-integer nonlinear programming (MINLP) model is established for synthesizing sub and aboveambient heat exchanger networks with multi-stream expansion, which explicitly considers the optimized selection of end-heaters and end-coolers to adjust temperature requirement. Our proposed method can commendably achieve the optimal selection of expanders and valves in a bid for minimizing exergy consumption and total annual cost. Four example studies are conducted with two distinct objective function (minimization of exergy consumption and total annual cost, respectively) to illustrate the feasibility and efficacy of the proposed method.     

A process fault diagnosis method using multi-time scale dynamic feature extraction based on convolutional neural network

Unlike many other techniques used in process control, which are widely applied in practice and play significant roles, abnormal situation management (ASM) still relies heavily on human experience, not least because the problem of fault detection and diagnosis (FDD) has not been well addressed. In this paper, a process fault diagnosis method using multi-time scale dynamic feature extraction based on convolutional neural network (CNN) consisting of similarity measurement, variable ranking, and multi-time scale dynamic feature extraction is proposed. The CNN-based model containing the fixed multiple sampling (FMS) layer can extract dynamic characteristics of process data at different time scales. The benchmark Tennessee Eastman (TE) process is used to verify the performance of the proposed method.     

基于混合时间序列卷积神经网络的轴承故障诊断

在传统滚动轴承故障诊断中,绝大多数方法采用了从振动信号提取特征的诊断模式,但是这种模式必然会使原始信号降维进而导致故障信息的丢失。卷积神经网络(CNN)通过权重共享和稀疏连接直接对原始信号进行操作,实现自适应特征提取,最大化保留故障信息。受CNN原理启发,开发出了一种基于工业振动信号特征的新型诊断框架,称之为混合时间序列CNN(HTS-CNN)。首先,利用估计总体比例的方法自适应确定模型训练样本数目;其次,通过对时间序列片段进行随机组合的方式,使模型能够提取非相邻信号特征;最后,利用Softmax激活函数在模型输出端执行多分类任务。通过对凯斯西储大学及CUT-2平台轴承数据进行分析,实验结果表明:该方法能够准确、有效的对滚动轴承故障进行分类。     

Distributed adaptive state estimation and tracking by using active-passive sensor networks

Heterogeneous sensor networks (HSN) find a wide range of applications in the field of military and civilian environments, where sensor nodes are utilized to estimate the position of a target with both dynamics and control input being unknown for the purposes of tracking. In the HSN, nodes are considered active depending upon their ability to sense the target output while the others are taken passive. Accurate estimation requires local information exchange among the spatially located sensor nodes, so that the active nodes as well as the passive nodes converge simultaneously to the same value. The local information exchange among the nodes is dictated by a connected graph. By using the criterion of collective observability, a novel distributed adaptive estimation scheme is introduced via adaptive observer where the nodes are allowed to have different sensor modalities. Using the estimated information, a subset of active and passive nodes, referred to as mobile nodes, can track the moving target. By using a constant state feedback controller at each mobile node, the state and parameter estimation as well as the tracking errors are shown to be uniformly ultimately bounded. Simulation results verify theoretical claims.     

Ultra-small RuPx nanoparticles on graphene supported schiff-based networks for all pH hydrogen evolution

As a promising candidate to replace Pt in hydrogen evolution reaction (HER), ruthenium phosphide catalysts have shown excellent performance and attract increasing attention. However, it is still a challenge to control the particle sizes of RuP_x were on the supports. In this study, the Schiff-based networks with 3D rigid organic networks are successfully grown on graphene, and the obtained composite was used as a new support for the RuP_x. This composite can not only provide more metal coordinating sites but also restrict the growth of RuP_x. Therefore, ultra-small RuP_x nanoparticles are uniformly dispersed via a subsequent phosphating and pyrolysis process. The final RuP_x/NPG catalyst exhibits low overpotentials at the current density of 10 mA cm⁻² (32 mV in 0.5 M H₂SO₄ solution, 13 mV in 1 M KOH solution, and 50 mV in 1 M PBS solution). This work provides a new strategy to solve the aggregation problem of the metal phosphides on the substrate by introducing Schiff-based networks into the graphene.     

Internal and external factors of competitiveness shaping the future of wooden multistory construction in Finland and Sweden

In transitioning to a renewable material-based bio economy, growing public and industry interest is apparent for using wooden multistory construction (WMC) as a sustainable urban housing solution in Europe, but its business implications are not well understood. In our study, we evaluate, which internal and external factors of competitiveness are shaping the future of WMC, especially in the context of Finland and Sweden. Based on a multi-level perspective of socio-technical transitions, we conducted a three-stage dissensus-based Delphi study. The identified internal and external factors affecting the future competitiveness of the WMC business emphasize the importance of skilled architects and builders and the role of standardized building systems. Based on our results, the key aspects influencing the future competitiveness of WMC in the region are related to the development of technical infrastructure and project-based business networks, while additional changes in regulatory framework are perceived as less important. We conclude that towards 2030, the strong cognitive rules founded in the concrete-based building culture in these countries is likely to inhibit the dynamics of the socio-technical regime level. A change is also needed in the WMC business culture towards more open cross-sectoral collaboration and new business networks between different-sized players.     

Latent communities of digital publications: The role of editors,followers, and advertisers

This research explores how digital publications may be turned into something more than just a meeting point for people who are seeking information and may become a latent community. The paper proposes that the network of editors, followers, and advertisers of digital publications determine the user sense of community towards the publication. While the pattern of connections among editors and followers reinforces the sense of community, advertisers act as intruders who weaken the feasibility of building a community. Data were collected from followers of digital publications. The dynamics of the sense of community is described as a process in which informational value and identification are essential to nurturing readers’ soft or hard commitment. Findings indicate that the strong ties between editors and followers as well as the diversity positively impacts on informational value and identification, whereas advertising intrusiveness diminishes the informational value.     

Multi-path connected network for medical image segmentation

In recent years, deep learning has been successfully applied to medical image segmentation. However, as the network extends deeper, the consecutive downsampling operations will lead to more loss of spatial information. In addition, the limited data and diverse targets increase the difficulty for medical image segmentation. To address these issues, we propose a multi-path connected network (MCNet) for medical segmentation problems. It integrates multiple paths generated by pyramid pooling into the encoding phase to preserve semantic information and spatial details. We utilize multi-scale feature extractor block (MFE block) in the encoder to obtain large and multi-scale receptive fields. We evaluated MCNet on three medical datasets with different image modalities. The experimental results show that our method achieves better performance than the state-of-the-art approaches. Our model has strong feature learning ability and is robust to capture different scale targets. It can achieve satisfactory results while using only 0.98 million (M) parameters.